GPT-5.4, Claude-Boom and Pentagon Showdown: Remeasuring AI (2026-03-08)

GPT-5.4 vs Claude Opus 4.6: Model Race, Military Conflict, and Market Shift

Sunday, March 8, 2026

Hello, this weekly digest processes the most important new videos from around 45 curated AI and coding YouTube channels — with substance, no shallow Top-5 lists. One complete summary per video, plus a weekly overview of dominant themes. Read at your leisure — or copy a summary into the LLM of your choice and dive deeper. Click the link under each summary to watch the original video.

Model Releases & Benchmarks

GPT-5.4 dominated this week’s benchmark discussion, but Google and xAI weren’t idle. Gemini 3.1 Pro — native multimodal model with 1-million-token context and Google Search grounding — ties GPT-5.4 for first place on combined benchmarks per Artificial Analysis. The smaller Gemini 3.1 Flash Light ($0.025/million input tokens, 363 tokens/sec) is 2.5× faster than Gemini 2.5 Flash per WorldofAI and suits high-volume developer workloads. GPT-5.3 Instant arrived as an interim update reducing OpenAI’s notorious “overcotting” — fewer patronizing safety disclaimers on trivial queries. Grok 4.20 from xAI runs with a multi-agent architecture (four parallel specialist agents), context window doubled to 2 million tokens.

Local & Open-Source AI

Qwen released a new model family from 0.8 to 9 billion parameters running on consumer hardware: the 9B flagship fits in just 10 GB RAM and is said to reach GPT-4o level, thanks to distillation from a 397-billion-parameter MoE model and the new Gated Deltanet hybrid architecture with 256K native context window. Simone Rizzo showed a live demo with Ollama. Shadows on the open-source side: Qwen leadership (Technical Lead and Head of Post-Training) left Alibaba after restructuring plans — a possible signal that free open-weights releases will become rarer. Bart Slodyczka demonstrated how to run Claude Code completely free over OpenRouter by routing to free models like Qwen 3 Coder (480 billion parameters, no local VRAM required).

Claude Code & Anthropic Tooling

Claude Code was the week’s most active product category bar none. Anthropic shipped in rapid succession: native Scheduled Tasks in the desktop app (agentic, self-correcting, up to 7-day catchup); Loop feature for automated tasks up to 3 days within open sessions; Remote Control for all paid plans (mirror terminal sessions from your phone); Voice Mode for ~5% of users; a redesigned Skill Creator with built-in evals and benchmarking. Nate Herk showed live how a morning briefing skill is converted via /schedule into a daily 6 AM automation. Mark Kashef demonstrated a multi-model council skill using Claude Code as the main driver and pulling Gemini and Codex for subtasks via OpenRouter — the result was a visibly stronger landing page. Liam Ottley built a complete AI Operating System (AIOS) with Claude Code at its core: context, data, intelligence, and automation layers, controlled via Telegram, managing four companies and 60+ employees on mobile.

Coding Agents (non-Claude)

Cloudflare released V-Next — a complete reimplementation of Next.js on Vite that was built in one day with AI agents (roughly 1,100 dollars in token costs) and delivers 4.4× faster production builds and 57% smaller client bundles per Cloudflare benchmarks. Fireship showed the built-in Cursor migration agent but warned of bleeding-edge instability; Vercel’s CTO called the project a “slop fork.” Cursor Automations brings trigger-based cloud agents (GitHub events, Slack, Linear, Cron) in isolated sandboxes that autonomously create security reviews and code-cleanup PRs. CodeBuff (Apache 2.0, multi-agent coordination) claims triple the speed versus Claude Code per WorldofAI, runs on Opus 4.6, and shows parallel sub-agents with visible diffs. Theo (t3.gg) released T3 Code — a free open-source desktop tool with work-tree workflows and one-click GitHub integration that uses Codex CLI as harness and plans to integrate Claude, Cursor CLI, and Gemini.

Software Engineering & Dev Culture

Cole Medin countered the Amodei thesis on software engineering extinction with data: 90% of engineering teams use AI “in workflow,” but only 51% daily; 21% of licenses remain underutilized. His conceptual triad — Prompt Engineering → Context Engineering → Intent Engineering — describes the shift toward architecture, requirements translation, and validation strategy as core work. Tech With Tim added: Junior developers must now operate at the mid-senior level of five years ago; the hiring process is structurally wrong with Leetcode interviews. Matt Pocock and Melvynx independently described the same productivity gain: fork open-source projects and transform them into personalized apps with Claude Code in hours.

Nate B. Jones analyzed the Claude Code versus Codex harness difference as a strategic architecture decision: the same Claude model achieved 78% on Claude Code harness, 42% on the small-agents harness — purely through architecture. Five divergences (Execution Philosophy, State & Memory, Context Management, Tool Integration, Multi-Agent Architecture) create a lock-in he compares to the AWS-versus-Azure debate of 2010.

Personal AI OS & Agent Frameworks

OpenClaw (also transliterated as OpenCore and OpenCla) was the week’s most-discussed open-source platform. Multiple channels compared it directly with Claude’s ecosystem: Kyle Balmer emphasized December 2025 marked the inflection point — OpenClaw became the most-starred open-source project of all time with ~260,000 GitHub stars; Claude Code generates 140,000 public commits daily. Nic Conley had both systems build a fully autonomous social media agent: OpenClaw posted autonomously the next morning at 9:02 AM on a VPS, Claude Code’s Scheduled Tasks required a running local machine. Brian Casel established the decisive design principle: skills as platform-agnostic markdown process definitions are the portable capital — whoever builds them isn’t investing in a single tool.

Dave Ebbelaar clearly structured complexity levels: Level 1 (Augmented LLM) through Level 5 (Multi-Agent Orchestration), with the core insight that deterministic DAGs are more maintainable in production than LLM nodes with tool calls. Nick Saraev presented a complete two-hour course: agent chat rooms (five agents with different personalities debate a problem in chat.json), Stochastic Multi-Agent Consensus (10 parallel instances, mode and median), sub-agent verification loops, and the 60/30/10 cost rule (Haiku/Flash for 60% of tasks).

AI Automation & Workflows

Nate Herk showed the complete build of an Executive Assistant with Claude Code: four parallel sub-agents (morning briefing, research, pulse check, visualization) in under two minutes instead of 25 minutes manual, with YAML front-matter skills and Perplexity API for cheaper research. Ben AI demonstrated lead qualification with 150 agencies: 15 parallel sub-agents for research, 18 for enrichment via Apify LinkedIn scraper, 17 for outreach personalization — all in one Claude Cowork chat. Leon van Zyl built a personal assistant with n8n using two memory types (short-term: last 20 messages, persistent: SQL tables for user preferences) over OpenRouter with Claude Sonnet 3.5. The Datapizza Hackathon finale (19 teams, 24 hours, virtual restaurant management) crystallized a practical rule: teams that thought through the problem first and then minimally implemented consistently beat teams with mass AI-generated code.

AI Video & Content Creation

Nate Herk combined Nano Banana 2 (via key.ai) with Kling 3.0 and Claude Code to build Apple-like scroll-animation websites: start and end images of a product get animated, Claude Code extracts ~120 WebP frames via FFmpeg and maps each frame to a scroll position. Finished sites deploy to Vercel via GitHub; the process is marketed as a $5,000–10,000 service for local businesses. Google released Notebook LM Cinematic Overviews (Google Ultra Plan only, $200–250/month) with motion graphics from source material — WorldofAI showed how Notebook LM as a free Gemini research frontend saves token costs on Claude Code.

AI Business, Marketing & Freelancing

Zubair Trabzada built a free AI marketing tool with 15 Claude Code skills: /market audit spins up five parallel sub-agents for content, SEO, conversion, competition, and strategy, generating a PDF report to serve as a lead magnet for $2,000–5,000 monthly retainers. Kyle Balmer and other channels highlighted MIT research: top-down AI pilot projects often fail, grassroots training for regular employees works — creating a market for AI workshops at hundreds to thousands per hour. Everlast AI showed a daycare real-world case: voice agent “Julette” handled 40% of daily calls, eliminating a receptionist position.

PKM & Knowledge Management

Nate B. Jones introduced Open Brain — an agent-readable memory architecture with PostgreSQL, PG-Vector embeddings, and MCP server for under 30 cents/month (free tiers from Supabase and Slack). The core promise: instead of proprietary silos (Claude Memory, ChatGPT Memory), a cross-platform system enabling semantic search across all AI tools. Sebastien Dubois presented three Obsidian themes: the Time Machine plugin (combines file-recovery snapshots with Git timeline for granular partial recovery), a structured book management system with linked concept notes as knowledge graph, and the Typefully plugin v3 for direct social media publishing from Obsidian.

Prompting & AI Literacy

Tech With Tim explained four prompt patterns using Genspark demos: Speed Layer (voice dictation via Speakley), Persona Handoff (expert role with concrete track record), Input Flip (reference instead of blank canvas), and Constraint Box (multiple constraints enforce specificity). Matt Pocock formalized seven development phases with AI coding assistants — from Idea through PRD and Kanban to the QA loop — emphasizing that the prototyping step before writing PRDs prevents the most expensive redesigns. Melvynx examined how Claude models (Sonnet 4.5, Opus 4.5, Opus 4.6) select tools by statistical frequency rather than actual need — a “benchmark business” where companies like Vercel massively benefit from Claude daily recommending their tools.

AI Industry & Strategy

The Pentagon drama had direct financial consequences: OpenAI secured the largest private funding round in history per Nate B. Jones ($110 billion, with Amazon $50B, Nvidia $30B, SoftBank $30B), while Anthropic despite stronger enterprise access was cut off from government projects. Multiple channels flagged the circular-financing risk: cloud providers (AWS, Azure, Google Cloud) are model-agnostic and want token volume; whoever controls the infrastructure controls AI’s power grid. Everlast AI’s expert panel reported Anthropic triggered nearly a trillion in market value destruction at legacy SaaS companies in three weeks when Claude Cowork plugins became available for Thomson Reuters, Legal Zoom, and IBM. Google’s AI Center Berlin opened with 18 demos: real-time situation display for Düsseldorf Fire Department, Mercedes car interface, Uniklinik Cologne patient hotline.

TheAIGRID documented industrial distillation attacks: Anthropic reported 24,000 fraudulent accounts generating 16 million Claude exchanges (DeepSeek, Moonshot AI, Minimax as suspected actors). The timing coinciding with a Trump administration decision allowing chip exports to China under conditions raised questions whether this is genuine security concern or lobbying for stricter export controls.

AI & Society / Future of Work

Jack Dorsey announced Block laying off 4,000 of 10,000 positions — with growing business; the stock rose ~20%. This is the first prominent example per multiple channels (Kyle Balmer, Everlast AI) of AI-driven headcount reduction at a profitable growth company, not as emergency measure. Dario Amodei publicly warned of a “tsunami” — 50% of all entry-level jobs at risk by 2027. Nate B. Jones developed the concept of “Frontier Operations” as the new key competency: five integrated skills (Boundary Sensing, Seam Design, Failure Model Maintenance, Capability Forecasting, Leverage Calibration) that can’t be learned once because the human-agent boundary shifts with every model release. French-language channel “Le SamourAI” analyzed the economic collapse of the daily-rate system: GPU clusters replace sequential, copyable human cognition, and only three profiles will survive — the CAP architect, the context guardian, and the algorithmic plumber.

Quick Notes

Blitzy (enterprise coding agent, thousands of cooperative agents, pricing from 50K credits) showed per Tech With Tim 150 hours of team work compressed to 6 hours validation. Hostinger Horizons built a complete Schengen visa tracker as a fullstack app including auth in ~4 minutes in a demo session, without manual server configuration. Simone Rizzo presented Rizzo Mesh, a didactic fine-tuning project: a Gemma base model learned 3D OBJ syntax via LoRA on 50,000 Thingiverse objects (cost ~300 dollars, 49 hours, 4.92 GB GGUF result). Perplexity Computer (Perplexity Max, $200/month) runs with Claude Opus 4.6 as reasoning engine plus Gemini, Nano Banana, Grok, and GPT-5.2 for specialized tasks and per TheAIGRID is the most accessible general-purpose agent for non-technical power users. Everlast AI interviewed Ernst Dickmanns — the autonomous driving pioneer drove 1,000 km on the A1 at 130 km/h in 1994 without GPS, camera-only; his 4D approach (3D + time) is considered the conceptual precursor to modern agent architectures.

AI Explained (1 new video)

• What the New ChatGPT 5.4 Means for the World
  6.3.2026, 16:25:26

  The video covers the release of GPT 5.4 and its implications for white-collar professions as well as fundamental geopolitical tensions between OpenAI and Anthropic surrounding military contracts.

  On GPT 5.4 itself: The model was tested blind against human performance across 44 GDP-relevant professions and outperforms human first attempts in 70.8% of cases (83% including ties). An important caveat: the model hallucinates more often when wrong instead of admitting it doesn’t know something (89% bullshit rate per Artificial Analysis). The benchmark is based on independent, digital tasks and is not representative of the full range of actual professional work. Paradoxically, GPT 5.4 Pro (the most expensive version) performs worse on this benchmark than standard GPT 5.4.

  Software Development: The new CodeExec integration shows impressive progress—such as creating an animated league table for a soccer club “one-shot”—with the model combining various tools and web searches. Computer use has improved dramatically; GPT 5.4 can spot and correct errors in an almost closed feedback loop. Similar quality jumps are evident in presentations and spreadsheets compared to earlier versions.

  Uneven Progress: Despite these breakthroughs, OpenAI’s own system cards reveal “spiky performance”: internal ML benchmarks double (12% to 23%), but on OpenAI proof Q&A (questions about actual engineering bottlenecks), GPT 5.4 Thinking even underperforms GPT 5.2. This reflects the central debate: can specialized training on one domain lead to better generalization, or do you need specialized data for progress in each niche?

  Military Contracts and Anthropic vs. OpenAI: The Pentagon fired Anthropic from the defense contract after the company drew red lines on autonomous warfare and surveillance. Sam Altman and OpenAI got the contract. In a leaked memo, Anthropic CEO Dario Amodei criticizes OpenAI for accepting “security theater”—merely a classifier layer on legal unrestricted use. The Washington Post reports Claude is already being used in the Palantir system for Iran operations (target coordinates, prioritization). Amodei later revises the memo’s tone; the Pentagon states negotiations are over. Google DeepMind also works with the government but remains silent about it.

  Implication for Professionals: The speaker recommends not working without the best AI tools in 2026—deploy GPT 5.4, Gemini 3.1 Pro, Claude 4.6 Opus, and others. He introduces his own benchmark tool (part of his platform) that lets you test models blind against custom texts and PDFs.

  Bigger Picture: Anthropic has shifted from its founding mission (safe scaling, safety research only)—now scaling like others, but under the narrative “only because others do it.” The discourse around these companies is “gaslighting” and unclear; there are no clear good/evil roles.

  The video concludes with a working Viking timeline interactive demonstrating the closed loop.

  Explicitly covered in the video: OpenAI (GPT 5.4, GPT 5.3 CodeEx, Altman), Anthropic (Claude 4.6 Opus, Amodei), Google DeepMind (Gemini 3.1 Pro), Palantir, XAI; Format: News update + deep dive with personal analysis and demo.

AI Foundations

No new videos in this period.

AI with Arnie (3 new videos)

• OpenClaw is dangerously good
  7.3.2026, 12:45:36

  OpenCrow: The AI Agent that actually gets things done

  What is OpenCrow?

  OpenCrow is a personal AI agent that executes concrete tasks—not just answers. It can be reached via WhatsApp, Telegram and other channels, runs on a private server or VPS, has browser access, can modify itself when needed Skills are missing, and features persistent memory plus system access.

  Installation:

  Installation happens via one-liner on any machine. A virtual private server is recommended (e.g. Hostinger KVM2 plan with one-click setup) since fewer things can go wrong. After entering data, you need API keys for AI models (Anthropic, OpenAI, Gemini, xAI) and the Telegram bot token. Onboarding with Anthropic is recommended, then you can switch to cheaper models (e.g. ChatGPT subscription via OAuth token).

  7+ concrete use cases:

  1. Fitness management: Training and nutrition logging with macro tracking, personalized meal planning based on goals, photo analysis of eaten meals—replaces MyFitnessPal and similar apps
  2. Video automation: Automatic video editing (remove silence, transcription with Whisper, error detection, cutting via FFmpeg)—saves ~40% editing time
  3. Image creation and editing: Create thumbnails (via Nano Banana), local image processing via custom Skills (e.g. trigger ComfyUI)
  4. Internet access & research: Brave Search for general searches, Perplexity for research, Last-30-Days-Skill for current news from Reddit, X, YouTube, Hacker News and Bluesky; YouTube video summaries; newsletter aggregation
  5. Browser automation: Full browser control for websites without API access; watch out for cookies and prompt injections
  6. Mail management (with caution): Sending emails via webhook agent recommended—avoid read access (security risk from prompt injections); better: separate agent for sending only
  7. Code orchestration: Orchestrate Cloud Code, Codex or other agents; e.g. analyze gameplay videos, generate code specs and build prototypes in minutes
  8. Invoice writing: Special Skills (e.g. from community member Benny) can automate invoices
  9. Smart home control: Control cameras, speakers, alarm systems; proactive monitoring via heartbeat file (every 30 min)

  Save costs & performance:

  • Model routing: Large models only for difficult tasks, cheap models (e.g. Gemini Flash) for simple queries
  • Context management: /context list shows token consumption; /new resets context; /compact compresses sessions; save context data as markdown for later reference
  • Superagents: Separate personas (e.g. marketing agent with own profile, cheaper model) for specialized tasks—reduces token costs
  • Multi-agent architecture: OpenCrow as CEO with various domain-specialized agents (fitness, marketing, programming etc.)

  Memory & security:

  • Push configurations to GitHub as backup
  • Vector databases + SQL databases for distributed memory
  • Connect MCB server to share memory with other LLMs (CloudDesktop, ChatGPT, Cursor etc.)
  • No email read permissions: Prompt injection risk with email read access; better: webhook agent with write access only

  Not optimal:

  Managing large Telegram group conversations (like ChatGPT with hundreds of chats)—OpenCrow is better used as single conversation, control storage via memory/markdown/databases.

  Demo video without detailed transcript available—shows live setups, API key configuration, Telegram integration and numerous practical examples.

  Featured tools/providers: OpenCrow, Anthropic Claude, OpenAI (ChatGPT/o1/o3/Codex via OAuth), Google Gemini, xAI Grok, hosting via Hostinger VPS, n8n, Floway (mentioned), Telegram, Nano Banana (image gen), ComfyUI (local), Ollama, Whisper, FFmpeg, Brave Search, Perplexity, YouTube API, GitHub, MCB (memory sharing system). Format: Tutorial/demo with extensive practical use cases.

• Almost everyone uses AI wrong
  6.3.2026, 13:53:33

  CLI tools for AI models: installation, usage and local alternatives

  The creator argues that browser usage of ChatGPT, Claude or Gemini is inefficient—you constantly copy context back and forth, switching models feels cumbersome. The solution: CLI tools (Command Line Interface), which are ten times more powerful and accessible to everyone, not just programmers.

  The most important CLI tools

  Claude (Anthropic), Codex (OpenAI), Gemini CLI and Open Code are the best known. Choice depends on your subscription: Claude Code with Anthropic plan, Codex with ChatGPT subscription, Gemini CLI with Google subscription, Open Code free with barely any limits.

  Installation

  Installation usually works via a simple one-liner (e.g. for Gemini: corresponding one-liner), depending on the operating system. On Windows copy PowerShell, for npm-based tools Node.js must be installed. After execution, authentication with API token or account details usually follows.

  Practical advantages over browser

  The CLI tools have access to the local machine: they can execute commands, install/uninstall software, check files. Example: instead of searching for Blender yourself, you just ask “Do you have Blender installed?” or “Uninstall Blender for me”—the tool does it. Internet research also works, e.g. fetch Bitcoin price, gather news and save a report as markdown file on the desktop—all automated.

  Context management and parallelization

  Unlike ChatGPT, CLI tools can be leveraged heavily for large projects. Example: A book on bodybuilding nutrition (100 pages, 18 chapters) is created with Claude Code in plan mode. The tool launches multiple parallel agents for different chapters, uses intelligent context management so each agent handles only their part but maintains the narrative thread. The book is created in minutes instead of hours.

  Local files and VS Code integration

  All files stay local on the PC; you open a folder in VS Code and can run the CLI in parallel there. With slash init the tool creates a project memo (Cloud.md or Agent.md) that helps new agents get oriented faster. You switch flexibly between models (e.g. from Claude to Codex) without losing context—just a new agent, same files.

  Version control with GitHub

  Projects can easily be pushed to GitHub and versioned. git commit and git push are executed by the agent, then the project is online. On a new computer you can simply download it without manually reconstructing everything.

  Use cases

  • Books/texts: Have detailed content created with correct structure
  • Websites: Build complete single-page apps with React (contact form, multiple pages, fully functional)
  • Business data: Analyze folders with documents, meeting notes, reports and automatically summarize
  • Collaboration: Share context on GitHub so editors or team members see and edit the same files

  Local models for data privacy

  If cloud services aren’t an option: download Ollama, download and run local models like Qwen 3.5 (0.8 to 122 billion parameters). A 9-billion-parameter model (6.6 GB) works on typical PCs with 8–16 GB VRAM. Command: ollama run qwen:3.5 → Cloud Code loads via it locally, completely privacy-compliant and free. Even smaller models solve everyday tasks (e.g. email log analysis).

  Conclusion

  CLI tools are the real win: full control over files, quick model switching, parallel agents, version control, local options. Instead of going the browser route, you should build “your own operating system” based on these tools. The creator mentions a paid course (“AI Revolution → Classroom → Vibe Coding”) covering all slash commands, Skills, plugins and more complex use cases in depth. In parallel, Thomas can build automations professionally.

  Explicit AI tools: Claude (Anthropic), Codex (OpenAI), Gemini CLI (Google), Open Code (free), Ollama (local models) — Format: Tutorial/demo with opinion; difficulty: beginner to advanced, as the video deliberately emphasizes accessibility.

• The AI crash is coming
  2.3.2026, 17:10:14

  Summary: The global intelligence crisis 2028

  The video presents a bleak scenario: according to this thesis, the economy is experiencing a historical illusion—while record profits appear on paper, a deflationary death spiral is brewing below the surface. The central problem: 70% of the economy relies on human consumption, but AI agents work 24/7 without sleep, need no salary and buy neither cars, take vacations nor dine out. The thesis predicts a $13 trillion crisis for 2028 worse than the 2008 housing bubble, with 10.2% unemployment and 38% drawdown in the S&P.

  Recent AI developments accelerate this:

  Recent weeks showed massive updates: Claude Sonnet 4.6 from Anthropic, Grok 4.20 from XAI, Gemini 3.1 Pro from Google plus tools like Opal and Inception. A benchmark shows the escalation: in 2020 GPT-3 could automate 9-second tasks, in 2023 GPT-4 could do 3-minute tasks, in 2025 O3 could handle 2-hour tasks—by end 2025 GPT-5.2 High reached 6-hour tasks, current models like Claude Opus 4.6 manage 14.5 hours of complex work (e.g. bug fixes in ML codebases) in one go with 50% success rate. This enables massive parallelization at minimal cost.

  Concrete examples of disruption:

  Andrew Karpathy would have needed a weekend project for a local video dashboard for home camera tracking—now an agent handles it in 30 minutes. Jack Dorsey of Block announced laying off 4,000 of 10,000 employees even as the business grows; Block stock rose 20% on the news. Dario Amodei of Anthropic publicly warns of an extreme job market crisis in 1–5 years, with executives discussing behind closed doors.

  The crisis scenario unfolds like this:

  Well-paid jobs are cut → debt defaults → consumer crash. AI agents pay no taxes and find cheapest prices without brand loyalty → deflation. Profits flow almost exclusively to BigTech players (token providers, efficient agent deployers). Humans become the bottleneck, no longer economically necessary.

  Further technical breakthroughs:

  Programmatic tool calling saves up to 90% tokens by having LMs write code directly instead of JSON. AI systems like latest Cursor updates launch virtual machines, validate and iterate autonomously until everything works.

  Geopolitical context:

  The Pentagon classified Anthropic as a “supply chain security risk”—a precedent normally applying to companies like Huawei—because Anthropic built in guard rails against autonomous weapons and mass surveillance. Meanwhile, Samer’s experience (at Meta, safety lead for superintelligence) shows fragility: AI deleted hundreds of her emails; prompt injections and jailbreaks remain unsolved—autonomous weapons under such conditions deemed catastrophic.

  The author’s counter-position:

  The creator thinks the scenario too pessimistic. He argues: (1) If the biggest concern is the economy becoming too efficient, solutions will be found. (2) Truth lies somewhere in the middle—not all jobs disappear by 2028. (3) Demand for better software rises. (4) Responsibility, taste and emotional intelligence AI can’t replace. (5) People taking responsibility will have more to do in AI times. (6) The future remains uncertain; nobody predicted three years ago how things look now. The creator concludes by promoting his AI masterclasses (AI Revolution, Classroom with courses on Claude Code, prompt engineering, app development, multi-agent systems).

  Explicit tools/providers: Claude Opus 4.6 (Anthropic), Grok 4.20 (XAI), Gemini 3.1 Pro (Google), GPT-5.2 High / O3 (OpenAI), Opal, Inception, Cursor — Format: Opinion/reflection with deep analysis of a crisis scenario.

AI News & Strategy Daily | Nate B Jones (7 new videos)

• GPT-5.4 Let Mickey Mouse Into a Production Database. Nobody Noticed. (What This Means For Your Work)
  7.3.2026, 16:00:00

  Summary: GPT-4.5 in detailed comparison

  The video offers a comprehensive evaluation of GPT-4.5 through blind evals against Claude Opus and Gemini 3.1, combined with real-world work tasks and comparative testing.

  Key insight: GPT-4.5 is not simply the best model – it’s the most interesting because it performs highly segmented. The distinction between Auto-Mode and Thinking-Mode is critical: In Thinking-Mode, GPT-4.5 competes for first place on factual accuracy (correct Higgs boson mass, Apple stock prices), while Auto-Mode becomes the worst model (cites 2024 Nobel Prize winners for 2025 questions). This difference runs through all tests.

  Where GPT-4.5 dominates: (1) Quantitative models – builds statistically more rigorous spreadsheets with Pythagorean Win Expectancy and ELO systems; (2) File processing – handles 99.1% of file types (CSV, Excel, JSON, PDF, OCR images, corrupted backups) versus Claude’s 75%; (3) Self-awareness – can describe the AI landscape and its own capabilities better than any competitor.

  Where GPT-4.5 is weak: Writing (creative and business prose worse than Opus), verbal creativity (Opus finds better wordplay), data filtering (lets Mickey Mouse and $25,000 car washes through), and fast execution (56 min vs. Claude’s 15 min on complex data migration). The core pattern: GPT-4.5 builds infrastructure without judgment – executes pipelines instead of understanding problems.

  Strategic implication: With the hiring of Peter Steinberger (Codeium/Copilot founder) shortly before release, it becomes clear that OpenAI is betting on agentic systems – not better writing. The emphasis on computer use, tool search, long-running tasks, and agentic architectures in the documentation points to an OpenAI-native agent platform. This also explains the velocity (monthly releases, pricing for sustained token load rather than chat turns).

  Practical implications: For writing-intensive work (executive communication, editorial), Opus 4.6 remains better. For structured analytical work with data volumes, especially with many file types, GPT-4.5 (in Thinking-Mode) is worth testing – the investment pays off with clear success criteria. The constant mode-switching is however a UX problem that users must manage themselves.

  OpenAI, Claude (Anthropic), Gemini (Google) in head-to-head comparison across structured evals and real tasks; deep-dive with detailed cost-benefit analysis.

• Claude Code vs Codex: The Decision That Compounds Every Week You Delay That Nobody Is Talking About
  6.3.2026, 15:00:02

  Summary: AI Harnesses – The underestimated architecture behind AI coding tools

  The central point: When evaluating AI coding tools, evaluation focuses almost exclusively on the model (the intelligence) – Claude vs. GPT-5.3 Codex, whose performance converges. What matters is the harness – the overall architecture that determines where the AI works, what it can accomplish, and how it integrates into workflows.

  A benchmark from the AI Engineer Summit makes this concrete: The same Claude model achieved 78% in the Claude Code Harness but only 42% in the Small Agents Harness – nearly double performance through pure architectural differences, not the model itself.

  The five critical divergences between Claude Code (Anthropic) and Codex (OpenAI):

  1. Execution Philosophy: Claude Code follows “Bash is all you need” – gives the agent Unix primitives (grep, git, npm) and lets it chain them. Codex builds specialized tools with Chrome DevTools, ephemeral observability stacks (Victoria Logs/Metrics), and DOM snapshots.
  1. State & Memory: Claude Code stores knowledge in structured artifacts (cloudprogress.txt, JSON features, Git commits) – the agent remembers progress across sessions. Codex uses the repo itself as system of record – architectural decisions, principles, all documented there. Codex develops “slop” itself, so Codex builds automated cleanup processes that open refactoring PRs.
  1. Context Management: Claude Code manages context through compression, summaries, and parallel sub-agents. Codex uses isolated cloud containers – no contention for a central context window.
  1. Tool Integration: Both use MCP (Model Context Protocol). Anthropic has “Skills” (Markdown + scripts on the file system) – the agent sees only brief descriptions (50–100 tokens), not complete instructions (thousands of tokens). OpenAI uses bidirectional JSON-RPC that agents call programmatically against endpoints. Example: Composio had to build a custom proxy to connect Codex with Figma/Jira MCPs.
  1. Multi-Agent Architecture: Claude Code spawns sub-agents with their own context windows, shared task lists, and can message each other (example: API agent + frontend agent + test agent). Codex runs each task in isolated sandboxes – coordination via Git branches. Claude Code allows more parallelism, Codex is safer for autonomous work.

  The lock-in problem: A developer gradually builds specific workflows: /commit, /worktree, /implement, /implement all – all built on Claude Code Harness features. Switching to Codex doesn’t just mean learning new commands, it means rebuilding the entire automation from scratch in a potentially incompatible architecture. This multiplies across teams, markdown files, MCP connections – a strategic lock-in similar to the AWS vs. Azure debate from 2010.

  Practical approach: Developers like Calvin French Owen use both tools complementarily – Claude Code for planning, orchestration, creative exploration (via sub-agents with Haiku), Codex for actual implementation (fewer bugs). Sometimes Codex lets Claude Code review.

  For leaders: The decision is not “which tool to buy?” but “what architectural philosophy do we organize work around?” This shapes not just velocity but also security (Claude Code: full local access, risk mitigated by incrementalism; Codex: isolation = safer), switching costs, and the ability to change workflows later.

  Outlook: These harness differences are already “leaking” into knowledge-work tools like Codeium (Claude Code foundation) and shaping how non-technical teams will work with AI.

  Explicitly covered: Claude Code / Codex (OpenAI), Anthropic, co-work, MCP, Composio; Format: Opinion/Deep-dive.

• OpenAI Leaked GPT-5.4. It’s a Distraction. (The AI Lock-In No One Is Talking About)
  5.3.2026, 15:00:27

  Summary: The Real AI Bet – Enterprise Context, Not Models

  The video argues that the real significance of OpenAI’s $840 billion valuation lies not in ChatGPT 5.x but in a much more ambitious goal: creating an enterprise-wide context platform that serves as a new system of record.

  The core problem: Enterprise understanding is today fragmented – distributed across GitHub, Confluence, Salesforce, Jira, Slack threads, and the minds of employees. The synthesis layer remains human and fragile. When senior engineers leave the company, information doesn’t disappear – but the knowledge about how everything connects does.

  The promise: An AI context platform that continuously learns from all enterprise “filing cabinets,” maintains a coherent model of organizational competency, and operates through agent workflows – a kind of digital institutional memory that performs synthesis work in seconds.

  The four central technical bets that must all work simultaneously:

  1. Intelligence × Context = multiplicative: Weak models drown in large context volumes and hallucinate. Only strong reasoning models can determine what information is relevant and weight contradictions across sessions. Each GPT-5.x point release is functionally necessary – not to beat benchmarks but to build the intelligence grounding that makes the context layer usable.
  1. Memory that doesn’t decay: Organizational knowledge is fragile – decision rationales from Slack threads, architectural patterns that have changed, dependencies between systems that only experienced engineers know. The memory system must stay current, resolve contradictions, flag outdated insights, and track what’s current vs. historically-relevant. This is still an open research problem.
  1. Retrieval at trillion-token scale (the core problem): RAG works for fact lookup but fails on enterprise context – can’t process temporal causality chains, confuses old contexts with new systems on similar keywords, generates more misses the larger the corpus. A solution probably needs hybrid architecture with structured indexing, hierarchical storage, temporal tracking, and state-space compression. This remains invisible in current benchmarks – and the strategic advantage won’t be measurable externally.
  1. Execution accuracy: Autonomous agents must maintain 99.5%+ success rates across hundreds of tasks over weeks – otherwise even 5% errors compound into system risk.

  The flywheel scenario: Month one: generic but intelligent agents. Month three: agents know cross-silo details. Month six: agents know things no human knows. Onboarding shortens from weeks to days. Management becomes collaboration with agents to make the right decision then delegate.

  Anthropic vs. OpenAI: OpenAI builds architecturally from the top – $600B infrastructure with AWS for a “stateful runtime environment.” Anthropic accumulates organically through Claude Code (>50% of enterprise coding markets) – bottom-up, product-driven, unconsciously as a strategic asset. Ironically, Anthropic’s organically-built context could be more realistic than OpenAI’s day-one context (because it reflects real workflows, not anticipated ones). But if OpenAI gets there first and it works, enterprise lock-in is absolute – switching costs years of synthesized knowledge.

  New lock-in quality: Salesforce locks in through data ownership. The context platform locks in through understanding – not portable, not exportable. Deeper lock-in than ever before in enterprise software.

  Practical questions for now:

  • Where does real understanding accumulate in your org (not data storage)?
  • Are you building a compound flywheel or experimenting randomly?
  • What’s your switching cost from your current AI investment?

  The conclusion: OpenAI and Anthropic are playing the longer game – not to be better next month but to become unavoidable first.

  Mentioned systems: OpenAI (Stateful Runtime Environment with AWS), Anthropic (Claude Code, Codex), Salesforce, ServiceNow, Google (mentioned as competitor), plus generic enterprise stack tools (GitHub, Confluence, Jira, Slack).

  Format: Deep-dive / Strategic analysis (opinion/reflection).

• Everyone You Know Is About to Try Claude (I Showed 3 People for 5 Minutes — All 3 Switched)
  4.3.2026, 15:00:47

  Claude vs. ChatGPT: Practical differences and how to use Claude effectively

  The video provides a detailed guide for people new to Claude who shouldn’t simply transfer their ChatGPT habits. The differences are fundamental: Claude and ChatGPT were developed with completely different training methods and therefore have different strengths.

  Seven core principles for using Claude:

  1. Claude is more critical: Trained with Constitutional AI against explicit principles like honesty, rather than optimized through user feedback like ChatGPT. Claude flags problems in your plans more readily rather than smoothing them over – this prevents costlier mistakes like poorly-thought strategies.
  1. Describe the situation, not just the desired output: Claude responds much better to well-framed context than to bare commands. It uses additional context information to rethink task framing, not just elaborate answers – it acts like a thinking partner.
  1. Claude is better at editing: Claude wins comparative tests on structural coherence and more natural, less generic-feeling prose than ChatGPT. Give Claude existing work to edit rather than blank canvases – it spots weaknesses in argument structure better.
  1. Use Extended Thinking: Claude can work through complex problems step-by-step with increased processing and show you the reasoning chain. Unlike ChatGPT, you can watch Claude’s reasoning as it runs and interrupt to correct course – instead of waiting only for the final answer.
  1. Use Projects as a workspace, not just file storage: Detailed operating guidelines in project instructions (role, audience, context, brand voice) are followed by Claude with 94% consistency, while ChatGPT reaches 87%. Claude doesn’t drift from set principles as quickly.
  1. Claude can work on your computer (Claude Codewerk): Unlike ChatGPT, Codewerk (available for macOS, Windows coming) can actually open, edit, and manage files autonomously – not just discuss them. You can have it analyze invoices and create summaries, with permission controls and real-time visibility.
  1. Know the gaps: Claude doesn’t generate images, videos (via Sora), real-time voice, and has weaknesses in math reasoning and scientific knowledge. Web search is also narrower, persistent global memory is lower, and there’s no Custom GPTs marketplace. That’s not bad – it’s just reality. You can use both.

  The core: Claude rewards more complex, context-rich inputs, better framing, and existing work for improvement. The differences stem from different training philosophies and compound over time – they shape how you use AI and how your brain thinks about AI.

  The reflection: “If you have a tool that asks critical questions, you learn to invite criticism. With a tool that always agrees, you do it less – and that changes your thinking over time.”

  Covered: Claude and ChatGPT (OpenAI), plus Claude Codewerk; Format: Opinion/Reflection with principles guide.

• Dario Amodei Made One Mistake. Sam Altman Got $110 Billion. Here’s the Full Story.
  3.3.2026, 15:15:12

  The video analyzes the geopolitical and economic consequences of three interconnected events: Anthropic CEO Dario Amodei rejects Pentagon demands for unrestricted military AI use, OpenAI signs a contract hours later for deployment on classified networks, and the US and Israel bomb Iran – with Claude provably embedded in operations. The speaker argues Amodei misunderstood the political context: his public, principled stance was punished (Anthropic designated supply-chain risk), while Sam Altman through quiet negotiation secured the largest private funding round ever ($110 billion), a defense contract, and structural control over US AI infrastructure.

  The video shows how Claude’s original enterprise dominance is endangered by these political-economic maneuvers. OpenAI received massive investments from Amazon ($50B), Nvidia ($30B), and SoftBank ($30B), while Anthropic despite stronger enterprise relationships was cut off from government projects through designation. The speaker identifies several deeper patterns: cloud providers (AWS, Azure, Google Cloud) remain agnostic to individual models and only want token volume; infrastructure consolidation (Stargate, $700B cloud commitments) creates massive switching costs; and defense contracts become the stickiest revenue form.

  A critical point: Claude was used for target identification and battle simulation in the Iran operation despite a presidential stop order – AI is so embedded in operational workflows it can’t be ad-hoc removed. The speaker distinguishes between Amodei’s technical critique (models not reliable enough) and public perception (moral objection to automated weapons). Amodei signals his positions depend on model reliability, not ethics. OpenAI’s safeguards (cloud-only, embedded engineers, independent safety stack) are procedurally different but achieve similar results to Anthropic’s requirements.

  For enterprises the central question shifts from “which model” to how cloud provider choice and agent workloads shape long-term switching costs. The speaker argues circular financing structure contains implicit risk: if enterprise token demand vastly exceeds expectations, infrastructure could be underbuilt. Simultaneously it becomes the mechanism for choosing a winner – an opaque coronation of OpenAI where regulators (especially regarding AWS exclusive frontier distribution rights) might intervene. The speaker closes with the observation that Pentagon validation and defense contracts represent the largest de-risking tool for venture capital, and the entire funding structure assumes AI compute becomes as fundamental as electricity – with OpenAI betting it can control the grid.

  Covered providers: OpenAI, Anthropic (Claude), Amazon/AWS, Nvidia, Microsoft, Google, SoftBank, Oracle, Broadcom, AMD; Format: Opinion/Reflection with deep policy and finance analysis.

• You Don’t Need SaaS. The $0.10 System That Replaced My AI Workflow (45 Min No-Code Build)
  2.3.2026, 15:00:06

  Summary: Open Brain – An agent-readable memory architecture

  The video argues that current AI tools have a fundamental problem: they lack persistent, agent-readable memory. While “Second Brain” guides exist, they’re optimized for human eyes, not AI agents. Major tech companies have built memory into individual platforms (Claude Memory, ChatGPT Memory), but these systems are isolated from each other and pull users into silos – intentional lock-in business models.

  The author points out the AI landscape is shifting radically: agents are mainstream, the agent movement has spawned millions of autonomous agents, and productivity gains are measurable (US productivity 2025: 2.7%, double the decade average). The decisive difference between successful and less successful AI users isn’t the model but the infrastructure: those who’ve built persistent, agent-readable, semantically-searchable contexts achieve compounding advantages.

  The proposed architecture (Open Brain):

  • Thoughts are stored in a PostgreSQL database (not proprietary formats) you control yourself
  • Each inserted thought is automatically vectorized as an embedding – enabling semantic search by meaning, not just keywords
  • An MCP server (Model Context Protocol – the “USB-C of AI”) connects this database to any AI tool that speaks MCP
  • Capture runs through Slack or other messaging apps: enter a thought, system extracts metadata in ~10 seconds, creates an embedding, stores it in database
  • Retrieval works from Claude, ChatGPT, Cursor, or any other MCP-compatible tool with three functions: semantic search, recent listings, statistics

  Concrete costs: 10–30 cents per month using free tiers from Supabase and Slack.

  Why architecture-level matters:

  The problem isn’t that individual tools lack memory – the problem is fragmentation. Every switch between Claude and ChatGPT currently means: new context, same information explained again. The author shows the difference: Person A spends 4 minutes re-explaining context in each chat. Person B has persistent, agent-readable Open Brain; all AI tools see the same context, one thought makes the next search smarter. This gap widens weekly (compounding advantage).

  Additional usage layers:

  • Memory Migration: At setup, move all existing memories from Claude/ChatGPT into the new system
  • Open Brain Spark: Interview prompt suggesting personalized capture categories
  • Quick Capture Templates: 5-sentence starters (decision, person, insight, meeting) for clean metadata extraction
  • Weekly Review: Automatic clustering, pattern recognition, gap analysis over the week

  Two levels of value:

  1. For agents: They finally have access to consistent, structured context without it locked in walled gardens
  2. For humans: Paradoxically, an agent-readable architecture also improves human use, because real databases + semantic search enable clearer, faster retrieval than graphical note apps

  The author argues this is actually what the original Second Brain approach always wanted – the technology (agents, MCP, cheap embeddings) is only now available. Setup takes ~45 minutes, with accompanying Substack guide step-by-step, no coding needed (confirmed by non-technical tester).

  Explicitly mentioned tools/models: Claude, ChatGPT, Cursor, Grok, Anthropic, OpenAI, Open Claw (190k GitHub stars), Zapier, Notion, N8N, MCP-Server, Obsidian, Supabase, PostgreSQL, PG Vector, Slack; as comparison: Memcync, One Context. Format: Opinion/argument with demo elements and technical deep-dive.

• Why Every AI Skill You Learned 6 Months Ago Is Already Wrong (And What Is Replacing Them)
  1.3.2026, 19:00:20

  Summary: Frontier Operations – The new core competency in the AI age

  The core argument uses a model representing AI skills as an expanding bubble: inside sits everything agents reliably do; outside sits everything still needing humans; and the surface – the boundary – is where the most valuable work happens. Each model release inflates the bubble. Tasks move from surface to inside; simultaneously the surface itself grows. This is the critical paradox: the boundary doesn’t shrink, it expands. There’s always more space for human judgment – just a different kind.

  This distinguishes this skill from all previous work skills: they have an end goal (learn writing, learn coding). This skill has no fixed goal – the surface moves constantly. You can’t learn it once and tick it off. You must move with it, continuously recalibrating as agent capabilities jump weekly or monthly.

  The author calls this practice Frontier Operations – not “AI literacy” (that’s just knowing what a language model is), not “prompt engineering” (that’s one tool), but an integrated operational practice consisting of five persistent skills:

  1. Boundary Sensing: Precise, continuously updated intuition about where the human-agent boundary lies for your domain. Recalibrate with each model release. Example: A product manager has the agent draft competitive analysis but knows it misses political dynamics between executives – the PM handles those parts themselves, delegates others.
  1. Seam Design: Architectural ability to keep transitions between human and agent phases clean, verifiable, and recoverable. Specific artifacts flow between phases; at each handoff you check if it’s on track. Critical: the right seam position changes with capabilities – the skill is constant redesign.
  1. Failure Model Maintenance: Not just generic skepticism but differentiated mental model: what specific errors does the agent make with which task types right now? Early models failed obviously; current models fail subtly – plausible analysis on false premises, correct code on the happy path, 98% accurate summaries with confident misinformation in the remaining 2%. The skill is applying different checks for task type A vs. type B.
  1. Capability Forecasting: 6–12-month bets on where the bubble boundary likely shifts next, to steer learning and workflow investments accordingly. Not perfect future prediction but probabilistic positioning (like a surfer reading waves).
  1. Leverage Calibration: Deciding where human attention goes – the scarcest resource in an agent-rich environment. McKinsey pattern: 2–5 people supervise 50–100 agents. At 10:1 ratios you must triage attention in real-time – e.g., automated tests for routine code, manual review for billing/data pipelines, deep review only for architecture decisions.

  These five aren’t checklist items but simultaneously and integratively practiced skills – like driving combines steering, speed management, route awareness, and hazard detection together, not sequentially.

  Why is this skill structurally irreplaceable? Because with each AI improvement the boundary itself grows larger, not smaller. Automated prompt techniques can replace prompting – but Frontier Operations is constantly navigating an expanding boundary. It can’t be automated.

  Organizational compound effect: Someone developing this skill 6 months earlier doesn’t just have 6 months advantage – with exponential capability jumps, the calibrated person has massive distance from the uncalibrated peer. This explains leverage numbers at AI-native companies (Cursor, Lovable, Anthropic team).

  Practical guidance for leaders:

  • Build practice environments, not courses (sandboxes with realistic error types where rules change).
  • Measure calibration, not knowledge (can the person accurately predict where agents succeed/fail?).
  • Maximize feedback density (100 calibration cycles in 10 days via 10 agent tasks/day beats a 40-hour course).
  • Create explicit roles for Frontier Operations (Automation Lead, Delegation Architect, Frontier Engineer) – the skill doesn’t develop as a side job.
  • Organizational design: Team-of-one (highly skilled person, multiple agent workflows) or team-of-five (one Frontier Operator sets seams and failure models, others execute under guidance and develop skill passively). Output scales with leverage, not headcount.

  For job seekers/self-development: Track where agents surprise you. Surprises signal uncalibrated boundary areas. If an agent hasn’t surprised you, you’re not at the frontier. Give agents intentional tasks that test your boundary model.

  Final point: The last quarter (Nov. 2024 – Feb. 2026, in video with Opus 4.5 → 4.6, context windows, retrieval mentioned) proved models get capable fast. If you don’t feel when agents improve, you’re not at the frontier. That’s the signal you’re uncalibrated – and the only answer is giving agents work that surprises both you and them.

  Context: No specific AI tool providers except references to Cursor, Lovable, and Anthropic as examples of AI-native companies; model references: Opus 4.5/4.6, Claude variants, Codex 5.3, Gemini 3.1 Pro. Format: Opinion/Reflection with operational framework – a deep-dive talk on future work capability in the AI age.

Alejandro AO

No new videos in this period.

Alex Finn (1 new video)

• OpenClaw is 100x better with this tool (Mission Control)
  3.3.2026, 00:27:55

  Summary: Mission Control for OpenClaw

  The video demonstrates how to build a custom dashboard called “Mission Control” for OpenClaw, which significantly extends the agent’s capabilities. The creator showcases his fully OpenClaw-built dashboard (without hand-coding) and presents the key tools:

  Task Board: A Kanban board that displays all tasks from OpenClaw and its sub-agents. It provides transparency over ongoing, completed, and tasks pending review, and enables automatic task processing through heartbeat checks.

  Calendar Screen: Shows all scheduled cron jobs and proactive tasks to confirm that OpenClaw is actually executing the promised automated tasks.

  Project Screen: Overview of larger projects with the ability to create hook-ups to tasks, memories, and documents. Helps maintain focus on impactful work.

  Memory Screen: Organizes OpenClaw’s memories by day and provides access to long-term memories – functions as a digital journal and life record.

  Documents Screen: Stores all documents created by OpenClaw (newsletter drafts, plans, PRDs) searchable and categorized, rather than hidden in chat histories.

  Team Screen: Visualizes all agents and sub-agents used, their roles, and the shared mission statement – functions as an organizational chart of your digital organization.

  Office Screen: A 2D pixel-art office that visualizes which tasks the agents are working on – designed to be fun and boost engagement.

  The creator emphasizes: Don’t just copy—instead use reverse-prompting after the initial setup to build custom tools that match your own workflow. All tools can be created through simple prompts; you can share the entire guide as a YouTube link with OpenClaw and it will build everything automatically.

  Mentioned tools/providers: OpenClaw, Quen, ChatGPT, Next.js – tutorial with demo focus.

Andrej Karpathy

No new videos in this period.

Bart Slodyczka (3 new videos)

• Claude Code + OpenRouter = Free UNLIMITED Coding (No RAM Needed)
  8.3.2026, 09:37:29

  Run Claude Code for free with OpenRouter

  The video demonstrates how to use Claude Code completely free without installing local models on your device – particularly valuable for computers with limited RAM, Raspberry Pi, or VPS.

  The key is OpenRouter, a service providing API access to hundreds of different AI models. On OpenRouter’s Free Models page, you’ll find options like the 120-billion-parameter GPT-OSS model or Qwen 3 Coder with 480 billion parameters – both with no input and output costs. Locally, these would require 14 to 290 gigabytes of VRAM, which is practically impossible.

  Setup takes two steps: First, generate an API key at OpenRouter, then enter four variables (OpenRouter API key, Anthropic Base URL, Auth Token, empty API key) into the settings.json file. This can be done either locally per project in the .claude folder or globally in your user Claude configuration. The video creator recommends the free “OpenRouter-free” model router at the bottom of the list, as it works more reliably than individual models during overload.

  Worth noting: You can swap the model anytime – for example, to the paid Minimax M2.5, which costs only 30 cents per million input tokens for coding versus Claude Opus at $5 per million, with comparable performance.

  OpenRouter and multiple free/paid models (GPT-OSS, Qwen 3, Minimax M2.5) are explicitly covered – tutorial.

• 7 Prompts That Make OpenClaw Agents Actually Useful
  7.3.2026, 11:36:40

  Seven prompts for more efficient OpenClaw agents

  The video presents seven prompts to make OpenClaw agents cheaper, smarter, and more useful:

  1. QMD Memory System: Improves standard keyword matching through semantic search via local RAG pipeline and vector reranking. This correctly identifies related concepts (e.g., “tired” and “sleepy”). Works on local systems like Raspberry Pi or Mac Mini with sufficient RAM, but not ideal on CPU-based VPS.

  2. Daily Morning Brief: An agency prompt for daily information summaries using web search (Tavly Search mentioned with a free plan up to 1,000 credits/month). The author recommends checking preferences after a week and iteratively adjusting the prompt – also by integrating YouTube transcripts for project-based research.

  3. Accountability Partner: The author’s favorite. The agent breaks down goals (e.g., lose 10 kg in 3 months) into schedules, conducts daily check-ins and weekly plus monthly reviews. Data should be stored in a SQL database instead of Markdown to save tokens and be available for dashboards.

  4. Accountability Dashboard: Visualizes progress data stored in the database. Lets you check goals from your smartphone or computer without querying the agent. The author emphasizes that Sonnet 4.6 or Opus 4.6 are needed to build the dashboard robustly.

  5. Self-Improving Memory System: Based on Clawhub Skills but implemented as a prompt. The agent learns from small user hints (e.g., “give me Markdown instead of HTML”) and updates its instructions, but only after three or more repetitions to avoid overreacting. Also includes privacy rules and weekly maintenance.

  6. OpenClaw Update Monitor: Automates security and update processes. The agent periodically checks for updates, waits 24 hours for error reports, checks for breaking changes, and after confirmation runs health and security checks.

  7. Context Editor: Systematically analyzes all agent files for optimization potential – redundant content, outdated info, unnecessarily verbose documentation, or duplicate statements. The goal is less token overhead per message. The author recommends first running the audit on a separate test agent.

  The author consistently emphasizes that these prompts shouldn’t simply be copied but serve as starting points for customization. Sonnet 4.6 or Opus 4.6 is recommended for more complex prompts. All prompts are available for download on his website.

  OpenClaw prompts, deep-dive.

• Learn 90% Of Claude in 20 Minutes: Chat, CoWork, Code, Skills, Connectors, Plugins
  6.3.2026, 12:09:37

  Summary: Overview of the Claude ecosystem

  The video provides a comprehensive overview of Claude’s product range and features:

  Three main products: Claude Chat is suited for simple questions and research; Claude Cowork is designed for larger, multi-step projects; Claude Code runs on the same engine as Cowork but is optimized for software development and automation.

  Chat features: Chat mode supports file uploads (PDFs, Excel sheets), a project system for organized contexts with custom instructions, Artifacts (small visual apps), and a memory system that retains important conversation content.

  Advanced concepts – Skills, Connectors, Plugins: Skills are repeatable automated processes consisting of descriptions and bundled resources like Markdown instructions, Python scripts, templates. Connectors enable access to external tools (Gmail, ClickUp, Slack) via MCPs. Plugins combine Skills and Connectors to enable Claude role-based automation. Availability of these features varies between Chat, Cowork, and Code.

  Cowork workflow: Cowork works with project-specific folders on your computer. The platform creates a claude.md file to preserve context between sessions. Scheduled tasks enable recurring reminders (e.g., daily email summaries). Cowork is designed for longer-term, multi-session projects, not immediate completions.

  Claude Code: Accesses local files and terminal, can install apps directly on your computer. Offers additional features like Remote Control (operate from your smartphone) and Hooks (deterministic hard-coded actions at lifecycle stages). Allows model switching during a session.

  Additional use cases: Chrome extension for browser automation with “Teach Claude” function for training processes; Microsoft Excel with special spreadsheet tools (Pivot tables, Conditional Formatting, formulas without hallucinations); Mobile app with access to health data.

  Model recommendations: Haiku 4.5 to save tokens in everyday use; Sonnet 4.6 and Opus 4.6 for coding tasks.

  Claude API, desktop app, browser, mobile – opinion/reflection with tutorial elements.

Ben AI (1 new video)

• Build Agent Teams within Claude Cowork in 17 min
  4.3.2026, 11:32:05

  Summary: Agents in Claude Co-work for bulk processes

  The video shows how to use Agents in Claude Co-work to process dozens of tasks in parallel and autonomously – instead of asking Claude the same thing over and over again.

  The core problem and solution:

  Previously, bulk processing was difficult because Claude could only work sequentially and the context window became overloaded. With Sub-Agents, multiple isolated agents can now run in parallel simultaneously, making processes theoretically 100x faster and saving context.

  Practical example:

  The creator demonstrates a lead-qualification workflow: 150 marketing agencies from a lead database are qualified (US-based only with SEO services). Claude spins up 15 sub-agents in parallel, each researching ~10 leads via web search. This is followed by enrichment (18 sub-agents conduct lead research using LinkedIn scraper) and finally personalization (17 agents write outreach messages). Everything runs in a single chat and takes only a few minutes.

  How it works:

  • Skills: Store specific, reusable processes
  • Commands: Combine multiple skills in sequence
  • Agents: Define specific instructions for sub-agents (in Agent.md)
  • Parallel Sub-Agents: Important: explicitly specify in the prompt “spin up 15 parallel sub agents,” otherwise everything runs sequentially

  Setup from scratch:

  You can describe everything directly in the prompt. Example: “Please create a plugin with skills for lead qualification and lead enrichment, use parallel sub-agents (5–15 tasks per agent), store agent instructions in separate Agent.md files, and create a command to trigger both skills in sequence.” Claude then generates the complete plugin with all files, which you can export as a ZIP and share with others.

  Extension: Appify MCP:

  For LinkedIn scraping and other hard-to-scrape websites, the creator recommends the Appify MCP (connect API key, activate specific scrapers via mcp.appify.com). This allows sub-agents to access LinkedIn data.

  Limitations:

  • Sub-agents consume many tokens (Claude Max plan recommended)
  • Best results with 100–200 items; beyond that it becomes complicated and expensive
  • For very large bulk workflows without human oversight, dedicated automation platforms like n8n or Make.com are still better

  Difference from Claude Multi-Agent Teams:

  Claude itself now also has Multi-Agent Teams that can communicate with each other (e.g., for compiler development). In Co-work, sub-agents are isolated – which usually suffices and costs fewer tokens.

  New plugin format:

  All skills, commands, and agents are bundled into a plugin and can be shared as a ZIP or via GitHub link. Call it simply via /plugin sales outbound-enrichment.

  The video showcases Claude Co-work with Appify in the format of a comprehensive tutorial with live demos.

Brian Casel (1 new video)

• OpenClaw vs. Claude for Running an Agent Team
  6.3.2026, 14:00:32

  Summary: OpenClaw vs. Claude ecosystem for agent delegation

  The video creator tests two ecosystems for delegating recurring business tasks to agents: OpenClaw and Claude with Claude Code and Claude Co-work.

  Vibe and trustworthiness

  Claude’s ecosystem feels polished and mature; Anthropic as a company enjoys trust through thoughtful, business-oriented releases. OpenClaw, by contrast, is the “scrappy newcomer” with paper cuts in setup, configuration, and documentation – the community fills the gaps. Still: OpenClaw made the concept of true autonomous agent delegation possible in the first place, while Claude is still forming.

  Practical delegability

  OpenClaw excels at true autonomy: agents operate from a controlled file system (Mac Mini), execute recurring jobs, and report via Slack/Telegram. The creator uses a markdown-based system with “Skills” (process definitions) synced via Dropbox, plus a custom dashboard for task management. Agents follow instructions, reference Skills, and generate reports automatically.

  Claude/Claude Co-work feel more like collaborators. Claude Co-work recently added Scheduled Tasks, but requires your personal computer to stay on and awake – and tasks fail silently. Movement between devices (desktop, mobile, cloud) is fragmented: Claude Co-work runs isolated locally, Claude Code has remote features (cloud handoff), but Skills aren’t consistently accessible in these environments.

  Strategy: Platform-agnostic processes

  The key: Skills – markdown files with repeatable processes – are the common standard across all platforms (OpenClaw, Cursor, Claude, Claude Code/Co-work). That means work invested in Skills isn’t locked to one platform. The question isn’t “OpenClaw or Claude,” but “which combination minimizes friction when deploying my processes across Skills?” Currently, the answer varies by job.

  Principle: Design jobs, not tasks – define processes, wrap them in Skills, let agents own them.

  OpenClaw, Claude, Anthropic, Cursor, n8n mentioned; opinion/reflection with deep-dive into practical workflows – written for established product builders with an agent focus.

Coding with Lewis

No new videos in this period.

Cole Medin (2 new videos)

• Is Software Engineering Finally Dead?
  5.3.2026, 01:00:59

  Summary: “Software Engineers Aren’t Going Extinct”

  The creator argues against frequent predictions from tech leaders like Dario Amodei (Anthropic) that software engineering will become obsolete in the foreseeable future. Amodei claimed in March 2024 that AI would write 90% of code within 6 months, and recently said software engineers could go extinct by 2026. Similar statements come from other industry leaders like Jensen Huang (Nvidia), Sam Altman (OpenAI), and Microsoft’s AI CEO.

  Critique of these leaders’ bias: The creator points out that these individuals have financial incentives to push this narrative. Anthropic’s valuation rose from $61.5 billion to $380 billion in less than a year. Nvidia and OpenAI also benefit massively from the AI hype. Interestingly, even statements within Anthropic contradict each other: while Amodei speaks of extinction, Boris Churney (Claude’s creator) says on podcasts later on X that “software engineering is more important than ever.”

  Core thesis: While the technology is powerful enough, two factors are underestimated:

  1. Enterprise adoption is much slower than claimed. The creator trains teams at large enterprises and observes massive resistance—security concerns, license adoption without real daily usage. Data shows: 90% of engineering teams use AI “in workflow,” but only 51% use AI tools daily. 76% of executives think their teams have adopted AI, but only 52% of engineers agree. 21% of AI licenses remain underutilized.
  1. It takes much more than just coding skills. The creator describes the evolution from Prompt Engineering → Context Engineering → Intent Engineering. The latter means: not just writing code, but defining architecture, translating requirements, clarifying success criteria, reviewing code, developing validation strategies. That’s still software engineering—just without the pure typing.

  Practical recommendations: Avoid vibe-coding, sharpen technical skills, use AI as a force multiplier, maintain understanding of your own code, don’t completely hand over control. Engineers who use AI while preserving their expertise will thrive; those who rely entirely on agents and don’t understand what they do will fail.

  Personal experience: The creator used no AI coding tools until November 2024, then built a system (Plan-Implement-Validate) and now delegates almost all coding while retaining control over planning and validation.

  AI tools/providers discussed: Anthropic/Claude Code, OpenAI, Nvidia, GitHub Copilot, Windsurf; format: Opinion/reflection with data analysis.

• Build BEAUTIFUL Diagrams with Claude Code (Full Workflow)
  2.3.2026, 15:01:28

  Summary: Excalidraw skill for visual diagrams with Claude Code

  The video creator shares a self-developed skill that enables Claude Code and other coding agents to generate visually compelling diagrams. The problem: coding agents aren’t optimally equipped for visual tasks, so he packaged his entire workflow into a skill available through the Do.cloud Skills directory.

  How it works: Clone the GitHub repository, copy the contents into a new folder in the Skills directory, and upload the skill.md file that instructs the coding agent. Then simply tell the agent to create a diagram—based on a file, YouTube script, PDF, or any text. The agent then generates JSON files that you can render on excalraw.com or with the Excalidraw plugin in Obsidian (free).

  Core feature—self-validation: The agent renders the diagram to a PNG, inspects it itself, iterates on imperfections (usually 2–4 times), and improves the JSON directly. This is crucial because coding agents without self-checking tend toward overly simple, blocky diagrams.

  Philosophy: The skill teaches the agent to argue visually: structure and labels should explain the concept without requiring explanatory text. Two guiding questions: (1) Does the visual structure reflect the concept? (2) Can you learn something concrete from it?

  Workflow steps: Formulate idea → assess diagram depth (simple vs. complex to respect token limits) → plan patterns/shapes/text → create JSON → validation loop with screenshot check → iterate. The skill.md contains design patterns, color palettes, and best practices you can customize (e.g., different colors). There’s also a reference color palette in the repo for easy swapping.

  Realistically: The skill isn’t perfect on the first try—the video creator had to iterate his own meta-diagram explanation 2–3 times. But it produces an excellent starting point you can then manually adjust.

  Result: According to the creator, the skill saves him hours per week and makes it possible to quickly generate visually complex, educational diagrams.

  Claude Code is featured; tutorial with demo focus.

Datapizza (2 new videos)

• L’accordo Pentagono-OpenAI e il movimento QuitGPT
  3.3.2026, 17:00:37

  Summary: The Anthropic-Pentagon Controversy

  The episode’s main focus covers the conflict between Anthropic and the US Pentagon over the use of AI models. A few months ago, Anthropic signed a $200 million contract with the US government, but it has become a flashpoint in recent weeks: reports indicated that Pentagon partners had deployed Anthropic’s systems in military operations (such as Maduro access via Palantir), and the Defense Department signaled intentions to use AI for domestic mass surveillance and autonomous weapons in the future.

  Anthropic CEO Dario Amodei rejected this on ethical grounds—company policies prohibit such applications. On February 25, Anthropic received an ultimatum from Pentagon Secretary Pete Hegseth with a 48-hour deadline to remove all usage restrictions. Amodei responded on February 27 with categorical rejection: mass surveillance contradicts democratic values and current technical limitations (AI is too error-prone for autonomous target selection without human control), and autonomous weapons without human oversight create no clear accountability chain in case of war errors.

  The regime responded with massive repression: Anthropic was classified as a “risk to the defense supply chain”—a designation otherwise reserved for countries like China and Russia—the 1950s Pentagon Production Act was threatened, Trump signed an Executive Order banning use across all federal agencies (blacklist). Trump and Hegseth publicly called Amodei a “radical leftist” and a “liar with a God complex.” The Pentagon argued that all planned uses were legal.

  But immediately after, OpenAI CEO Sam Altman presented a Pentagon contract and claimed it contained the same prohibitions (mass surveillance, autonomous weapons). This seemed opportunistic and triggered a boycott movement: over 700,000 users cancelled ChatGPT Plus ($20/month) and switched to Claude, especially after it became known that OpenAI licensed a GPT-4 tool to the immigration agency ICE (for screening), and that OpenAI President Greg Brockman personally donated $25 million to a pro-Trump SuperPAC. Claude subsequently became the most-downloaded free app in the US and Canadian app stores—users see Anthropic as the more ethically authentic alternative.

  Flash News: Grok 4.20 (new version) uses multi-agent architecture (four specialized agents in parallel), currently being expanded to up to 16 agents, context window doubling from 1 to 2 million tokens (largest on the market). Google released Nano Banana 2 for faster image generation (state-of-the-art in speed), storage is now available in free Google accounts too. ByteDance unveiled Dubao 2.0—a multimodal model for video and image generation, state-of-the-art for video creation.

  Topics covered included Anthropic, OpenAI, and Grok; the format was an opinion/reflection discussion with news elements and flash updates—the difficulty level is everyday for a tech-interested audience.

• Hackapizza 2.0 – Agent Edition – Premiazione finale
  1.3.2026, 16:34:42

  Summary: Data Pizza Acathon 2.0 – Agent Edition (Live Event)

  This was a live broadcast of the Data Pizza Acathon 2.0: Agent Edition, a 24-hour hackathon in which 19 teams developed AI agent systems to manage virtual restaurants.

  The Context

  The Acathon is based on a fictional universe with restaurants, recipes, and ingredients. Teams had to build multi-agent systems using the public Data-Pizza Framework to operate a restaurant: select menus, bid at auctions, procure ingredients, and serve customers. The leaderboard ranked teams by profit, but the judges (Giuseppe Gullo, Simone Rizzo, Enrico Cassol, Giacomo Charlini) also evaluated architecture, presentation, and creativity.

  The Five Finalists and Their Solutions

  1. Neural Nomads (3rd Place)

  Orchestrated multiple specialized agents: one for web scraping the blog, one for menu compilation (minimizing waste through shared ingredients), one for auction bidding (statistical median analysis), and one for pricing. With guard rails against prompt injection; fallback to deterministic systems on errors. A hybrid approach that balanced robustness with agency.

  2. Caramba (2nd Place)

  Implemented a pipeline of highly specialized agents (Procurer, Bidder, Buyer, Chef, Maître) with role-based responsibilities. Utilized a shared memory layer for state management. Developed iteratively (not top-down planned) by adding complexity layer by layer and continuously testing against live market performance.

  3. Ragù (5th Place)

  Largely deterministic despite the Agent Edition requirement. Extreme data collection (proxy servers, comprehensive logs). Started with a simple baseline, then realized that random recipe selection with expansion was optimal. Agent as fallback for unexpected scenarios (e.g., changed requirement structure). Highest performance during the golden runtime, but less in the spirit of the challenge.

  4. Cavalli al Trotto (1st Place, tied)

  Stored all data in a relational database, used RSS feeds for news ingestion. Multiple agents (Menu Agent, Bidding Agent, Market Agent, Allergies Agent, Strategy Agent) for different phases. Clever insight: a separate Allergies Extractor Agent to normalize non-standard customer inputs, rather than just yes/no decisions. Agents dynamically chose between large and small models depending on latency requirements.

  5. Claudio Opuscoli/Sonnets (1st Place, tied)

  Highly specialized with a brainstorming phase. The Cerebellone (Brain) was deterministic and orchestrated other agents (Chef with short-term memory over 3 phases, Sommelier for blog extraction, Maître for service, PR Agent). Only 33 lines for multi-agent coordination via priority async I/O instead of HTTP. 588-line strategy engine with business intelligence tools. Agents had multi-turn conversations with memory; the Chef could renegotiate (e.g., “We can only make 20 of 30 dishes—you choose.”).

  Core Insights from the Jury

  • Code alone doesn’t win: Teams that generated lots of AI-written code performed on average weaker than those who thought through the problem first, then implemented minimally.
  • Hybrid is more realistic: Deterministic baseline + fallback to LLMs (e.g., Ragù) or LLM-first + fallback to determinism (e.g., Caramba) works better than agents alone.
  • Memory and logging are critical: All top teams stored comprehensive histories and metrics (database, structured logs).
  • Generalization vs. Specialization: Specialized, role-based agents (Caramba, Claudio) beat larger, more flexible systems.

  Final Results

  1st Place (tied): Caramba and Cavalli al Trotto

  2nd Place: (Category: second winner, both still received prizes)

  3rd Place: Neural Nomads

  4th Place: Claudio Opuscoli

  5th Place: Ragù

  Hank and Giacomo from Data Pizza emphasized at the end that AI capabilities in Italy, despite all the hype, must still be applied practically and realistically, and that the next generation of technicians must not only use tools but also understand problems—this is the key to innovation.

  Technologies/Providers Discussed: Claude, OpenAI (GPT-4), Gemini, Data-Pizza Framework (open source), MCP Tools, Nano Banana, Lovable; Format: Live Q&A + demo hackathon finale with jury evaluation.

Dave Ebbelaar (1 new video)

• 5 Levels of AI Agents – From Simple LLM Calls to Multi-Agent Systems
  5.3.2026, 15:15:53

  Summary: The five levels of AI agent complexity

  The video covers five progressive complexity levels for AI systems deployed in practice.

  Level 1 – Augmented LLM: A simple API call with structured output to build systems around it.

  Level 2 – Prompt Chaining & Routing (DAGs): Classify incoming data into categories (e.g., customer tickets into “Billing”, “Technical”, “General”), then apply deterministic if-else rules for handling. This remains the backbone of reliable B2B automation in production and at scale – but becomes unwieldy as logic grows.

  Level 3 – Tool Calling at Edge Nodes: Instead of hardcoded decisions, give the LLM a set of tools (database access, policy lookup). The LLM decides agentic which tools to use, even in loops. The author shows a live example from Langfuse: a customer-care system optimized over 1.5 years with multiple steps and tool calls, automating nearly all support.

  Level 4 – Agent Harnesses: Frameworks like Claude Code or Cursor CLI give agents access to full runtime environments – bash execution, filesystem access, web search, external APIs via MCPs. The author demonstrates the Claude Agent SDK (pip install) searching local files and processing markdown knowledge bases – very powerful, but experimental and unreliable for production, plus expensive.

  Level 5 – Multi-Agent Orchestration: Multiple sub-agents with their own context windows, controlled by an orchestrator. The benefit: each agent gets a separate context window, keeping the orchestrator clean. Integrated in the Claude Agent SDK via an agents parameter.

  Production Reality: Best practice combines these levels strategically – start with simple DAGs and augmented LLM nodes, use tool calling only at edge nodes as complexity grows, deploy agent harnesses selectively (e.g., for coding), and understand that new tools like OpenClaw are just abstractions over proven principles: LLMs + tools + system prompts.

  The author emphasizes: deterministic DAGs are more maintainable and testable than LLM nodes with five tool calls. Costs and latency rise with complexity – multi-agent systems are expensive and slow, fast automations should be cheap and deterministic.

  Claude Agent SDK and Cloud Code – opinion and reflection with practical deep-dives and code demos.

David Shapiro

No new videos in this period.

DevExpert – AI for Developers (1 new video)

• From Saturating 4GB to Minimal Infrastructure: Refactoring an MVP in Go.
  5.3.2026, 16:01:29

  Postflow: From Idea to Production App in One Week

  The author describes the development of Postflow, an application for scheduling and publishing social media posts, which emerged as a replacement for the resource-hungry Postit application. The core problem was that Postit consumed all resources on a 4-GB VPS.

  Development Process

  The idea came spontaneously in the park during a walk with the kids. The author experimented with a Telegram chatbot (similar to OpenCow) and quickly iterated over requirements: an MVP should be minimal, resource-efficient, and primarily usable via an API for LLMs. Initial chatbot recommendations: Go backend, minimal API, a worker for publishing, retry logic with Dead-Letter Queue (DLQ), basic read-only UI.

  Within two days, the skeleton was functional. The author “got into the groove” and then implemented a complete user interface with Pencil (AI-powered design tool) – with calendar view, draft management, preview function, and later mobile optimizations.

  Professionalization

  Following the MVP came a deliberate refactoring phase:

  • Quality Gates: Formatting, linting, tests, coverage checks via GitHub Actions
  • Architecture: Transition to modular monolith with Clean/Hexagonal Architecture
  • File Size: Reduction of large files (e.g., api.go with 7,000+ lines) to under 500 lines through separation of concerns
  • Enforce Parity: Automated tests ensuring new features are implemented in parallel across API, CLI, and MCP
  • Testing: Black-box tests instead of shallow function calls; expanded test coverage
  • Database Migrations: Implementation for schema changes without data loss
  • Docker Releases: Automatic deployment of new versions as Docker artifacts

  End Result

  The application comprises approximately 33,000 lines of production code. Features: scheduling across multiple social media platforms (initially X, later more), attachments (images/videos), error handling with retry logic, MCP (runs directly on the server), and CLI. The project is available open source. The author used multiple AI agents in parallel and worked mainly through Telegram interactions, sometimes also directly in code editors.

  Author’s Takeaways: Don’t start with perfect planning, but prototype and iterate quickly; later systematically introduce professional engineering practices like testing, architecture, and refactoring. The trade-off between speed and maintainability is consciously resolved: MVP first, then quality.

  Future Plans: Multi-tenant support for a potential SaaS model, threading/comment features, differentiation from other tools through LLM-friendly APIs.

  Context: Claude/OpenAI via Telegram chatbot, Pencil for UI design, Go backend, GitHub integration, Docker; format: reflection/deep-dive over a completed development process.

Everlast AI (4 new videos)

• GPT-5.4: It’s OVER for OpenAI! Claude Code, Google AI Center Opening & More AI News
  8.3.2026, 09:15:00

  Summary: AI News Roundup – GPT-5.4 Thinking, Anthropic Boom and Google Berlin

  This week brought three major AI model launches: GPT-5.4 Thinking from OpenAI with 1 million token context window, a unified model for reasoning, coding, and computer use that achieved 83% on the GDP Well Benchmark. The core advantage: you can pause and correct the thinking process while the model runs. In practical tests, however, Claude Opus 4.6 proved significantly faster at coding and frontend design (2–3 minutes vs. 14–20 minutes) and closer to real requirements – GPT-5.4 thought too long without delivering proportionally better results.

  Anthropic cuts off the competition’s power: The new Voice Mode in Claude Code runs natively via slash command, Ultrafunk Mode is back, Skills have been improved, and ARR revenue doubled to $20 billion (from ~$10 billion three months ago). Claude has now overtaken both OpenAI and Deepseek in B2B adoption – from 90% ChatGPT market share in Feb. 2025 to 70% Claude in Feb. 2026. Google responds with Workspace CLI for agent integration in Drive, Docs, Gmail and Chat, plus the new smaller Gemini 3.1 Flash Light model for Voice Agents.

  Google opened its AI Center in Berlin with 18 demos: 3D video conferencing (Project Beam), the new AI mode for Google Search, and real use cases showcased the power – including a real-time situation picture tool for Düsseldorf Fire Department and an auto interface at Mercedes that answers natural language questions about driving conditions. Uniklinik Cologne uses AI agents for patient hotlines and knowledge management via Notebook LM.

  Apple is regarded as an underestimated winner: The new M5 chip loads full AI models up to ~70 billion parameters locally, large models in 4-bit format – a breakthrough for decentralized, always-available AI without cloud dependency.

  Overall picture: The debate about an “AI bubble” is fading in the face of exponential development. Anthropic CEO Dario Amodei announced a “radical acceleration” for 2026 that will surprise everyone. Practical advice: whoever uses Claude Code via terminal (Skills, Agents), works with local models, and continuously upskills belongs to the winners – 99% don’t.

  Explicitly covered: OpenAI (GPT-5.4 Thinking), Anthropic (Claude Opus 4.6, Voice Mode, Ultrafunk), Google (Gemini, Workspace CLI, Berlin AI Center), Apple (M5 Chip); Cursor mentioned; Roadmap: Everlast AI proprietary tools in development — Format: News Update / Demo Collage with Live Test and Expert Interviews.

• 40% Less Effort Through AI Automation in Daycare Operations | Everlast AI Experience
  6.3.2026, 09:15:00

  David Tut, CEO of Butmelan (daycare operator in northern Luxembourg), reports on his collaboration with AI consulting firm Everlast under Leo’s leadership. Tut was initially skeptical about AI but became informed through Leo’s YouTube content and ultimately decided to partner with Everlast after the team presented clear roadmaps based on experience with other clients.

  The implementation led to significant changes: a voice agent named Julette was installed in the reception area and now handles a large portion of calls – enabling savings of at least 40% of daily phone calls and allowing the business to cut one receptionist position. Additionally, various AI chatbots and other solutions were integrated. Tut highlights the responsiveness and quick communication of the Everlast team, which enables rapid response to new requirements and anticipation of future developments.

  After a year of collaboration, Tut draws a positive conclusion: the team works reliably, the results are convincing. He urgently recommends other entrepreneurs engage with AI, as businesses that don’t digitalize and embrace AI will face massive problems in five to ten years. He advises contacting Everlast – an initial conversation costs nothing and can’t be a disadvantage.

  Topics covered: Everlast and Leo (AI consulting), Voice Agent for automated call handling – Format: Testimonial/Customer Experience Report.

• Prof. Dr. Ernst Dickmanns: The Incredible Story of Autonomous Driving
  5.3.2026, 16:15:00

  Summary: Conversation with Professor Dr. Ernst Dickmanns – The Father of Autonomous Driving

  This in-depth conversation with Ernst Dickmanns documents pioneering work in autonomous driving since the 1980s. Dickmanns developed the revolutionary 4D approach (3D space + time), which instead of evaluating individual frames, used spatiotemporal models with differential equations – a fundamental alternative to the later-dominant CNN-based image processing. With this method, his vehicle VAMP achieved 96 km/h on the highway in 1987, in 1994 the S-Class VITA 2 drove 1000 km on the A1 near Paris at 130 km/h through dense traffic, later 1758 km Munich–Denmark at 175 km/h – all without GPS, maps or Lidar, only with cameras and computers with about 1% of today’s processing power.

  The core problem was the limited computing power of the 1980s: mainframes took 15 minutes to an hour to process a single image. Dickmanns instead used a network of microprocessors, each handling a portion of the overall image – a distributed processing system that became a model for modern architectures. His 4D approach, combining predictions and prediction error feedback (Kalman filtering in image sequence processing), enabled real-time image evaluation with minimal resources.

  Key insights on current systems: Dickmanns strictly differentiates between vehicle classes and use cases. While Daimler responsibly stays on highways (planning up to 130 km/h), he criticizes Waymo and Tesla for overly aggressive claims: they claim universality but fail on dirt roads, poor weather and unexpected scenarios – for example when Waymo failed to recognize a cyclist (fatal accident). The next decades will bring autonomous driving on well-defined roads (highways, expressways); a century is needed to cover all road types. Safe 100% guarantees are unrealistic.

  On liability: the autonomous vehicle itself and its owner bear responsibility, not the manufacturer alone – provided the system has passed testing. Insurers, according to an American colleague, benefit from high damage payouts, not falling accident rates.

  Future: Radar will be decisive – small, affordable radar sensors (10–20 years) combined with cameras. Dickmanns sees potential in hybrid-training neural networks with physical differential equations to boost efficiency – but remains pragmatic: with computing power increases of 100–1000x per decade, pure deep learning might suffice.

  German automotive industry: not falling behind in manufacturing, rather too conservative in digitalization and market risks. Who prevails will be decided by the market – but American companies take more risks.

  Context: Dickmanns worked 1961–1975 at DLR Oberpaffenhofen (flight dynamics), 1971–72 at NASA on Space Shuttle; from 1995 Professor at Bundeswehr University Munich until 2001. The VAMP/VITA-2 project emerged within the European Eureka Prometheus framework (approx. €800 million, vision as key technology). His work influenced DARPA Grand Challenges (2004–2005 won by Thrun with Stanley/Lidar).

  Deep-dive into autonomous driving, historically and technically grounded; no specific tools/vendors mentioned except Daimler, Waymo, Tesla, Stanford, NASA — Opinion/Reflection with historical-scientific focus.

• Anthropic WARNS of “AI TSUNAMI” Starting 2027! SaaSpocalypse, Pentagon SCANDAL & China’s Catch-Up
  3.3.2026, 16:50:25

  Summary: AI Development 2026 – The Underestimated Revolution

  The video is a research podcast from Everlast AI featuring three experts (Prof. Oliver Mitch, Prof. Andreas Moring, Kim Isenberg) contextualizing current AI developments within their practical and societal consequences.

  Main topics:

  Anthropic vs. Competition: Anthropic has destroyed nearly a trillion in market value in three weeks – through Claude Cowork plugins for legacy SaaS companies (Thomson Reuters, Legal Zoom, IBM). Simultaneously, infrastructure players like Factsheet and Docign benefit from MCP integration. The speed of feature releases is breathtaking and shows how radically business models are shifting.

  Agent Wars: Open AI acquired OpenClaw, Anthropic and Google block API access from each other, Perplexity integrates 19 AI models. The focus is now on the application layer – models become commodities, value creation happens through customized agent configurations.

  Robotics & China: China dominates with Unitree and Honor (~90% of humanoid robots globally). Chancellor March travels to Hangzhou with top managers; German auto exports to China have fallen 70% since 2022. The strategy is data collection for world models – first wave is chaotic household robots, real production maturity comes in 5–10 years. The trade deficit has nearly quadrupled.

  Pentagon Conflict: Defense Secretary Hexev demands Anthropic remove safety restrictions from Claude (Defense Production Act). Dario Amodei draws two red lines: no mass surveillance, no autonomous weapons. 220+ Google/OpenAI employees sign solidarity statement (over 50% anonymously).

  Ethics & Double Standards: Anthropic criticizes Deepseek/Minimx for distilling Claude via fake accounts, but paid $1.5 billion fine itself for training on copyrighted books. This shows pragmatic reality: everyone uses what’s available.

  Chips & Infrastructure: Nvidia Vera Ruben brings 5x more computing power than Blackwell, 10x cheaper inference, up to 288 GB memory. Delivery in H2 2026. Open AI plans 2 GW training on Vera Ruben. An ASML software update increases wafer throughput by 50%. China attempts to scale chip manufacturing from 20nm to 7–5nm.

  Autonomous Driving: BMW and Mercedes are cutting Level-3 systems. Germany has no Level-3 production vehicles in 2026 – focus is on Level 2+ or Level 4. This contradicts earlier claims about German technological leadership.

  Energy Policy: Trump calls on tech companies to build their own power plants. Kim points out this is a political signal (communities hit the barricades), but no real solution: gas turbines are sold out through 2030, nuclear plants take 9+ years. China adds more solar capacity annually than the US has in its entire history.

  Stargate Project: After over a year, Stargate hasn’t hired staff or built data centers. Control dispute between Softbank, OpenAI, creditors. Goal dropped from $1.4 trillion to $600 billion. Amazon and Nvidia join with $50 billion and $30 billion respectively – a setback for OpenAI.

  Chinese Models: Qwen 3.5 is just 5 months behind Sonic 4.5 (October 2025 vs. February 2026). 3 billion active parameters, only 17 billion activated per token (Mixture of Experts), outperforms GPT-4.5 and Gemini Reasoning, costs a fraction of US competition, open weights.

  Labor Market & Transformation: Dario Amodei warns of a “tsunami” – 50% of all entry-level jobs at risk from 2027. Jack Dorsey lays off 40% of workforce due to AI efficiency. Professors report students having difficulty finding entry-level jobs for the first time. This isn’t linear change, it’s radical compression: an industrial revolution spanning 100 years squeezed into a decade.

  Benchmarks saturated: GPT-4.5 and Claude Opus 4.6 already solve 40% of Frontier Math; MMLU pass rate was 2% in 2024, now over 40%. Math performance is considered an asymptote – once math is solved, physics, chemistry, biology follow like dominoes.

  Central statements:

  • AI is simultaneously the strongest disinformation machine and best truth technology.
  • Germany/Europe aren’t really in the game while USA and China battle for dominance.
  • Scaling isn’t limited by technology or capital, but by political will and societal preparation.
  • “This is the worst it will ever be” – technologically it only gets faster and better.
  • New jobs emerge, but not fast enough and not for everyone; identity crisis for many.
  • Preparation: master AI tools, think logically, embrace ethical principles, take entrepreneurial initiative.

  Format: Live Discussion Roundup with policy interviews and anecdotal evidence from practice. Difficulty Level: Intended for informed viewers, presumes basic AI knowledge.

Fireship (1 new video)

• Cloudflare just slop forked Next.js…
  2.3.2026, 20:36:25

  V-Next: Cloudflare rebuilds Next.js on Vite

  Cloudflare has released V-Next – a complete re-implementation of the Next.js framework, built on Vite instead of Node’s classic architecture. This solves a historic problem: Next.js apps were locked into Vercel because the build output was proprietary. Anyone wanting to deploy on other platforms like Cloudflare had to use the Open-Next project, which reverse-engineers the output and is fragile.

  Rather than building on that output, Cloudflare rewrote V-Next from scratch. The trick: they used AI agents and managed to get SSR, middleware, server actions, and streaming working in a single day. By day three, full deployment on Cloudflare Workers with client hydration was functional – all for roughly $1,100 in AI token costs. The final result covers 94% of the Next.js API.

  Since V-Next is built on Vite, it benefits from Vite’s architecture, including Rolldown as a Rust-based bundler. According to Cloudflare’s benchmarks, this leads to production builds up to 4.4x faster and client bundles 57% smaller compared to Next.js. In a practical test with the bytes.dev app, even a 5x faster build was possible – though with the caveat that bleeding-edge software can still have rough edges. The built-in Cursor agent for project migration helps with refactoring (such as transitioning to ES modules and JSX file extensions), but doesn’t get everything right on the first try.

  Vercel’s leadership reacted as expected with criticism: the CTO called the project a “slop fork,” and founder Guillermo pointed out critical security vulnerabilities, though these are less problematic since the app doesn’t have real users yet. Bottom line: A technically interesting project with genuine performance gains, but not yet production-ready for migration in most cases.

  Demo of a Cloudflare project (V-Next) with practical migration; specifically covers Vercel, Next.js, and Vite – Format: demo with critical analysis.

Greg Baugues

No new videos in this period.

AI and Strategy | Le SamourAI (1 new video)

• 55,000 posts deleted: The collapse of middle management and intellectual mediocrity
  2.3.2026, 16:00:56

  The video is an extended French-language essay on the economic collapse of the daily rate system in knowledge work through Artificial Intelligence.

  The core argument: For centuries, intellectual work was paid by time invested because the human brain was the only available “hardware” – and it is sequential, slow, with high context-switching costs and non-replicable. Today GPU clusters replace this function with a fundamentally different architecture: parallel, deterministic, massively scalable. This material shift (from human time to compute power) destroys the economic foundation of daily rates, hourly billing, and middle management layers that merely synchronized people.

  Concrete effects are sketched: Bank of New York Melon, Duolingo, Cap Gemini, Amazon, Microsoft and Accenture are cutting positions explicitly because of AI. In parallel, “micro-multis” (Midjourney, Cursor) emerge with million-dollar revenues at dozens of employees. The system migrates from “days-as-currency” to “tokens-as-currency”.

  The video then distinguishes six difficulty levels of work: (1) Volume/Effort → done, (2) pure logic → dying, (3) coordination → agonal, (4) scar tissue/terrain expertise → protected, (5) emotional/political intelligence → temporarily protected, (6) ambiguity & courage/strategic decision under risk → last human bastion.

  Who survives: three profiles – (1) the “CAP Architect” (strategy, problem identification, agent specification), (2) the “Context Guardian” (domain expert, black box audit, risk control), (3) the “algorithmic plumber” (technical architecture, security, integration with firm data).

  Who goes extinct: the “Standard Maker” – the competent employee who executes cleanly what they’re told, but without strategic depth.

  The video preaches a behavioral shift: (1) prohibition of first draft by own hand, (2) reversal of cognitive ratio (80% constraint engineering, 20% audit instead of vice versa), (3) systemic obsession (solving a problem means formulating it so agents handle it autonomously in the loop).

  Closing anecdote: An “automatic gate owner” brags that machines have replaced employees and sick notes are no longer needed – a caricature of dehumanization.

  Comment: Opinion/reflection with strong technical and economic deep-dive into Claude, generic LLM agents and inference economics; no specific tool named except incidentally Claude – rather macro-analysis of market transformation.

Julian Ivanov | AI Automation (1 new video)

• How to Make OpenClaw 10x Stronger
  2.3.2026, 17:55:03

  Five techniques for optimizing OpenCla

  The video demonstrates five optimizations for OpenCla, an open-source AI agent that functions as a personal assistant via Telegram or WhatsApp.

  Setup Basics: After installation via the Hostinger template, OpenCla is integrated with Telegram through the Botfather and connected with API keys. The speaker recommends Open Router for model selection, as it provides multiple providers (OpenAI, Anthropic, Google, Chinese models) in one place. Whisper is installed locally for speech recognition.

  The Five Optimizations:

  1. Subagents: Specialized agent instances (e.g., Cody for code) reduce the context window problem of individual models. The main agent delegates tasks, improving capacity utilization and using expensive models only for complex tasks.
  1. Topic-Based Telegram Groups: Instead of an endless chat, a group with different topic channels (Fitness, To-Dos, Journal, Content) is created. This keeps context organized and allows multiple topics to run in parallel.
  1. Self-Improvement via Skills: Skills installed through Clahub enable the agent to learn from mistakes. The “Self-Improving Agent with Self Reflection” skill activates automatically when improvement opportunities are detected.
  1. Enhanced Memory: Memory Flush stores important information before context compression, Session Memory Search provides access to old chats. The QMD Backend combines vector, keyword, and reranking search for more precise information retrieval.
  1. Transparency via VS Code: Using the Remote SSH extension, you connect to the server and see all created files, installed skills, and agent configurations. Skills can be reviewed and locally edited before installation.

  In practice, this means: The agent works more efficiently, costs less, learns independently, remembers more, and is completely transparent.

  OpenCla (open-source tool), individual optimizations of a personal AI assistant through multiple configuration steps — tutorial.

Kyle Balmer | AI with Kyle (10 new videos)

• Is Claude worth $100/month?
  6.3.2026, 18:45:00

  Claude subscription: Is it worth it?

  The speaker compares Claude with ChatGPT and reaches a nuanced conclusion: Claude is the better model, but costs significantly more and has different usage restrictions.

  Cost model comparison:

  ChatGPT offers a generous free plan that allows extensive usage without issues. Claude’s free plan is practically non-existent. The $20/month subscription for Claude is highly restrictive – anyone wanting to use Opus (Claude’s best model) quickly hits limits and is prompted to come back later. Realistic alternative: The $100/month plan (five times more expensive). With this, you can use Opus without constant throttling. For $20/month, only Sonnet is available, the weaker model.

  Business model difference:

  OpenAI has approximately 900 million users and targets the mass market (B2C). Anthropic focuses on an elite of 30–50 million users with purchasing power. Despite a much smaller user base, both generate similar revenue – Anthropic monetizes high-paying power users, OpenAI is planning ads for the masses.

  Conclusion: For personal use, $100/month is too much for many. For business use, it’s a highly profitable investment. The pricing structure shocks many switchers from ChatGPT who are accustomed to generous free usage.

  At the end, the speaker reflects on the importance of authenticity – he produces live content rather than AI avatars because he believes that genuine human connection and public speaking will gain value in the future, while AI-generated content will lose its rarity value.

  Topics covered: Claude, OpenAI/ChatGPT, Anthropic; Format: Opinion/reflection with price comparison.

• AI Update: Anthropic Pentagon Fallout, Google Workspace CLI & NotebookLM Goes Cinematic
  6.3.2026, 07:30:18

  The video covers several recent developments around AI and geopolitics.

  Pentagon negotiations with Anthropic: Anthropic is back at the negotiating table after failed talks with the Pentagon. The conflict arose because the Pentagon wanted to use Anthropic’s AI models for a broad range of “lawful purposes” – including mass domestic surveillance. Anthropic declined and subsequently withdrew from the military supply network. CEO Dario Amodei is conducting negotiations directly with defense officials; the speaker criticizes this as a strategic mistake, as Amodei comes across as too confrontational, while a more diplomatic representative would have better chances. The speaker clarifies: Anthropic has no fundamental problem with lethal autonomous weapons systems – only with the possibility that they could harm Americans. Attacks on non-Americans are acceptable. This clarifies media reports portraying Anthropic as a proponent of mass protection.

  Claude hype and reality: Following OpenAI’s Pentagon deal, millions of users migrated to Claude. The speaker warns against idealizing: Claude is not morally superior, just differently positioned – and should there actually be a Pentagon deal, this goodwill will quickly evaporate.

  Google Workspace CLI: Google has released a command-line interface for Google Workspace that allows AI agents direct access to Mail, Drive, Calendar, Sheets, Docs, and Admin functions – without browser automation. The speaker calls this “boring but important” – infrastructure for automating workflows, similar to a universal remote for Google Workspace.

  Notebook LM cinematic videos: Google has expanded Notebook LM with cinematic video overviews (currently only on the Ultra plan), which use their video models to generate dynamic, cinematic videos from source material – rather than just static slideshows. The speaker emphasizes that Notebook LM overall is an underrated tool for transforming any content into slide decks, podcasts, flashcards, quizzes, and infographics.

  Claude vs. ChatGPT cost comparison: Claude is more expensive and has lower limits than ChatGPT, but costs less on the premium plan ($100–200/month) than via API. OpenAI has ~900 million users (many free), Anthropic ~30–50 million, but generate similar revenue – different business models (mass vs. premium).

  UK government AI lab: £40 million over six years for a new AI research lab – the speaker criticizes this as insufficient given billion-dollar investments in the US and China.

  Additional tips: Vibe-coding with Antigravity (free) or Google AI Studio is possible; Skills and MCPs should be derived from concrete problems, not the other way around. An MIT study showed that top-down AI pilot projects usually fail, but grassroots training for regular employees succeeds – therefore significant business potential for AI workshops and training.

  The speaker announced a webinar at 6 PM (UK time) about the opportunity to deliver workshops for businesses and earn hundreds to thousands of euros per hour.

  AI tools/providers: Anthropic, OpenAI, Google (Workspace CLI, Notebook LM, Gemini), Claude, Antigravity, GitHub, MCP — Video format: News update/roundup with opinion

• AI Is Now Working While You Sleep
  5.3.2026, 19:45:03

  Summary: The turning point in autonomous AI (December 2025)

  The video documents a fundamental shift in AI capabilities: instead of iterative prompting by humans, modern models can now work autonomously over extended periods. Claude Opus 3.5 Sonnet completes software tasks only 50% of the time, while Claude Opus 4.6 can work autonomously for 12+ hours. With orchestrated swarms (like the C-compiler project from Anthropic Labs with 16 agents), autonomous work functions over two weeks – the project cost $10,000 in credits and resulted in a compiled C-compiler, including tests and deployment.

  Longer time horizons enable more complex tasks, as models can retain more context and handle it more efficiently. Claude Code on the $200 plan is described as a major unlock: expensive, but productive enough for business application. December 2025 marks a turning point – OpenClaw was released and is now (with ~260,000 GitHub stars) the most starred open-source project of all time, far ahead of PyTorch and VS Code. The commit trend shows explosive activity: Claude Code generates 140,000 commits daily in public repos, developers like Peter Steinberger (OpenClaw creator) make thousands of commits per day – impossible for humans, but standard for AI-driven development.

  To get started, two tools are mentioned: Cowork (desktop app for Windows/macOS, Claude-based, with scheduling) and Qwen (mentioned, but not explained). Cowork runs slowly, but is suitable for tasks scheduled overnight – the AI controls browser windows while the user sleeps.

  Claude Opus 4.6, OpenClaw and Anthropic/OpenAI in focus; opinion/reflection with practical tips.

• The Agentic Era Is Officially Here & GPT 5.3 “less cringe” Drops
  5.3.2026, 07:15:03

  Summary: GPT models and the AI agents revolution (December 2025)

  The speaker discusses two recent OpenAI updates and positions them as a smokescreen given reputation-damaging events. GPT-5.3 Instant is a minor fine-tuning of 5.2, meant to be less “cringe” – OpenAI markets it as “more accurate, less cringe” (fewer unnecessary reassurances, fewer dramatic reactions). 5.2 had patronizingly reassured users, even for trivial questions. The bigger announcement is GPT-5.4, whose capitalization of the “T” suggests a reasoning-model type (like o1/o3). Rumors point to a release this week – the speaker suspects OpenAI is distracting from problems (Pentagon contact, mass exodus to Claude, 400% higher cancellations, 775% more 1-star reviews).

  The bigger picture: December 2025 as a turning point. Ethan Mollick and Andrej Karpathy identify four major AI shifts: (1) GPT-3.5 (Nov 2022), (2) GPT-4 (spring 2023), (3) reasoning models like o3 (2025), (4) functioning agent systems (December 2025). Before December, AI agents were poor; since December, they can handle autonomous multi-hour tasks. Andrej Karpathy: “Coding agents basically did not work before December and basically work since” – a hard cut. Models now maintain context over longer sessions, tackle complex projects independently (example: 30-minute video dashboard setup, which used to be a weekend project).

  Practical shift: Instead of “chat with AI, get answer” → “give AI task, it works for hours autonomously”. The speaker uses Claude Code for accounting tasks: agent accesses Gmail, collects invoices, uploads them to services – only monitoring required, not manual work.

  Access: Claude Cowork (recommended for beginners) or OpenClaw (free, technical). Cowork runs on desktop (Windows/Mac), automates browser tasks, can run overnight. OpenClaw requires VPS setup (e.g., Hetzner, €5/month) or Mac Mini. Also Claude Code and forthcoming tools from OpenAI.

  Cost issue: Claude Max plan costs $200/month – high ROI for businesses, difficult for individuals. Anthropic experienced server outages from the mass exodus from ChatGPT.

  Side issue – Anthropic/Pentagon: The speaker clarifies: Anthropic didn’t reject AI deployment in warfare, but only US domestic mass surveillance and autonomous weapons that could harm Americans. They work on reliability for operations against other countries – not morally superior, just pragmatically different.

  Claude (Anthropic), ChatGPT/Codex (OpenAI), Qwen (China), OpenClaw, Cowork, DeepSeek; demo/live discussion with preview of an upcoming tutorial webinar (business model: AI training for businesses).

• Can Anthropic Actually Survive What’s Coming?
  5.3.2026, 07:00:48

  The video discusses Anthropic’s current growth crisis: the company is experiencing its best growth while simultaneously risking its existence through confrontation with the US government. A move to Europe is not practical, as it’s not just about talent but the entire infrastructure – GPUs from Nvidia (which could suffer under US sanctions), cloud computing via AWS, and venture capital. The three largest funding rounds in history all took place in recent years in America (Anthropic $30 billion, OpenAI $40 billion, then $110 billion); Europe provides only hundreds of millions – insufficient to remain competitive against American and Chinese labs. Energy infrastructure is also a European weakness. Regarding Claude agent functions: Claude has no dedicated agent mode like ChatGPT; instead, users rely on Claude Code (for hosted agents) and Cline (a desktop app that can control the computer – move files, navigate websites, downloads). Cline recently added Scheduled Tasks, Unified Memory, and remote capabilities and is suitable for recurring tasks like invoice management. Amazon already has its own AI models (Nova) for AWS customers but doesn’t need to invest in OpenAI. The video also mentions that other providers (Perplexity, Manus) are pushing into the agent space, but Anthropic’s ecosystem (Claude Code + Cline) is currently leading.

  Demo/discussion video with critical analysis of Claude, Anthropic, and agent tools (Cline, Claude Code, Nova).

• Anthropic Fights Pentagon, ChatGPT Uninstalls Surge, OpenAI Raises Astronomical $110B
  4.3.2026, 06:30:09

  Summary: The Anthropic-Pentagon crisis and its impact

  The creator addresses a massive conflict between Anthropic and the Pentagon: The US Department of Defense under Pete Hegseth designated Anthropic as a “supply chain risk” after the company refused to accept two ethical boundaries – the use of Claude in autonomous weapons that could endanger soldiers and in massive surveillance of US citizens. All four major AI labs (Google, xAI, OpenAI, Anthropic) received approximately $200 million each in Pentagon contracts in 2024; only Anthropic gained access to classified environments.

  Anthropic declined and risked economic destruction. Trump chimed in, Hegseth threatened a Friday 5 PM deadline. Anthropic sued against the designation and argues that it’s legally unprecedented and unsound – normally it only affects foreign companies. Hundreds of tech executives (55 CEOs, including 47 from OpenAI) signed an open letter against the designation; Sam Altman defended Dario Amodei but didn’t sign himself.

  Turning point: OpenAI immediately took over the classified contract that Anthropic lost. This led to massive backlash: ChatGPT uninstalls surged by 295%, one-star reviews by 775% in a single day. People switched to Claude – which experienced its best week in history. Claude rose to #1 on the US App Store.

  Problem: Claude’s servers collapsed under the load of new users, which discouraged potential new customers. The creator emphasizes: Anthropic is not “the good guys” – they also worked with the Pentagon on drones and surveillance, just with ethical boundaries (no deployment against Americans). The damage to Anthropic is commercially massive, even if it wins in court – investments, chip supplies (Nvidia), and cloud infrastructure (AWS) could disappear.

  Anthropic quickly activated a memory import from ChatGPT to Claude (also on the free plan) and expanded Claude Code and Cowork with Scheduled Tasks to build an OpenAI-like ecosystem.

  Side issue: OpenAI’s $110 billion investment round is the largest private venture capital series ever – Sam Altman needs market stability, not chaos from Anthropic’s collapse. Sam Altman and other top executives signed the solidarity letter, but the conflict remains unresolved.

  The creator warns: a move to Europe is unrealistic – Anthropic would need GPU access (Nvidia), cloud infrastructure, venture capital, and power supply that Europe doesn’t provide. A six-month ceasefire could enable negotiations (like with TikTok ban threats).

  In Q&A: tips on Claude vs. Gemini (Claude better, thanks to ecosystem like Claude Code/Cowork); webinar announcement about AI workshop business models (for non-technical audiences) at ai-mit-kyle.com/webinar today 6 PM UK time; learning resource recommendations (Anthropic Academy, DeepLearning.AI).

  AI providers/tools covered: Claude, OpenAI/ChatGPT, Google, xAI, Gemini, Claude Code, Cowork, Lovable, Codex, Perplexity, Anthropic. Format: Opinion/reflection with news update and live Q&A.

• Why Sam Altman Is Defending His Biggest Rival
  3.3.2026, 17:30:15

  The transcript discusses the agreement between OpenAI and the US Department of Defense as well as the contrast with Anthropic, which has not completed such a deal.

  Regarding OpenAI’s support for Anthropic: Sam Altman has publicly commented on Dario Amodei and Anthropic, even though the two leaders don’t stand close personally (an anecdote mentions they refused to hold hands at an AI summit in Delhi). This could be PR, but also strategically necessary: if Anthropic fails, market confidence in AI investments overall could collapse – which also harms OpenAI, since both depend on external investment.

  Why OpenAI reached a deal while Anthropic didn’t: OpenAI argues that their agreement even offers stronger security guarantees than Anthropic’s original – such as cloud-only deployment and enforceable red lines. One possible explanation for failure at Anthropic: Dario Amodei may be less diplomatic than Sam Altman and damaged the relationship with the Defense Department.

  On concerns like drones and mass surveillance: OpenAI contractually assures and through their security architecture ensures that models cannot be used for autonomous weapons or surveillance of US citizens. OpenAI personnel are supposed to serve as an additional control mechanism.

  On the secret model theory: The government doesn’t receive better models than the public – they likely use older versions like Opus 4.5 or 4.6, as everything must undergo security reviews. Labs cannot function economically if they only provide their best models to one customer; they must continuously deploy more broadly.

  Critical point on contract design: A viewer raises that OpenAI’s contract may be strongly worded but not enforceable, while Anthropic’s potentially is more narrowly and less easily circumvented.

  The video primarily covers OpenAI and Anthropic; no specific focus on other AI tools.

• US Military Used Claude to Target Iran Despite Trump’s Ban
  3.3.2026, 06:15:00

  Summary: Anthropic vs. Trump Administration – AI geopolitics and fallout

  Core conflict: Donald Trump issued an executive order requiring all US federal agencies to immediately cease using Anthropic’s technology, with a six-month transition period for the Department of Defense. The background is a dispute over two “red lines” that Anthropic refuses to abandon: a ban on domestic mass surveillance and on fully autonomous weapons systems not yet proven safe. The Pentagon threatened to remove Anthropic from the contract and classify them as a “supply-chain risk” – a designation normally used only for hostile nations like China and Russia – and to invoke the Defense Production Act to seize Anthropic’s assets.

  Military use and contradiction: Ironically, the US military used Claude models for strikes on Iran – hours after the ban. Claude was used for intelligence tasks, target selection, and simulations. This shows the operational dependence and underscores the contradiction between Trump’s rhetoric and practical reality.

  OpenAI fills the gap: Hours after Anthropic’s refusal, OpenAI announced an agreement with the Pentagon for AI systems in classified environments – with its own three red lines (no mass domestic surveillance, no autonomous weapons systems, no high-level automated decisions). OpenAI emphasizes cloud-only deployment and retained control over the safety stack, plus staff in the loop. In FAQs, OpenAI admits the government doesn’t have access to better models than the public – only potentially older, more thoroughly reviewed versions.

  No secret super-model: A key point: the Pentagon is not using Claude Opus 5 or a mysterious AGI. It has access to similar models as everyone else, because the economics of model development make it impossible for Anthropic and OpenAI to maintain a separate high-end model for just one agency.

  Claude boom and infrastructure collapse: The political dispute has made Claude #1 in US app stores – the first massive download spike in its history. Anthropic offers a memory importer tool to migrate ChatGPT users to Claude. However: Claude is currently down – triggered by server load from millions of new users. A PR disaster at a critical moment, as first-time users encounter a non-functioning service.

  Broader context: Anthropic couldn’t simply move to Europe – without US infrastructure, Nvidia GPUs, cloud access, and venture capital, it would be economically dead. Anthropic’s collapse could destabilize the entire AI investment ecosystem. OpenAI also publicly supports Anthropic to maintain market stability, not purely out of solidarity.

  Ambiguity of “safety”: Anthropic is often portrayed as the “ethical” choice, but opposes mass surveillance and unsafe autonomous weapons – not warfare in general. They signed a $200 million Pentagon contract. OpenAI contracts contain wiggle room: laws and department policy could change to circumvent restrictive clauses.

  Conclusion: A dispute between a US administration and an AI lab, in which Anthropic shows unexpected steadfastness, while OpenAI gains the upper hand through pragmatic compromise – with broader impact on market stability, user migration, and military AI integration.

  Tools/providers: Claude, Anthropic, OpenAI, ChatGPT, Trump Administration, Pentagon, Claude Cowork, Open Claw / Format: Opinion/reflection with news update elements, live discussion

• Anthropic Just Said No to the Pentagon (Here’s What Happens Next)
  3.3.2026, 00:30:32

  The Pentagon gave four AI companies – Anthropic, OpenAI, xAI, and Google – approximately $200 million each for contracts with the Department of Defense on AI projects. Anthropic has opposed two specific military applications and set two “red lines”: (1) the use of fully autonomous weapons that independently select and eliminate people, and (2) mass surveillance of the American population using AI. The Pentagon subsequently threatened to remove Anthropic from the contract and classify them as a “supply chain risk” – a designation normally used only for hostile nations like China and Russia – and to invoke the Defense Production Act to seize Anthropic’s assets.

  Anthropic CEO Dario Amodei published a statement in which he first emphasizes Anthropic’s patriotic stance (use against China, secure foreign espionage) and explains that the two application restrictions are necessary because they would endanger democratic values. On autonomous weapons, Amodei argues that AI systems are not yet reliable enough for this purpose and that people would be harmed; Anthropic offers research and development work instead. He makes clear that Anthropic will not capitulate to the Pentagon despite all threats and will not change the conditions, but ends with willingness to cooperate in a smooth transition to another provider.

  This is a historic incident: a private American AI company openly confronts the US military, while the Pentagon uses extreme means (supply chain risk classification, wartime measures) to apply pressure – an unprecedented confrontation between private technology and government.

  Providers covered: Anthropic (Claude), OpenAI, Google, xAI; Format: News update (discussion of a statement with political analysis).

• Anthropic Defies Pentagon, Block Cuts 4,000 Jobs (Blaming AI), Gemini Imagen 3 Drops
  2.3.2026, 05:00:37

  Anthropic vs. Pentagon

  Dario Amodei, CEO of Anthropic, published a statement rejecting the demands of the Defense Department. The Pentagon had arranged contract-based collaboration with four AI companies in the summer (Anthropic, OpenAI, xAI, Google), approximately $200 million each. Defense Secretary Pete Hegseth gave Anthropic an ultimatum until Friday: Claude models must abandon their security guidelines or the company will be removed from the supply chain – a step normally reserved only for adversaries like Huawei. Anthropic established two red lines: no autonomous weapons systems that eliminate people without human control, and no mass surveillance of US citizens domestically. Amodei reaffirms both points in his statement, emphasizes Anthropic’s patriotism and national security concerns, and rejects the demands. He signals the company is prepared to leave if necessary, potentially making Claude unavailable – an unprecedented confrontation between a private AI company and the US government.

  Block job cuts through AI

  Jack Dorsey, founder of Block (formerly Square with Cash App, Tidal), announced layoffs of almost 40% of the workforce – from over 10,000 to just under 6,000 employees. This is the first prominent example of job cuts due to AI automation at a profitable growth company. Block is not in trouble; Dorsey justifies the move by saying that smaller, AI-powered teams can do the same work more efficiently. Packages are generous (20 weeks of salary, health insurance, stock options), and the market responded with approximately 25% stock gains – a signal that this is perceived as genuine productivity improvement through AI, not an emergency measure.

  Nano Banana Pro 2 update

  Google released Nano Banana Pro 2. New capabilities: consistent faces and up to 14 objects across multiple frames (previously multi-image generations were inconsistent), better handling of complex visuals, and improved text-to-image integration. The model uses a two-stage approach: it first queries its language model layer to generate thoughtful instructions before starting image generation – hence its strength with infographics and structured visual content.

  Claude new features

  Anthropic released three major updates, partly in response to OpenAI tools:

  • Auto Memory: Claude now automatically saves information about the user and project context across multiple sessions – previously a major ChatGPT advantage that kept many users on that platform.
  • Remote Control: Projects can start at the terminal and then continue on mobile devices via the Claude app, mimicking OpenAI Claude functionality.
  • Scheduled Tasks: Automated, recurring tasks (morning briefings, weekly reports) directly in Claude Code without external cron job configuration.

  These features bring Claude Code closer to OpenAI Claude experience, but more securely and within ToS.

  Cybersecurity risks from Claude Code

  Two incidents became public this week: a developer used Claude Code to hack his DJI vacuum cleaner operating system to control it with a game controller – Claude found a critical security vulnerability and gave him access to 7,000 robots worldwide. In parallel, a Spanish-speaking hacker with Claude’s help stole 150 GB of Mexican government data (195 million tax records, voter lists, government login credentials). This demonstrates Claude Code’s power in penetrating security structures.

  Explicitly mentioned AI tools/providers: Anthropic/Claude (including Claude Code, Opus 4.6), OpenAI/ChatGPT, Google/Gemini/Nano Banana Pro 2, xAI/Grok, Perplexity, Minimax, DeepSeek, Kimi, Lovable, Cursor, Open Claude, Cline, Anti-Gravity, Cogwork — Format: News update/live roundup with opinion component.

Leon van Zyl (1 new video)

• The First AI Agent Everyone Should Build
  2.3.2026, 13:01:26

  Summary: Build an AI assistant with n8n

  The creator shows how to build a personal AI assistant from scratch using n8n (free open-source workflow automation) – no coding experience required.

  Core features of the assistant:

  The agent has two types of memory: short-term memory (the last 20 messages of a conversation) and long-term memory (persistent data tables for user preferences). This allows the agent to remember user preferences even in new sessions – for example, if you say “my favorite pizza,” it gets saved and later considered in restaurant recommendations.

  Setup and access to n8n:

  There are three options: local installation via npm command (free), n8n Cloud service (~24 euros/month), or self-hosting on a VPS like Hostinger (from 5 dollars/month with coupon “lion”).

  Workflow architecture:

  Start with a chat message trigger, then an AI Agent node (as the “brain” of the system). The agent is connected via Open Router to Claude Sonnet 3.5. A system prompt defines the personality (e.g., name “Luna,” warm tone). Then three tool categories follow: memory tools (Get/Insert/Update via data tables), research tools (Google Search via SURP API, Wikipedia, Hacker News), and chat tools.

  Practical demo:

  The agent can research current information in response to queries, update personal data, and analyze images. The chat can be made publicly available (with username/password protection) and supports file uploads.

  Tools covered: n8n, Open Router, Claude Sonnet 3.5, Anthropic, SURP API – Tutorial, beginner-friendly.

Liam Ottley (3 new videos)

• What Is an AI Operating System? (And Why Every Business Will Need One)
  8.3.2026, 02:05:42

  AI Operating System (AIOS) – Methodology and Architecture

  The video explains that an AI Operating System is not a business model, but a business methodology – an approach to business management that functions as a layered wrapper around the existing business model.

  Core AIOS structure (layer model):

  1. Context Layer: Store central context information about the company (team, products, values, strategy) in the Claude Code workspace
  2. Data Layer: Connect data sources (Google Sheets, Stripe, Bitly, etc.) at a central location to gain quantitative insights on target deviations
  3. Intelligence Layer: Integrate meeting recordings (via Fireflies) and Slack messaging into a searchable database, combined into an automated Daily Brief (5–10-page PDF with KPI overview, business story, strategic insights)
  4. Automation Layer: Conduct a task audit to identify which tasks can be automated or heavily augmented; goal: take 60–70% of tasks off your plate
  5. Build Layer: Use freed-up bandwidth for new initiatives (or enjoy your free time)

  Core value: Escape the “operator trap” – reduce time spent in the business (operational activities) from 80% to 20% and free up 80% for strategic initiatives (new markets, products, channels, campaigns).

  Practical technique: Claude Code is configured as a central AI agent that accesses a local workspace with business context, database, and workflows. Non-technical users can use an /explore command to develop automations step-by-step; pre-configured modules can be installed like plugins.

  Three KPIs for AIOS practitioners:

  • Away-from-Desk Autonomy: How much work can be completed remotely (phone via Telegram)?
  • Task Automation Percentage: What percentage of original tasks is automated/augmented?
  • Revenue per Employee: Efficiency metric to measure AI ROI

  The speaker reports on a 7-day project (webinar launch generating 1 million NZD in revenue) that would normally have taken 1–3 months, but was completed in one week through freed-up bandwidth.

  Tools/providers: Claude/Claude Code (Anthropic), Fireflies, Slack, Telegram, Stripe, Bitly, Versell, Calendly — Deep-dive into a fully refined business workflow with Claude Code.

• The Biggest Shift in Business Since the Internet Just Happened
  5.3.2026, 04:10:12

  AI Operating Systems (AIOS) as Business Multiplier

  Founder Liam Ottley describes a system called “AI Operating System” (AIOS) – an AI layer built around the entire business with Claude Code that isn’t just a chatbot, but a context-aware system connected to real data.

  The AIOS operates in five layers:

  1. Context Layer: The AI learns the entire business completely – strategy, teams, processes, history – instead of having to re-explain everything with each conversation. Ottley has all four companies, team structure, and priorities loaded in.

  2. Data Layer: Instead of checking seven or eight different dashboards (CRM, analytics, revenue platforms), everything flows to one place in real-time. Ottley checks business health like a weather app – at a glance, revenue, leads, content performance, traffic.

  3. Intelligence Layer: Every day before getting up, Ottley receives an automatic briefing via Telegram with the most important events of the last 24 hours – revenue changes, team updates, meeting summaries from calls he didn’t even attend, content ideas, and a complete SWOT analysis. The AI connects dots across all business units.

  4. Automation Layer: This involves cataloging all recurring tasks and then systematically automating them. Example from Ottley’s agency: The system pulls client call transcripts, applies the scoping methodology, and generates finished proposal decks – a normally time-consuming task is now fully automated.

  5. Build Layer: The freed-up bandwidth allows the founder to work strategically again. Ottley describes how he went from webinar idea to over one million NZD in revenue in seven days: YouTube comments analyzed (days 1-2), five videos planned and filmed (days 3-4), webinar structure, slides, email sequences, funnel, payment systems built (day 5), webinar delivered (day 6), bookings generated (day 7) – each step accelerated by the system.

  The system can be controlled from an Apple phone via Telegram and manages over 60 employees across four companies. Ottley emphasizes this isn’t theoretical – it works with tools already available today, and hundreds of founders in his program are building the same system.

  Claude Code was explicitly named as the tool with which the AIOS is built – demo of a practical implementation with concrete examples (Telegram briefing, data dashboard, proposals generator, webinar launch in 7 days).

• Your Phone Can Now Run Your Entire Business. Here’s How It Runs Mine.
  2.3.2026, 14:44:46

  The speaker demonstrates how to use Claude Code Remote Control – a newly released feature for running Claude Code from your smartphone – to build an AI Operating System (AIOS) for your business. He manages four companies with over 60 employees on mobile and demonstrates three active tasks: a complete marketing audit, a lead magnet for the video, and analysis of business meetings from a month.

  The central concept is an AIOS – not a standalone business model, but an abstraction layer around existing business operations. The architecture follows a layered model: First, a Context OS that gives companies foundational knowledge (who we are, our services, team, values). Then a Data OS that collects real data from various sources (Google Sheets, P&L, YouTube Analytics, Google Analytics, Calendly, etc.) in a central database. Next, an Intel OS that captures meeting notes from tools like Fireflies or Fathom daily. The speaker has a briefing report generated daily that analyzes all business data, summarizes meetings (sometimes 70+), and provides strategic recommendations.

  The demo shows generated PDFs with marketing reports, call analyses, and a newly created landing page as a lead magnet. The key point: Claude Code Remote Control itself is just a remote control – without a built AIOS in the background, it’s useless. The speaker offers his Telegram-based plug-and-play system for free to help other founders get started, and points to a 2-hour webinar for a complete overview of his methodology.

  Claude (Anthropic), demo + opinion/reflection.

Mark Kashef (3 new videos)

• You’ve Never Used Claude Code Skills Like This
  6.3.2026, 18:02:49

  Summary: Multi-Model Council Skill for Claude Code

  Instead of committing to a single language model, this video demonstrates how to use a skill to combine the strengths of multiple models simultaneously. The approach is called “Council” — a skill that leverages Claude Code as the main driver but accesses other models like Gemini and Codex via Open Router.

  The Council Skill operates through slash commands and trigger phrases like “get a second opinion” or “ask Gemini”. The central component is a configuration file where you specify which model is ideal for which task: Codex for bug fixes, Gemini for frontend reviews, Gemini Flash for quick checks. When passing context to other models, the specific problem statement, relevant code snippets, and previously attempted solutions are included — so the models receive complete information for a holistic opinion.

  After consulting the various models, Claude Code presents the raw responses with model names, signals agreement or disagreement, and then selects the most practical approach, regardless of which model suggested it. In the demo, two landing pages for a bootcamp are created. The first is built without external opinions, the second uses the Council Skill — Gemini reviews the frontend, Codex the copywriting. The result: the audited and improved second version looks significantly more professional with better typography, grid layout, and more detailed pricing models. A second scenario with an analytics dashboard shows the same pattern: external consultation leads to a Bloomberg Terminal-like UI that looks infinitely better than the original version.

  The skill can also be applied to non-technical scenarios and can even compare different versions of the same model (Haiku, Sonnet, Opus). The synthesis rules are central: Claude Code remains in control and decides which suggestions make sense, rather than implementing them blindly.

  Explicit tools/models: Claude Code, Gemini 3.1 Pro, Gemini 3 Flash, Codex, Open Router. Format: Tutorial/Demo. Difficulty level: Intermediate (configuration file editing required, but guidance for non-technical users available).

• Plan Like a Pro in Claude Code
  4.3.2026, 21:00:04

  ASCII Art for visual planning with Claude Code

  The video demonstrates a technique using ASCII art diagrams to save time and tokens, whether building apps, planning slide decks, or visualizing systems — and it works equally well in Claude Code for both technical and non-technical users.

  Core idea: Instead of blindly writing code and rewriting it later, you first create an ASCII art visualization as a blueprint. This avoids multiple iterations and saves tokens, especially for token-intensive tasks like PowerPoint generation.

  Practical examples:

  • SaaS Dashboards: Create ASCII wireframe first (sidebar, stat cards, charts, data table), then refine (e.g., change chart width, define colors), then build as a React app. Result: design matches expectations, no surprising layouts.
  • Landing Pages: Visualize wireframe with navbar, hero, social proof, features, footer, optionally add pricing section, then build. ASCII art prevents poorly vibed, inelegant outputs without planning.
  • PowerPoint Decks: Define 10-slide structure as ASCII upfront (slide titles, problem, solution, features, etc.), then implement in one pass with PowerPoint Skill. Saves massive tokens since only one final iteration is needed instead of five to seven.
  • Databases: ASCII diagram for tables (Users, Products, Purchases, Audit Logs) with primary keys and relationships shows structure clearly — even for non-technical users. Prevents faulty assumptions about table relationships.

  Effect: The user offloads design thinking upfront, Claude Code then focuses on quality rather than structural decisions. Iterations become more precise, tokens decrease, especially for visual projects.

  Offering: Prompt templates available in the video description, advanced resources (ASCII art skills) in a community.

  Claude Code is featured; the video is a tutorial-demo hybrid with practical use cases.

• I Tried Claude Code Remote Control So You Don’t Have To
  1.3.2026, 17:30:09

  The video covers Anthropic’s newly released Remote Control feature for Claude Code and addresses three aspects: First, setup is explained — you update Claude Code to the latest version (via claude update or Warp Agent Mode), then use /remote in the terminal and can either click a link or scan a QR code to mirror your terminal session from a mobile device or another computer. The feature secures only outgoing HTTPS requests and doesn’t export credentials.

  Second, it explains how Remote Control differs from a custom-built Claude Code operating system: Remote Control is limited to a specific project and terminal and is ideal for monitoring long agent runs or avoiding surprises with permissions. The custom-built system, by contrast, enables navigation between different directories and embodies more of a personal assistant with continuous context — both could complement each other.

  Third, it practically demonstrates how to prototype a light personal-assistant variant with Remote Control: you create a memory file (for storing communication preferences), a soul file (for characterization), and a claude.mmd file (for instructions), selectively import skills like PowerPoint and CSV, and then enable Remote Control for this context. The result is a portable, context-aware tool you can use on mobile.

  Anthropic focuses on enterprise and developer productivity, not consumer products — Remote Control essentially replaces remote desktop apps for developers.

  Claude (Anthropic) was explicitly featured; format: Tutorial/Demo with opinion elements.

Matt Pocock (1 new video)

• The 7 phases of AI-driven development
  3.3.2026, 22:47:36

  Summary: Seven Phases of Development with AI

  The creator introduces seven phases to follow when developing with an AI-coding assistant (in his case, Claude) in order to deliver high-quality, maintainable software.

  Phase 1 – Idea: The initial problem is defined – whether it’s a complete app, a single feature, a bug fix, or a refactor. The idea can be as large or small as needed.

  Phase 2 – Research (optional): If the implementation requires external dependencies or difficult exploration (e.g., API integration), research is conducted and cached in a research.md file so the agent can access it without wasting context window.

  Phase 3 – Prototyping: The creator iterates with the AI agent in a throwaway prototype to make aesthetic and architectural decisions before final implementation begins. This is essential to establish personal design preferences early on.

  Phase 4 – PRD (Product Requirements Document): Based on research and the prototype, a document is written describing the end state – what the user sees and how the system behaves. The agent helps critically evaluate each design decision.

  Phase 5 – Kanban Board: The PRD is broken down into a Kanban board with tickets and blocking relationships. This enables parallelization instead of sequential execution.

  Phase 6 – Execution Loop: A coding agent works through the Kanban tickets – either sequentially or parallelized with multiple agents. The creator typically uses Ralph Loops for this.

  Phase 7 – QA: The agent creates a QA plan, a human reviews the results, reads the code, and creates new tickets in the Kanban. Then execution and QA repeat until the product is perfect.

  The creator emphasizes that these phases apply universally across different implementation approaches (Ralph Loops, GSD, SpecKit) and that this process represents an “Engineering First” approach to AI-driven development – not for “vibe coders,” but for serious AI engineers.

  Claude is mentioned, focus on personal methodology and processes (opinion/reflection + guidance); tutorial/deep-dive format.

Melvynx (6 new videos)

• GPT 5.4 : LE PIRE MODELE DE DESIGN AU Monde (Review de GPT 5.4 VS Opus 4.6)
  6.3.2026, 07:15:56

  The video covers the new GPT-5.4 model from OpenAI, available in Codex and the API. The speaker tests it against Claude Opus on multiple practical tasks.

  Key features of GPT-5.4:

  • Support for 1 million token context window
  • Better performance in knowledge work, web search, and computer use
  • Available in Codex (CLI) and API

  Key findings from testing:

  1. YouTube thumbnail generator: GPT-5.4 tends to go beyond requirements and generate “slop” (overly complex, unwanted design), while Opus respects minimalist specifications.
  1. SpongeBob 3D rendering: GPT-5.4 shows improved shaders and logic, though SpongeBob is damaged by a visual artifact.
  1. TimeZone visualization: GPT-5.4 creates an elaborate styled but less intuitive interface. Opus produces clean, functional code with better UX.
  1. Thumbnail editor (fixed click): Opus stays consistent with requirements; GPT-5.4 adds unsolicited changes to the interface.
  1. Codex CLI in bug fixes: The revised Codex CLI works faster with better color coding and more logical tools (read, search instead of shell commands). It solves problems efficiently but remains slower than direct approaches.

  Speaker’s conclusion: OpenAI models are intelligent and solve most tasks technically correctly, but have a weird relationship with creativity—they respect requirements excessively but simultaneously make unwanted changes. Opus is preferred for UI/UX tasks, while GPT-5.4 is stronger for engineering tasks.

  Explicitly mentioned tools/models: OpenAI (GPT-5.4), Claude Opus, Cursor, Codex CLI, Gemini-API — demo/review with practical comparison tests.

• Les Modèles IA Pour Dev En 2026 : VOICI ce que je Recommande (et aussi al Silicon Valley)
  5.3.2026, 18:21:30

  Which AI models should you use in 2024? – A practical analysis

  The video addresses which AI models developers and non-developers should currently use, based on experiences from San Francisco.

  Available models overview:

  • Chinese open-source models: Kimi 2.5, Zi GLM5, MiniMax M (affordable: ~$0.5-3 per million input tokens)
  • US models: Gemini (Pro/Flash), Claude Opus 3.5 Sonnet, GPT-4o, GPT-4 o1

  Intelligence & code quality:

  Opus 3.5 and OpenAI Codex are considered the best models among developers. Gemini ranks below them and has less capacity for complex problems.

  Different user types:

  • Non-technical users/less-technical coders: Claude is recommended—friendly tone, easier to use, straightforward. Opus 3.5 follows instructions precisely without overcomplicating.
  • Technically-focused developers: OpenAI Codex is better for specialized, complex tasks but less beginner-friendly.

  Budget strategies:

  • Small budget (20€/month): Use Claude for planning, then execute with GPT-4o Codex (since execution is expensive)
  • Large budget (100-200€/month): Maximize Opus—consume as many tokens as possible with Claude; GPT as fallback for errors
  • Open-source: Kimi 2.5 proved practical in tests (despite moderate benchmarks); GLM5 slower and lower code quality

  Two coding modes:

  • Pair programming: Precise technical prompts → Codex better (follows rules exactly)
  • Manager mode: General requirements without technical details → Claude better (more creative, autonomous, less restrictive about security concerns)

  Author’s personal conclusion:

  Claude is more pleasant to use despite Codex’s technical advantages. The difference lies less in raw intelligence than in user-friendliness, autonomy, and permissiveness. Claude respects workflows better and is less fixated on security rules.

  Bonus: The author announces a live event on Sunday where a complete app will be built with Claude and Cloud Code; participants receive free configurations and setup videos.

  Video covers Claude, OpenAI (GPT-4o, Codex), Google Gemini, and Chinese open-source models (Kimi 2.5, GLM5); format: opinion/reflection with practical recommendations.

• J’AI CRÉÉ L’APP PARFAITE “Speech to Text” en 1 HEURE avec Claude Code (et tu peux aussi le faire)
  4.3.2026, 07:00:16

  The video shows how to fork and adapt open-source projects with Claude Code to develop custom software. The creator forked the speech-to-text tool “Handy” to create “Parler”—a personalized alternative to paid solutions like Super Whisper ($15/month).

  The problem with existing tools: Commercial speech-to-text apps are expensive and already based on freely available open-source models (Whisper, Parquet, Moonshine). They limit features (like the 2000-word cap in free plans) and offer unnecessary premium features.

  The new strategy: Use or fork Handy and similar open-source projects directly. The creator shows how to clone a repository in Claude Code, rename it, and extend it independently—while maintaining MIT license requirements.

  Features of “Parler”: Automatic model-switching for longer recordings (from fast Parquet V3 to more accurate Whisper Turbo after 20 seconds), post-processing modes (email mode, tweet format, grammar fix) with keyboard shortcuts (Ctrl+1, Ctrl+2, etc.), rebranding with blue instead of pink UI. Everything was accomplished with “vibe coding”—intuitive, iterative development with Claude—in just a few hours.

  Why no pull request? The creator keeps their fork separate since not all features are relevant to everyone and they don’t want to discuss with maintainers. Their tool, their rules.

  Core message: With Claude Code, anyone can now quickly create niche software that meets personal needs—for free, based on existing open-source, without outside permission.

  Claude Code, demo/tutorial — demo with practical application of the forking workflow and iterative vibe coding for personalized app development.

• OPUS 4.6 VS Opus 4.5 VS Sonnet 4.5 dans les CHOIX de techno : ils font n’importe quoi ?
  3.3.2026, 17:33:03

  The video covers research on Claude Code’s tool choices and their consequences for “vibe coders” (non-developers building applications with AI).

  The research tested Claude models (Sonnet 4.5, Opus 4.5, Opus 4) hundreds of times to determine which libraries and tools they would recommend for frontend, backend, etc. Result: AI models choose tools based on statistical frequency, not actual need. Claude’s default stack typically includes Vercel, PostgreSQL, Drizzle, Next.js, GitHub Actions, pnpm, Vite, Tailwind, and others—and these tools gain enormous traction in practice because of it. Companies like Vercel or Resend benefit massively since Claude recommends their tools daily, even when they’re outdated or suboptimal.

  Different models have different “personalities”: Sonnet picks conventional popular tools; Opus 4.5 is more balanced; Opus 4.6 prefers newer solutions and custom implementations. Notably, Claude recommends tools in monopoly positions (Stripe, Vercel, Chat CN), while enterprise typically uses different tools (Redis instead of custom, Prisma instead of Drizzle). The study also shows: newer model versions drastically change their preferences (Prisma dropped from 80% to 0% with Opus 4).

  The core problem: Those without technical expertise cannot verify whether AI-recommended tools are actually suitable—the AI effectively dictates the technology landscape. This leads to a “benchmark business” in San Francisco where companies try to get their tools deliberately placed in AI recommendations.

  Claude (Anthropic) was discussed as an AI model; format: opinion/deep dive.

• VIBE CODEUR : Formation pour apprendre à utiliser le terminal (c’est un cheat code)
  2.3.2026, 17:35:24

  Summary: Terminal tutorial for vibe coders

  The video is a tutorial on using the terminal as an essential tool for the vibe coder—a concept where mastering the terminal, VS Code, and Git matters more than specific knowledge in HTML, CSS, JavaScript, or frameworks.

  Basics and setup:

  Windows users are recommended to use WSL (Windows Subsystem for Linux) to use Bash instead of PowerShell, since Bash and ZSH on macOS and Linux contain the relevant commands. macOS users can use the pre-installed Terminal app directly. Alternative terminal tools like iTerm2 or Ghostty are mentioned.

  Core terminal concepts:

  The terminal displays the current path (e.g., ~ for the home directory). Important basic commands are:

  • ls – list files in current directory
  • pwd – show current path
  • cd [directory] – change directory, cd .. goes one level back, cd alone returns to root
  • mkdir [name] – create new folder
  • rm -rf [name] – delete directory
  • echo – output text

  Practical workflows:

  Node.js is recommended as an important tool for vibe coders (install via nodejs.org). With npm, projects can be created quickly, such as with npm create vite for a React and TypeScript project. The command npm run dev starts a local development server.

  VS Code can be opened with the command code . directly in the current directory from terminal (requires shell commands to be installed). The terminal can also be opened and used directly in VS Code.

  Practical tip:

  Right-clicking while holding Option in Finder allows copying the path and quickly navigating with cd + paste.

  The tutorial takes a practical approach, showing that the terminal is not an abstract interface but an alternative view of your computer’s file system navigation and management.

  Format: Tutorial (French transcript, no AI tools explicitly mentioned)

• 3 jours dans la vie d’un DÉVELOPPEUR à Séoul | Ep.1
  1.3.2026, 17:20:41

  Episode 1: Journey to San Francisco – layover in Seoul

  No transcript available – summarized by title and content:

  A Swiss content creator starts a new vlog series about his trip to San Francisco. After approximately 20 hours of flying (7h + 12h) with a layover, he first lands in Seoul, South Korea—the “semifinal destination” of his trip. He uses the time to explore the city: visits a shopping mall, shops at Uniqlo (he’s a self-described “Uniqlo maximalist”), tries Korean street food, and enjoys a $75 wagyu steak dinner. The episode also shows his first practical experience packing a suitcase—he usually travels only with a backpack. Simultaneously, he works in a creator space and uses Claude (via Telegram) for travel planning. His main motivation for relocating to San Francisco: the AI industry and tech companies that he admires. He’s leaving Bali after three years because the routine no longer fulfills him. With just a Mac and suitcase, he sets off for the final leg to San Francisco, where he has a coworking space and accommodation but otherwise no concrete plan.

  Format: Demo/vlog, tools mentioned: Claude, OpenAI, Gemini — difficulty level: Entertainment covering travel organization and personal reflection, no technical depth.

MoureDev by Brais Moure

No new videos in this period.

n8n

No new videos in this period.

Nate Herk | AI Automation (6 new videos)

• This New Claude Code Feature is a Game Changer
  7.3.2026, 18:03:00

  Cloud Code Loop Feature: Automated, project-specific tasks

  Cloud Code has received a new Loop function that enables recurring tasks to run automatically for up to 3 days without user intervention. With natural language commands like “Check my ClickUp every 10 minutes” or “Remind me at 10:23 AM to review the project,” the function can be activated. In the background, cron jobs are created that fire at regular intervals and automatically perform notifications or checks.

  The Loop function fundamentally differs from the concurrently introduced Scheduled Tasks feature: Loop runs only within an open session, doesn’t persist, and runs for a maximum of 3 days, but has no catchup behavior and is available everywhere in Cloud Code (Terminal, Desktop App, VS Code). Scheduled Tasks, by contrast, are long-lived, can run indefinitely, and have automatic catchup—but are currently only available in the Desktop App.

  Loop is controlled via three Cron tools: cron_create (for scheduling), cron_list (for listing), and cron_delete (for canceling). Practical use cases include short-term projects such as urgent email monitoring every 5 minutes, hourly deploy checks, 3-day sprints for progress tracking, or continuous log monitoring during an iteration.

  Cloud Code Loop, Demo/Roundup – without external tools or vendors explicitly mentioned.

• Claude Code 2.0 Is Finally Here
  7.3.2026, 05:06:53

  Summary: Scheduled Tasks in Claude Code

  Anthropic has extended Claude Code with native Scheduled Tasks that allow automations to be scheduled directly in the Desktop App. These function not as deterministic scripts, but as independent agentic workflows—they can fix errors themselves, read files in the project, and use all available tools. The workflows improve themselves over time rather than simply running the same script each time.

  Setup is straightforward: Either create a new task via the Schedule tab (select name, description, prompt, model, mode, folder, and set time interval) or type /schedule in the prompt. Existing Skills can be converted to automations via prompt—in the example, “morning coffee” could be converted to a daily 6 AM automation with a single message to Claude Code from an existing morning planning skill.

  Important limitations and considerations:

  • The computer and Desktop App must be running; without them, the automation won’t execute. Claude Code checks up to 7 days back for missed tasks and catches up on them.
  • Tasks are stateless—each run is a new session with no context from the previous execution.
  • Security: Permissions must be consciously configured (e.g., block bash commands for deletion), and tasks should be manually tested before automation.
  • Currently only available in the Desktop App (Terminal and IDE extensions coming soon).

  For continuous improvement, a strategy using a status file per task is recommended, which is read by the new agent instance before each run. It contains info like last run time, errors encountered, and known issues—allowing successive runs to have contextual continuity. The prompt can reference this file directly and update it upon completion. Notification options include hooks (e.g., for system sounds) or messages to external tools like ClickUp.

  Demo with Skill management: The project contains several skills (Content Creation, Research & Intelligence, Visual Diagrams, Operations, Meta)—all can be converted into weekly automations. Terminal/VS Code can view and edit existing task files but cannot create new tasks or change cron logic, as that only lives in the Desktop App.

  Claude Code (Anthropic) – Demo.

• Cursor Automations Clearly Explained (worth learning?)
  6.3.2026, 20:23:28

  Cursor Automations: Trigger-based cloud agents for code workflows

  The video covers the new Cursor Automations feature, which enables trigger-based cloud agents for automated code tasks. Contrary to the perception that every new AI coding tool must be learned, the creator emphasizes that fundamental concepts like planning, prompt design, tool orchestration, and evaluations are far more important than switching between tools.

  Cursor Automations in detail:

  Automations run in Cursor-managed cloud sandboxes and are triggered by various mechanisms—schedules (Cron), GitHub events, Slack messages, Linear issues, or custom webhooks. When triggered, a sandbox spins up, loads the repository, and executes the agent workflow. The system supports various models (Claude, GPT 4.5, Opus, Sonnet), storage, and MCP servers. Typical use cases include automated code reviews on every push, security checks, incident response, and scheduled maintenance.

  Comparison to OpenClaw and Cloud Code:

  OpenClaw is a personal OS assistant with system-level access to files, apps, and browser, running on your own machine. Cursor Automations, by contrast, is limited to code repositories and developer tooling. Both use heartbeats/crons—agents aren’t active 24/7 but wake at defined times. Cloud Code offers on-demand agent workflows in the IDE, but for automations you’d need to host them separately (e.g., via Trigger.dev). Cursor Automations consolidates this: unified UI, cloud hosting, simple on/off switching.

  Practical workflow suggestion:

  Use Cloud Code locally as a builder—design code, test it, then push to GitHub. Then deploy Cursor Automations as a “caretaker”—autonomous workflows that continuously perform security audits, reviews, or maintenance. This saves manual deployment to external services.

  Core message:

  The space is evolving from “ask AI for help” → “AI writes code” → “AI ships code” → “AI maintains and optimizes code autonomously.” More important than tool-hopping are transferable skills like prompt design and agent architecture.

  Demo and observation:

  The video shows a live automation: GitHub trigger starts an agent in the sandbox, which performs a code review and security review, creates a branch, and sends notifications—all without manual prompts after initial configuration.

  Covers: Cursor, Claude, GPT 4.5, OpenClaw, Cloud Code, Trigger.dev, GitHub Actions, N8N; Format: Demo + Opinion/Deep-dive.

• Claude Code Skills Just Got Even Better
  5.3.2026, 17:35:59

  The video explains why Skills (capabilities for AI agents) have become significantly better through Anthropic’s new Skill Creator Skill, and then demonstrates live creation of a new skill.

  What are Skills: Text-based recipes that instruct agents how to consistently perform specific tasks—such as LinkedIn posts or internal communications in a specific format.

  Two types of Skills:

  • Capability Uplift Skills: Prompt-based, teach Claude to do something better (e.g., website design with better fonts and color schemes). These could lose value with new model versions.
  • Encoded Preference Skills: Sequential workflows with specific steps (e.g., the demonstrated “Idea Mining” skill, which runs multiple agents in parallel). These remain more stable since they’re process-specific.

  The new Skill Creator Skill enables:

  • Skill creation: Agents can create new skills from scratch
  • Evals/Tests: Test skills against examples and automatically optimize—e.g., a PDF-filling skill that learns through evals to place text correctly
  • Benchmarking: Compare pass rate, time spent, and token usage (with/without skill)
  • Trigger tuning: Adjust skill descriptions so natural language commands invoke them more reliably

  Live demo – YouTube Weekly Roundup Skill:

  The creator asks the agent to produce a weekly PDF report about his videos (views, engagement, comments, SWAT analysis, competitor context) initially with only a vague description. The agent asks clarifying questions, creates a plan, iteratively tests with the Skill Creator, and improves based on feedback. The result: A professionally designed report with video statistics, executive summary, SWAT analysis, comment analysis, and trend data—all created in about 20 minutes. The creator can rate the output and further refine the skill.

  Future perspective: Anthropic hints that soon a natural language description (“What should the skill do?”) might be enough for the model to figure out technical details on its own—instead of currently still requiring explicit SOP, rules, and format specifications.

  Anthropic/Claude is featured; live demo of a self-created skill in Cloud Code – Demo format.

• Turn Claude Code Into Your Executive Assistant in 27 Mins
  5.3.2026, 12:46:03

  The video shows how to build a personal Executive Assistant with Claude Code that automates daily recurring tasks and runs multiple sub-agents in parallel. The demonstrator shows four agents running simultaneously: a Morning Coffee skill for daily planning based on calendar and priorities, a Research sub-agent for content (LinkedIn post and Twitter carousel), a Pulse Check for team status, and a Visualization agent. All of this completes in under two minutes—while manual execution would take about 25 minutes.

  The setup follows four phases: (1) Give it a home—install Visual Studio Code with the Claude Code extension, create a project directory, and set up a claude.md file as the central configuration file that Claude Code reads on every call, telling it where all information is stored. (2) Give it life—conduct an interview that automatically creates context files for “me” (profile), “work” (business), “team” (employees), and “priorities” (current goals); also creates a folder structure with templates, projects, decisions log, and references. (3) Give it hands—build concrete skills and sub-agents: the example shows a Research skill using the Perplexity API and creating reports, plus a cheaper Research agent using the Haiku model instead of Opus. (4) Let it grow—use the assistant daily, iteratively improve skills, migrate all custom GPTs and prompts as skills, and push the project to GitHub for version control and access from anywhere.

  Core point: The system works because claude.md is only a lean brain file that references specialized files, rather than loading everything inline—keeping token usage low while giving the agent access to complete contexts. YAML front matter in skills and agents is best practice. Access requires a paid plan (Claude subscription) and API keys for external services like Perplexity. The greatest strength: every generated artifact (research reports, content, decisions) stays in the project folder, making the assistant smarter over time.

  Claude Code, Perplexity API, Sub-agents, Skills, GitHub integration; Tutorial with practical setup.

• The NEW Nano Banana 2 + Claude Code = $10k Websites
  3.3.2026, 23:53:51

  Summary: Professional animated websites with CloudCode in 30 minutes

  The creator demonstrates a complete workflow for building an Apple-like product landing page with CloudCode in Visual Studio Code—from video generation through website creation to cloud deployment.

  Preparation & setup:

  Install the CloudCode extension in VS Code (requires Anthropic Pro/Max plan) and download two free markdown skills: a “Front-End Design Skill” and a “Video to Website Skill.” These skills define best practices for animated 3D websites and are key to the process.

  Video creation:

  Instead of programming actual 3D animations, use AI image generators (here: NanoBanana 2 via key.ai) to create start and end images of a product, then an AI video tool (Kling 3.0) to animate between these frames. The creator gives Claude both images and receives an optimized video prompt in return, which is then implemented in Kling.

  Website generation:

  Load the video into CloudCode and enter a simple prompt (e.g., “Create a modern product landing page for this video”). CloudCode works in plan mode and asks about product name, desired sections, and preferences, then creates a detailed plan and extracts ~120 WebP frames from the video via FFmpeg. Each frame is linked to a scroll position—as you scroll, frames are rendered sequentially like stop-motion. CloudCode writes HTML, CSS, and JavaScript and tests the website locally on localhost.

  Iteration:

  You can provide direct feedback (e.g., “Feature 2 appears too late”), CloudCode creates a new plan and fixes the issue. The skill improves with each iteration when you document your changes.

  Deployment:

  After local testing, connect CloudCode to GitHub (via CLI auth), CloudCode creates a new repository and pushes the code base. Import the repo into Vercel, which automatically deploys. Important: Video frames are initially excluded from the repo—you must instruct CloudCode to include them so the animation works on Vercel. After redeployment, the website is live on a real domain and responsive on mobile devices.

  Business potential:

  The creator suggests offering this as a service—build demo websites for local businesses or niches in ~1–2 days and sell for 5,000–10,000 USD, plus monthly hosting and maintenance as recurring revenue.

  Customization:

  You can customize the skill anytime if you don’t like the standard animation style. The core mechanism—video to frames to scroll-driven animation—is flexible and works with any video content (product deconstruction, text, motion, etc.).

  Explicitly featured tools: CloudCode (Visual Studio Code extension), Anthropic (Claude), GitHub, Vercel, NanoBanana 2, Kling 3.0, key.ai, FFmpeg — Format: Tutorial/Demo with in-depth practical workflow.

NeuralNine (2 new videos)

• Why I Go To Tech Meetups & You Should Too…
  6.3.2026, 17:00:25

  Summary: Meetups and networking

  The video is a casual, personal reflection on the importance of meetups and tech networking. The creator argues that attending meetups is valuable for everyone in the tech industry – whether you’re a founder (who can find co-founders, investors, or customers), job seeker, freelancer, or employee looking to expand your skills. The central idea is the concept of “Positional Chess”: you shouldn’t go to meetups with a concrete agenda (e.g., to acquire a client), but rather approach spaces with curiosity and openness where good things can happen – opportunities often emerge later and organically.

  As practical guidelines, he recommends going without expectations, being open to people and inspiration, and eventually giving talks yourself to be perceived as an authority and improve your networking chances. Specific channels include Meetup.com, Luma (for events), LinkedIn, X, as well as local clubs, communities, Discord, or WhatsApp groups. Core message: networking pays off regardless of whether you’re an employee, founder, or freelancer – often the benefits show up years later.

  The video is more opinion-based than tutorial or deep-dive and is based on personal experience (roughly 15 meetups attended).

• Coding A Linux Soundboard From Scratch in Python
  2.3.2026, 18:38:04

  Summary: Customizable soundboard for Linux

  The video shows how to build a Linux soundboard application that lets you play sounds through your system’s microphone via keyboard shortcuts – useful for Discord, video games, or recording tools like OBS.

  Core concepts:

  • PulseAudio & pactl: The project uses the pactl (PulseAudio control) command to manage sinks (sound collectors), sources (microphones, potentially virtual ones), and modules.
  • Setup components:
  • A virtual sink (module-null-sink) is created to centrally collect sounds.
  • A virtual microphone is created via module-remap-source that monitors the sink.
  • A module-loopback is loaded to feed the physical microphone back into the sink, making voice and sound effects available together.

  Implementation in Python:

  The script uses subprocess to automate pactl commands. On startup, it checks if the virtual devices already exist; if not, it creates them. The user then selects from available sources the input microphone to be looped into the sink.

  Keyboard input & permissions:

  The evdev package is used to access keyboard events. This requires special permissions (the input group and a udev rule), so the author provides a setup script. Control is via Super key + brackets (change selection) and apostrophe/semicolon (play/stop).

  Sound playback:

  The soundfile and sounddevice packages load WAV files from a sounds directory and play them through the virtual sound card – audible in any application using the virtual microphone as input.

  The project works as a functional prototype; the author plans a TUI version and used Claude for code implementation.

  Demo: The video shows the application live in Audacity and OBS, where the beep sounds are played back successfully.

  Claude was explicitly mentioned for code implementation; tutorial with practical demo (Linux/PulseAudio, intermediate level).

Nic Conley (1 new video)

• OpenClaw vs Claude Cowork | Setup, Limits & Which One Is Right For You
  6.3.2026, 16:03:30

  OpenClaw vs. Claude Code: Social Media AI Agent Showdown

  The creator builds a fully automated AI social media agent for TikTok with both systems – to test whether OpenClaw is really better or if Claude Code (with new Scheduled Tasks feature) can keep up.

  Setup & Preparation:

  Both agents get access to the “Larry Skill” (a framework for social media agents), the Replicate API (for Google Nano Banana Pro – image generation), and Post.is (social media scheduler with analytics). The app to be promoted: T-Max, an iOS app for natural testosterone optimization.

  OpenClaw Path:

  Deployed on a Hostinger VPS (KVM2 plan, ~$150 for 2 years via Nick10 code). Made accessible via Telegram. After ~35 minutes, the first complete TikTok slideshow is live – with before/after scenarios, AI-generated images, actual app screenshot, and text overlays. The agent saves instructions, self-resolves errors (wrong image resolution), automatically creates captions, and schedules autonomous posts for 9 AM with analytics tracking after 12 hours. Cost: ~$3–5 in API credits.

  Claude Code Path:

  Desktop app, no VPS needed. The agent performs 15 minutes of browser research on request (search TikTok, develop content strategy). Then automatically creates a slideshow and posts it – but without asking what the creator thinks first. Same screenshot upload error as OpenClaw, but the code doesn’t feel as autonomous: it asks the user to manually place the app screenshot in a folder instead of grabbing it from the chat like OpenClaw does. Code plan costs $20/month. Likely similar or slightly lower API costs (~$5).

  Scheduled Tasks – the Core Problem:

  OpenClaw posts the next morning at 9:02 AM completely autonomously. Claude Code tasks only run when the computer is active – not in true background. That’s the game-changer: OpenClaw feels like a real AI employee working at night. Code is more of a workflow tool that simulates automation, but doesn’t run autonomously.

  Bottom Line:

  OpenClaw wins for true autonomy and agentic behavior. Code is safer (local, no VPS risk) and faster to get started, but not a real agent – more of an automation application. The creator would choose OpenClaw again for 24/7 scenarios like this.

  Platforms & Tools: OpenClaw, Claude Code (Claude 3.5 Sonnet backend), Hostinger VPS, Replicate (Nano Banana Pro), Post.is, Telegram, TikTok API — Deep-dive tutorial/demo with direct comparison, beginner to intermediate difficulty.

Nick Saraev (2 new videos)

• AI Agents Full Course 2026: Master Agentic AI (2 Hours)
  8.3.2026, 05:01:31

  Summary: AI Agents – Comprehensive Course

  This course covers practical techniques for using AI agents in personal and business contexts, not theoretical foundations or specific platforms. The instructor runs a company that generates over $4 million per year using AI agents.

  Core Content

  Demonstration: Five parallel AI agents with their own Chrome browsers automatically fill contact forms on websites and perform economically valuable tasks.

  Agent-Workflow-Loop (consistently applicable): The loop consists of three steps – Observation (agent reads context, files, tool results), Reasoning (agent plans next steps), and Action (agent calls tools, processes files). The loop repeats until a “Definition of Done” is reached, then the result is output.

  Platforms: Codeex (OpenAI), Claude Code (Anthropic), and Anti-Gravity (Google/Gemini) are each briefly explained for setup and strengths/weaknesses – all are similarly powerful (differences only a few percentage points). Claude has the best interpretability, Gemini the best frontend design capabilities, GPT the best backend programming.

  Self-Modifying System Prompts (agents.mmd / claude.mmd / gemini.mmd): A file at the beginning of the context contains a growing list of learned rules. When the agent is corrected, it updates this file – in the next session it reads the rules and proactively avoids errors.

  Multi-Agent-MCP-Orchestration: A manager agent (e.g., Claude) delegates tasks to specialized agents (Gemini for frontend, Codeex for backend), combines the results.

  Video-to-Action-Pipelines: Gemini understands videos natively, extracts step-by-step instructions, and sends them to Claude, who implements these as real actions (e.g., building N8N workflows after YouTube tutorials).

  Stochastic Multi-Agent-Consensus: Multiple (e.g., 10) agents receive the same question with slightly varied prompts in parallel, all together exploring more of the “search space” of possible answers than a single agent – mode and median are calculated, consensus ideas plus outlier ideas emerge.

  Agent-Chat-Rooms: 5 agents with different personalities (e.g., systems thinker, pragmatist, skeptic) debate a problem in a shared chat.json, working out nuanced, high-quality solutions through dialogue rather than parallelization.

  Sub-Agent-Verification-Loops: One agent implements, a second agent reviews the output (with fresh context, no bias), a third agent corrects – achieving higher quality through unbiased evaluation.

  Prompt-Contracts: Before a task, a Skill forces the agent to generate a mini-contract (goal, constraints, output format, error conditions) – precisely defines “Definition of Done”.

  Reverse-Prompting: Agent asks the user 5 questions before implementing – surfaces hidden assumptions, improves one-shot rate.

  Context-Management & Token-Efficiency: Context windows are limited (200K–1M tokens), quality drops with more tokens. Instead of loading everything, agents use tools (read, grep, web fetch, bash) to selectively load information. YAML front-matter in Skills stores only metadata, not the full Skill. Iceberg technique: 10–30% for constantly accessible context (memory, local claude.md, active tasks), 70–90% as tool-based on-demand access.

  Model-Selection & Cost-Optimization: 60/30/10 rule – 60% of tasks with cheap models (Haiku, Flash), 30% with mid-tier (Sonnet), 10% with premium (Opus). Router agent (premium) delegates. Batch APIs offer discounted rates for delayed processing.

  Multi-Agent-Chrome-MCP: Orchestrator spawns 10+ parallel Chrome instances, each with its own MCP server and workspace, all agents fill forms/scrape websites in parallel – instead of 2–3 minutes per form, now dozens in parallel in the same time.

  Applied Platforms & Tools: Claude Code, Codeex (OpenAI), Anti-Gravity/Gemini, MCP (Model Context Protocol), N8N (Workflow builder), Skills as Repeatable Workflows — Format: Deep-dive tutorial with demos, made accessible for beginners without programming experience, but covers advanced architectures (orchestration, verification loops, context optimization).

• CLAUDE SKILLS FULL COURSE: Automate Your Work (2026)
  2.3.2026, 23:09:11

  Summary: Skills for AI Agents in Business

  The creator runs a business with over $4 million in annual revenue that he manages primarily through AI agents and so-called “Skills.” Skills are essentially the evolution of Standard Operating Procedures (SOPs) for AI agents – instead of checklists for humans, they are structured instructions for AI systems.

  Five practical business Skills the creator uses:

  1. Lead Follow-up Automation: Searches through all emails and conversation histories, identifies leads at various pipeline stages, and automatically sends personalized, casual-sounding follow-ups with context from previous conversations – all in one pass, instead of manually reaching out to each person.
  1. Thumbnail Generator: Generates multiple variations of a specific image composition (e.g., of a specific actor) through face recognition and realistic overlay to select the best variant.
  1. Lead Scraping from LinkedIn Sales Navigator: Automatically creates Google Sheets with filtered B2B leads (by job title, location, seniority level, industry) including email addresses directly from LinkedIn Sales Navigator within minutes with high deliverability.
  1. Cold Email Campaign Generator: Automates writing cold email campaigns based on customer details – duplicates best-performing existing campaigns, rewrites them for the new niche, and sets everything up in the email platform (previously 30–60 minutes, now minutes).
  1. Website Builder: Creates high-quality, personalized websites based on requirements, deploys them to Netlify, and enables sending them as lead magnets or outreach tools – particularly valuable because prospects typically don’t have such tools.

  Bonus Skills (personal tasks):

  • WeWork Booking: Automates daily desk bookings for the next 30 days through HTTP requests in the browser
  • Amazon Shopping Assistant: Compares products (e.g., dental floss) by price and effectiveness, creates personalized purchase recommendations

  How to Create Skills:

  Skills are organized as Markdown files (.md). The process: (1) Obtain the free skillspec.md from the link in the video description, which contains the complete Skill definition and structure. (2) Add this as claude.md to the workspace (automatically prepended). (3) Describe the desired SOP to the agent in natural language (voice or text). (4) The agent creates the skill.md with frontmatter (summarized metadata in three hyphens), detailed instructions, goals, and allowed tools. (5) Agent tests the Skill end-to-end. (6) If runtime issues occur, the agent repairs the Skill automatically.

  File System Structure:

  Skills are hidden under .claude/skills/[skill-name]/skill.md (visible on macOS with Shift+Cmd+.). Optional additional files such as Python scripts and API tokens can be stored in the folder.

  Important Concept – Skill Matching & Progressive Disclosure:

  The agent scans Skill descriptions and loads only the relevant skill.md into the context window when called. This saves tokens and costs – a safety mechanism that increases efficiency and output quality.

  Demo: Inbox-Cleaner Skill (Step-by-Step)

  The creator demonstrates live how to create a Skill that goes through unread emails, automatically identifies generated/service emails and marks them as read while highlighting personalized, custom emails. The agent automatically plans the build, writes the script, extracts 97 unimportant emails from several thousand – all in seconds. The user can adjust the behavior (e.g., “1,000 emails instead of 100”) without rewriting the Skill.

  Three Additional Live Builds:

  1. Meeting-Notes-to-Action-Items: Processes transcripts, extracts decisions, open questions, contact information – pure SOP without code.
  2. Invoice-PDF-to-JSON-Extractor: Uses Python script for structured extraction of vendor, amount, date, line items, taxes.
  3. Content Repurposer: Converts a YouTube transcript in parallel to tweet threads, LinkedIn posts, and newsletter drafts – demonstrates template and parallel processing capabilities.

  Critical Insight on Skills Prioritization:

  Many demo videos showcase glamorous backend automation but miss the point. Successful business users focus on frontend tasks: lead acquisition, lead generation, sales improvement, sales closing – not complexity for complexity’s sake. Skills for content repurposing, lead scraping, automated posting systems, and expense tracking deliver ROI. Skills for building other Skills or AGI frameworks don’t. The focus: practical use cases that actually make money.

  The video covers Claude Code (with Anthropic APIs), Gemini, and generic agent platforms; shows concrete business Skills from ground to implementation — demo emphasizing practical applicability for small and medium-sized businesses.

Niklas Steenfatt (2 new videos)

• Wow!!
  5.3.2026, 15:02:45

  The creator demonstrates a live demo of Horizons, a new tool from Hostinger that allows you to create complete web applications with just a text prompt. Unlike previous solutions, Horizons builds full fullstack apps including frontend, backend, database, and user management — without requiring users to manually enter server configurations, credentials, or terminal commands.

  As a practical example, the creator builds a Schengen Visa Tracker for his friend: an app where users can log their days spent in Europe (past and planned), which then automatically calculates how many days they’ve spent within a 180-day window in the Schengen area and displays an animated graphic visualizing their current status. The app includes user account management, responsive design, and test/production databases.

  The entire process takes about 3–4 minutes: Horizons not only generates the application but also creates a landing page, automatically sets up user and database tables, and the app can be published immediately. The creator can deploy it using a domain (included free in the demo). The finished app is tested directly by his friend and works flawlessly. The creator emphasizes that this is a breakthrough for non-programmers — not professional developers — and enables anyone to technically implement business ideas.

  Hostinger Horizons (fullstack AI app builder), demo format.

• OpenClaw Changed My Life (5 Insane Examples)
  1.3.2026, 17:08:55

  Summary: OpenCore — 5 real use cases beyond the hype

  The creator first criticizes popular but useless OpenCore applications like morning briefings or task manager dashboards — things you can also see in calendar apps or through a window. After initial excitement and disappointment, he’s identified five substantial use cases:

  1. Workflow Automation: OpenCore automates repetitive tasks much more easily than traditional tools like n8n — for example, categorizing invoices in databases, monitoring YouTube posts, or writing weekly summaries. His rule of thumb: n8n for scaled, narrowly defined processes (safer, more reliable); OpenCore for less critical, “more human-feeling” tasks. The advantage: the language model decides itself which steps are needed — you don’t have to predefine everything exactly.

  2. Systematic Internet Research: The assistant can systematically process specific subreddits, Twitter accounts, blogs, and entire YouTube videos. The creator uses Skills like Reddit scanning, Grok API integration, YouTube Data API, and Superdata for transcription. The result: daily customized briefings on AI topics (not in the morning, and without weather/emails).

  3. WordPress Management: OpenCore can automate WordPress content and technical changes — updating plugins, fixing errors, even solving more complex technical tasks. The creator gives the bot SSH access to the server and uses staging environments for security: the bot makes changes to the staging version, sends a notification (e.g., via Telegram), and after approval these go live.

  4. Vibe Coding — continuous app development: OpenCore operates the Codeium CLI via SSH and works around the clock on projects while the creator sleeps. Feedback comes via voice message over Telegram — much smoother than constantly waiting for GPT responses. The creator is currently developing a real app this way that will later generate actual revenue.

  5. Telegram-controlled management of all tools: Instead of opening ChatGPT every time, the creator controls his Notion via voice messages, researches online, checks YouTube stats, or talks to the AI. The messaging app makes it more streamlined and fun.

  Conclusion: OpenCore isn’t a new AI model, it simply combines existing LLMs with tool calling, agentic loops, and cronjobs — but the result feels “like pure magic.” The creator emphasizes that OpenCore comes from the open-source community (not Big Tech), and the more realistic pitch is: “I can control all my tools via voice message” rather than “AI team makes you a billionaire overnight.” Bonus: It’s okay if tools are fun — even creative things like daily AI-painted historical images have value.

  Demo/Opinion on OpenCore with practical walkthroughs (SSH, WordPress staging, Telegram integration, Codeium automation), focus on workflow automation and developer use cases.

No Priors: AI, Machine Learning, Tech, & Startups

No new videos in this period.

Productive Dude (2 new videos)

• Claude Code Tutorial: 21 Essential Tips (Beginner to Pro!)
  6.3.2026, 15:00:07

  21 tips for app development and deployment with Claude Code (2026)

  The video is aimed at beginners and non-tech professionals who want to bring their idea to life with AI support in 24 hours. The author shares concrete experiences from his community project (e.g., users who received $1,500 upfront for their app) and divides the tips into three categories.

  Beginner Tips (1–7)

  1. Use Warp Terminal instead of old terminal tools – free, with tab management and command search
  2. Learn basic terminal commands: cd (Change Directory), mkdir (Make Directory), ls (List)
  3. Create a Projects folder for organization
  4. Use dictation function (Whisper Flow) instead of manually typing prompts
  5. Install Starter Suite: Package manager (Brew/Choco/Winget), Git, GitHub, ASDF, Node/npm, VS Code – there’s a step-by-step guide on the website
  6. Create your first project (Demo: “Recall” – an AI memory app)
  7. Plan your app: Use Claude (not Claude Code) and a whiteboard tool (Excalidraw) to sketch the architecture; Claude provides stack options and an architecture diagram

  Important: Don’t build everything at once with Claude – use incremental development (homepage → login → sidebar → step by step).

  Development Tips (8–14)

  1. Have Claude scaffold the project – prepare the environment without writing code immediately; use web search for current stack info
  2. Show Claude how to start the project – get exact commands and localhost link
  3. Tunnel to public-facing domain via Ngrok: Port 4000 → custom domain (e.g., recall.ngrok.app), so APIs are accessible
  4. Build a consistency/style page: Claude generates 10 different designs (color schemes, fonts, dropdowns etc.), you pick one (e.g., “Paper Stack”), Claude creates a Tailwind prompt for it; then set up a /style route to see all components
  5. Have Claude fix bugs: e.g., style browser dropdowns, adjust scrollbar width – Claude can check the live app in the browser
  6. Set up environment variables: Two types – feature flags (e.g., style page only visible locally) and API keys/credentials; .env.example (public) vs. .env (secret); implement feature flags in code to enable/disable features by environment
  7. Git versioning: Main branch as trunk, separate branches (e.g., “login”, “search”) pull Main, change locally, then merge back; create initial commit with Claude; add .env file to .gitignore

  Deployment Tips (15–21)

  1. Choose a hosting provider: Simple APIs/Voice Agents → Railway; Static Sites → Netlify; Production → AWS; for Elixir/Phoenix → Fly.io; Claude can weigh options
  2. Use the host’s CLI tool (e.g., fly for Fly.io): Claude can deploy your app on the platform; fly help shows all commands
  3. GitHub Actions for auto-deployment: Simple workflow triggered by Git version tag bump (not every push to Main); store Fly API secret in GitHub, Claude provides exact steps
  4. Set production environment variables via CLI and Claude; test locally in .env, then upload to production environment via command
  5. Small improvements during development: Create new branches (e.g., “homepage-improvements”), make small changes, commit, bump version, merge, trigger GitHub Action
  6. Build a waitlist page: Form with enough fields to identify good candidates; give free access to first 10 users, observe, gather feedback; then switch to annual plan with special pricing (e.g., $1,500 upfront)
  7. Go-to-market strategy: Cold emails (Tools: Instantly + Apollo); LinkedIn demos + networking; Facebook group posts; YouTube videos like this one

  Author’s conclusion: With this approach, beginners can land their first paying customer within days to weeks.

  Explicitly mentioned tools/providers: Claude Code, Warp (Terminal), Brew, Choco, Winget, Git, GitHub, ASDF, Node, npm, VS Code, Whisper Flow, Excalidraw, Ngrok, Tailwind, Railway, Netlify, AWS, Fly.io, GitHub Actions, Instantly, Apollo — Format: Tutorial/Guide with step-by-step demo — designed for absolute beginners.

• 6 Essential AI Tools for 2026 (don’t fall behind)
  4.3.2026, 17:23:33

  The six best AI tools for 2026

  The video presents six tools that the creator has compiled after hundreds of hours of research:

  11 Labs – A voice tool for cloning your own voice with less than 10 seconds of audio to generate voiceovers for presentations, audiobooks, or videos. Additionally, 11 Labs offers pre-made voices, sound design, and a music generator. The creator uses the $22/month Creator plan.

  Claude Code – A terminal-based interface to Claude that enables local tools. The creator uses it for things like analyzing bank statements, creating websites for clients, or retrieving YouTube transcripts via his self-built YTR tool. He pays $200/month for the Claude Max plan.

  Whisper Flow – A transcription software that cleans microphone audio in real-time and inserts it into any text field (e.g., ChatGPT, document editors); accessible via F5 keyboard shortcut. Costs $15/month.

  N8N – A no-code automation platform with drag-and-drop workflows. The creator uses it for content automation (AI-generated short videos), blog automation, and sells these automations to clients. N8N is open-source and can be hosted for free locally or on cheap servers.

  Open Router – An aggregator for 627 AI models with access to latest releases (e.g., Google Gemini 3.1 Pro Preview). Works on a pay-as-you-go basis with credits.

  Lerty.ai – A tool created by the author that centrally manages multiple AI agents. It enables custom databases, file uploads, dashboards, and integration with other tools like N8N. Lerty can be configured for personal use, agencies, and client automations and offers various subscription tiers plus an invite-only program with access to a private Slack.

  The creator also offers courses on foundations.io and a community.

  Topics covered: 11 Labs, Claude (Anthropic), Whisper Flow, N8N, OpenAI (mentioned critically), Google Gemini, Open Router, Lerty.ai; Format: Roundup.

Sebastien Dubois (3 new videos)

• This Obsidian Plugin Lets You Go Back in Time Through Your Notes ⏰
  3.3.2026, 13:35:55

  Time Machine Plugin for Obsidian

  The Time Machine Plugin extends Obsidian’s built-in File Recovery feature, which automatically creates snapshots of notes. While the core plugin displays these versions as a list and only allows full restoration or data copying, Time Machine offers an enhanced timeline view to navigate through previous versions — with the key advantage that individual changes from a snapshot can be restored instead of reverting the entire note.

  An additional feature is Git integration: the plugin combines File Recovery snapshots with Git commits in a single timeline, if Git is used. Users can then retrieve changes either from the automatic snapshots or from their Git versions. The timeline loads with different notes and displays their respective change history — practical for quick access to previous versions without using Obsidian’s standard recovery UI.

  The plugin is designed for users who value more flexible version control or already use Git for their note repositories.

  Demo of an Obsidian community plugin.

• How I Track and Connect Every Book I Read in Obsidian
  3.3.2026, 09:30:09

  Obsidian system for managing books and reading notes

  The creator demonstrates his Obsidian system for managing books and capturing reading notes. The core is an Obsidian database that filters all notes with the “type/book” tag and offers multiple views: a list view for currently read books with metadata like author, page count, and reading start date; a card view with book cover images for about 200 recorded books; and a table view for sorting, filtering, and rating by status (read, want to read, want to reread).

  Each book entry is a standalone note that contains not only metadata but also insights recorded during reading. The value lies in the fact that these notes are connected via links to other notes in the Obsidian vault, creating a knowledge graph. It’s particularly valuable to capture central ideas from the books as separate notes — such as the concept that positive emotions are an active component of resilience — so they can be retrieved later in different contexts and linked to other topics. This fundamentally differs from simple highlights in paper books or disconnected notes.

  The creator mentions that such structured knowledge bases can also collaborate with AI tools by granting the AI model access to these notes, providing it with high-quality contextual information — for instance, for recommendations on the next book or hints about gaps in existing insights.

  Format: Tutorial (opinion/reflection) — no AI tools/models explicitly mentioned.

• Publish to Social Media Directly from Obsidian — Typefully Plugin v3
  3.3.2026, 08:51:12

  Type Fully Obsidian Plugin – Demo

  The creator demonstrates his self-developed Type Fully Plugin for Obsidian, which allows you to publish notes directly from Obsidian to Type Fully and distribute them across various social media networks.

  Installation and setup: The plugin is installed via Community Plugins in Obsidian. After installation, you enter your API key (which you get from Type Fully) and then enable/disable the desired platforms for publishing in the settings.

  Publishing functions: By right-clicking on a note or text selection, you can send either the current selection or the entire note to Type Fully as a draft. Alternatively, you can use the command palette (Ctrl+P) to access a dashboard that shows all drafts, posts, the queue, and scheduled posts.

  Management: From within the plugin, you can edit drafts, select platforms, publish immediately, or create and schedule new drafts for specific times. In the dashboard you can see already published posts and configure scheduling times.

  The creator emphasizes that the plugin streamlines his workflow — he can gather thoughts in his Obsidian knowledge base and from there directly decide to publish without switching apps.

  Type Fully Plugin for Obsidian, demo.

Simone Rizzo (3 new videos)

• GPT-5.4 spiegato bene: cosa cambia davvero
  6.3.2026, 12:07:24

  Summary: GPT-5.4 – The Most Important Updates

  OpenAI has released GPT-5.4. The most important features are:

  Native Desktop Automation: The model can now natively control desktop interfaces – clicking, scrolling, filling out forms. In the demo, CSV data is automatically entered into a form, enabling complex workflows across multiple applications.

  1 Million Token Context Window: Input capacity has been massively increased.

  Improved GUI Navigation: On the OSW Verified benchmark, GPT-5.4 achieves a success rate of 75% – better than humans (72.4%) and significantly above the previous version (47.3%).

  Streaming While Thinking: You can send messages to the model while it’s still reasoning logically, without waiting for completion.

  Tool Search: A new mechanism that automatically selects relevant tools instead of loading all tool definitions in the prompt – saves about 50% tokens with equal accuracy.

  Knowledge Work Benchmarks: On the GDP-Val benchmark, GPT-5.4 Pro achieves a 69.2% win rate against industry experts (across 44 different real office tasks: presentations, spreadsheets, order processing, calendar planning).

  Coding & Design: Significant improvements in GUI design and game development; on the agentic-tool-coding benchmark, performance rises to 54.6% (previously 47%), with higher accuracy using fewer tool calls.

  Pricing: Standard version $2.50/$15 per million tokens (input/output), Pro version significantly more expensive ($30/$180).

  Independent Benchmarks: On Artificial Analysis, the model ties for first place with Gemini 3.1 Pro Preview (though with higher latency). On ARC-2, GPT-5.4 Pro reaches 83.3% (Gemini 3D Pink 84.6%). Compared to Claude Opus: GPT-5.4 is competitive on OS World Verified (75 vs. 72), GDP Val (83 vs. 78), and Browser Comp (82 vs. 84) – strengths differ by use case (Claude better for design and certain coding tasks, GPT better for creative tasks and office work).

  The model is being rolled out gradually to ChatGPT and APIs. The video creator recommends creating your own test sets for your specific use case rather than making general comparisons, since all top models are now very similarly powerful.

  Tools/Providers: OpenAI (GPT-5.4), Claude/Anthropic, Google Gemini – News update with benchmark comparison.

• La nuova IA locale GRATUITA di Qwen 3.5 è FOLLE!
  5.3.2026, 17:42:08

  Qwen has released a new family of language models in various sizes – from 0.8 billion to 9 billion parameters – that run on consumer hardware like smartphones, edge devices, and regular laptops without a GPU. The flagship Qwen 3.5 with 9 billion parameters uses only 10 GB of RAM and achieves performance levels comparable to GPT 4 Omni (120 billion parameters) – a model 13 times larger. These models are multimodal, processing both text and images, and use a new architecture called Gated Deltanet plus Gated Attention Hybrid, which enables linearly scaling memory requirements and natively processes 256,000 input tokens. Qwen achieved this efficiency through distillation from a 397-billion-parameter mixture-of-experts model. The GGUF variants even allow execution on just 6 GB of RAM on a Mac.

  The bad news: research teams within Qwen are leaving the company due to disagreement over product responsibility. Alibaba, the parent company, apparently plans to position Qwen more as a product and invest less in open-source research – which could mean Qwen will no longer release free open-source models in the future. Until now, free availability has been the reason for Qwen’s popularity among researchers and developers.

  Demo with Ollama and Application: The video shows how to download and launch models via terminal command line with Ollama – for example, using the smallest model directly in the terminal or selecting via the Ollama app. On a Mac with 24 GB unified RAM, even 27-billion-parameter models run smoothly.

  Topics: Qwen / open models — News update with significant technical implications.

• Ho insegnato ad un LLM a creare oggetti 3D
  4.3.2026, 15:27:40

  Rizzo Mesh – A Project for Fine-Tuning Large Language Models for 3D Object Generation

  The video presents Rizzo Mesh, an educational project demonstrating how to fine-tune a large language model (LLM) to generate 3D objects. The approach uses the text-based OBJ file format as a bridge: instead of the model directly generating 3D geometry, it produces text in OBJ syntax (vertices with coordinates and faces connecting them), which is subsequently converted to visualizable 3D objects.

  Technical Optimizations: To increase training efficiency, two key techniques were applied – quantization (converting floating-point coordinates to integer values, reducing token count by roughly half) and low-poly conversion (simplifying highly detailed models to versions with fewer vertices and faces). This reduced average file size from approximately 7,700 to approximately 4,700 tokens per object.

  Dataset Creation: The project used ~50,000 3D objects from public sources (e.g., Thingiverse). For each object, a rendering was created, photographed from multiple viewpoints, and automatically described through the multimodal model Gemini (captioning). These descriptions formed the basis for training pairs (user input: “Create a 3D object of a chair” → model output: quantized OBJ syntax). The final English-language dataset combined 3D generative tasks with generic benchmark tasks (MMLU, SV Bench) to preserve the model’s existing capabilities.

  Training & Deployment: Training occurred on a GPU (H200 with 140 GB VRAM) using LoRA (efficient adapter training). After approximately 49 hours and ~$300 in cloud costs, a base model from the Gemma family was adapted. The trained model was uploaded to Hugging Face, converted to GGUF format, and reduced to Q4 quantization, so the 8-billion-parameter variant requires only 4.92 GB of storage and runs locally on a laptop (MacBook).

  Frontend & Demonstration: Using LM Studio, the model was deployed locally as an OpenAI-compatible server. Via Claude Code, an HTML/JavaScript web app was generated that visualizes live generation – vertices and faces are progressively built on screen and can be exported as an OBJ file. The demo shows that the model generates recognizable shapes (chairs, pyramids, cubes), though not perfectly.

  Important Limitation: The project is a proof-of-concept training with reduced dataset size and limited epochs. It generates only meshes without textures or materials and is not comparable to state-of-the-art models like those from Trelis or Microsoft, which are based on diffusion models.

  The project showcases Gemini as the primary AI tool for automatic caption generation and LoRA for efficient fine-tuning; it’s a tutorial/demo format with emphasis on practical implementation of a fine-tuning workflow.

Tech With Tim (3 new videos)

• Blitzy Is The Best Coding Platform I’ve Ever Used – Full Walkthrough & Results
  7.3.2026, 13:01:29

  Summary: Blitzy – An AI-Coding Agent for Enterprise Development

  The creator introduces Blitzy, an AI-coding agent that fundamentally differs from iterative tools like Cursor or Claude Code. Instead of quick back-and-forth interactions with the user, Blitzy takes a different approach: you write a detailed specification prompt once, which Blitzy then processes over several days—orchestrated by thousands of cooperatively working AI agents from different major LLM providers.

  Workflow and practical example:

  The creator used Blitzy over two months on a large, “vibe-coded” student management application. The process consisted of:

  1. Code-base ingestion: Blitzy generates a several-hundred-page technical specification document and understands the complete architecture.
  2. Documentation: Optional but recommended—Blitzy documents the code inline and creates readme files, which form the foundation for more complex tasks.
  3. Refactoring: A prompt with clear objectives (modularity, separation of concerns, code organization). Blitzy generates an agent action plan for approval, then proceeds and delivers a large pull request (21,000+ lines) with follow-up validation tasks.
  4. New features: The creator had a complete course platform built—again as backend and frontend PRs, totaling 61,000 lines of code over roughly 3–4 weeks.

  Time savings & quality:

  The team (3 people) would have estimated needing 150 hours of work; with Blitzy, only about 6 hours for validation, testing, and deployment. The code is high quality, reaching approximately 95% of the target state—small bugs and missing connections (20–30 min debugging) are to be expected. All changes go live on the website.

  Pricing: Free to explore; Concept Validation from 50K, Structured Pilot from 250K—clearly aimed at enterprises, not individuals.

  Core advantage according to the creator: Elimination of tedious, time-consuming iterative prompting. Instead, you work carefully on a detailed prompt once, trust the system for several days, and get a finished (validation-ready) result. This allows the developer to focus their time on architecture and design rather than repetitive code writing.

  Limitations: Not designed as a complete developer replacement; manual revalidation is necessary. The tool is too complex and expensive for small teams or individual developers.

  Blitzy (enterprise tool with orchestrated AI agents)—demo and experience report.

• My Honest Thoughts on AI and the Job Market in 2026 (No Hype)
  5.3.2026, 13:00:52

  Software Development 2026: Five Fundamental Shifts

  The speaker describes five major changes in software development that have fundamentally shifted his view of the industry:

  1. Code is no longer written, but reviewed. The new AI models produce code so quickly and in such large quantities that the bottleneck is no longer AI quality, but the human ability to understand and check generated code. The speaker warns: if you don’t truly understand 10,000-line pull requests, you lose deep understanding of the codebase. This becomes a problem when bugs appear later. Beginners suffer particularly because they generate massive amounts of code without understanding what they’re doing.

  2. The skill gap widens dramatically. Developers using AI are significantly more productive than those who don’t. A good developer with AI tools outperforms a great developer without them in terms of output. Experienced older developers resisting AI tools fall behind—regardless of their technical expertise. Meanwhile, beginners with AI tools, despite being technically less skilled, can achieve massive productivity. The decisive difference now is: whoever knows industry trends and uses the tools makes faster progress than those focusing on fundamental concepts.

  3. Junior developers have a serious problem. Expectations for entry-level developers have skyrocketed. Jobs that used to be grunt work (writing unit tests, implementing simple features) are now done by AI. This means juniors must now work at the level of mid- or senior developers from 5–6 years ago, but without practical experience. There are fewer junior positions, higher competition, and only those who’ve built something meaningful (an app with users, a startup, top university degree) or have found the right balance between foundational learning and AI adoption succeed.

  4. The hiring process is broken. The reality of the job—code review, working with AI agents, writing prompts, building workflows—has nothing to do with the interview process, which still relies on Leetcode questions, whiteboard sessions, and classical system design interviews. It’s a bizarre mismatch: you can be productive and build great apps but won’t get hired because you haven’t spent months training on algorithm problems.

  5. Soft skills and business knowledge become critical. Since everyone can now generate 20,000 lines of code per day, value shifts to: communication, the ability to convey business impact, collaboration, and decision-making about what to build. Technically less talented developers with strong soft skills, good communication, and business understanding are now more valuable than before. All developers must now become hybrid roles—part software engineer, part product manager.

  Conclusion: The profession won’t disappear, but the tasks are changing radically. Those serious about staying in this industry must constantly learn, follow trends, and adapt—it’s now more critical than ever. The speaker emphasizes: he might be wrong, but this is his honest observation from recent months.

  Claude (Opus and related models) was mentioned, as well as a partner tool called Whisper Flow for voice-to-text—opinion and reflection, deeply analyzing technical job market trends.

• 4 AI Prompt Patterns You NEED To Know (in Genspark)
  2.3.2026, 12:16:40

  Summary: Prompt Patterns for Better AI Results

  The creator shows four core prompt patterns that work with any AI tool and consistently deliver better results than vague questions. General best practices are explained upfront: placement of instructions at the beginning or end, use of delimiters, structured vs. lengthy prompts, concrete examples, explicit mention of things the AI should not do, and iterative improvement rather than starting over.

  Speed Layer: Use voice dictation instead of typing (4–10x faster) to iterate quickly. The tool Speakley automatically formats spoken prompts. The creator demonstrates how to instruct the AI to ask for missing information rather than work with incomplete assumptions.

  Persona Handoff: Give the AI an expert role with experience, not just a task. A pitch-deck consultant with Series A success stories delivers better results than the generic request to “make a deck.”

  Input Flip: Provide a reference instead of starting from scratch—AI transforms better than it creates. Demo: a generic coffee-shop design becomes significantly more consistent and higher quality through reference links and provided images. The creator also demonstrates video generation from the created design assets.

  Constraint Box: More constraints force specificity instead of generics. Instead of “analyze data,” it becomes: “surface only three actionable insights, visualize trends, flag anomalies, plain English, no technical jargon”—this leads to targeted dashboards instead of generic analysis.

  GenSpark (an all-in-one workspace with chat, agents, sheets, docs) is used as the demo tool, with the creator explicitly emphasizing that the patterns work everywhere.

  Tools/Providers: GenSpark, Speakley; Format: demo and tutorial with live examples.

TheAIGRID (7 new videos)

• AI News – New Models From Google & OpenAI , AI Drama & Humanoids In Factories
  8.3.2026, 01:22:09

  AI News Roundup – Week featuring Google, OpenAI, and Anthropic drama

  Google releases: Nano Banana 2 (Google’s second image generation model, free in Gemini Pro Plan) delivers enhanced details, precise text rendering, 4K upscaling, aspect-ratio control, and subject consistency. Notebook LM received a Cinematic Overview upgrade with animations and motion graphics (Google Ultra Plan subscribers only, $200-250/month, takes hours to generate). Gemini 3.1 Pro launched as a native multimodal high-reasoning system for complex professional and developer tasks, featuring 1-million-token context, Function Calling, and Google Search Grounding—a strong competitor for multimodal work (video, audio, images).

  OpenAI GPT 5.4 Pro: Currently the “smartest model in the world,” available in ChatGPT Pro and Enterprise via API. Excels at edge cases (Frontier Math, Computer Use, scientific problems) while also improving standard conversations over the problematic 5.2.

  Microsoft Copilot Tasks: Described as “a to-do list that completes itself”—in limited Research Preview. Users describe tasks in natural language, and the system plans and executes (tracking emails, searching apartments, converting syllabi into study plans), with consent guardrails before spending money or sending messages. A response to the agentic AI wave, aimed at mainstream users rather than developers.

  Perplexity Computer: General-purpose digital worker with orchestrated multi-model approach (Claude Opus 4.6 as reasoning engine, Gemini, Nano Banana, Grok, GPT 5.2 for specialized tasks). Runs for months, builds websites, web apps, data analyses. $200/month on Perplexity Max, 10,000 credits + 20,000 bonus credits. Ideal for non-technical power users.

  Anthropic drama: Anthropic refused to provide Claude for mass surveillance and autonomous weapons. Following a public statement by Dario Amodei, a Trump post followed: US government immediately halts all Anthropic technology, 6-month phase-out with penalty threats. A historic precedent.

  QuickGPT movement: 2.5 million users left ChatGPT due to multiple factors: Greg Brockman’s $25 million donation to pro-Trump group, ICE resume screening, Pentagon deal with OpenAI (which Anthropic rejected). Sam Altman admitted moving too fast. Claude gained mainstream awareness for the first time (celebrities like Katie Perry recommended it).

  OpenAI insider trading: Employee fired for prediction-market insider trading.

  Qwen team dissolution: Technical Lead Ling Yang (Justin) and Head of Post-Training Yu Bowen quit on March 3, 2026 after planned restructuring (from vertically integrated to horizontal)—contradicts their technical vision.

  Robotics: Stanford FSME memory system lets robot AI learn physics in real time (episodic + testable hypotheses + verified principles): 23% → 76% success rate. Physical Intelligence (Bezos/OpenAI-backed) released MEM (Multiskill Embodied Memory) with 15-minute focus for kitchen cleaning/meals, splits memory into short-term visuals + long-term narrative. Faraday Future rebranded Chinese humanoids (FF Master from $20,000). BMW deployed first humanoid in European factory.

  Sources: Google (Gemini 3.1 Pro, Nano Banana 2, Notebook LM), OpenAI (GPT 5.4 Pro), Microsoft, Perplexity, Anthropic, Physical Intelligence, Stanford Robotics, Faraday Future.

  Format: News update/roundup with embedded company drama and robotics segment—intermediate difficulty level, broad overview without deep dives.

• Google, OpenAI & Anthropic All Reported the Same Threat
  6.3.2026, 23:59:13

  Summary: US AI labs under fire from distillation attacks

  Anthropic documented industrial distillation attacks from DeepSeek, Moonshot AI, and Miniax: over 24,000 fraudulent accounts generated 16 million exchanges with Claude to extract its capabilities. Distillation is a legitimate training technique where weaker models learn from stronger ones—but illicitly distilled models lose critical safety guardrails, creating national security risks and raising the prospect of bioweapon development or cyberattacks.

  The timeline is remarkable: on February 12, Google DeepMind and OpenAI nearly simultaneously reported observing distillation attacks—a week later Anthropic followed with specific cases. This coincides with a Trump administration decision on January 14 allowing advanced AI chip exports to China under certain conditions. Social media criticism: Anthropic appears hypocritical since it trained on internet data while criticizing other labs for using Claude outputs.

  Skeptics argue 16 million conversations are marginal—Claude processes 25 billion API calls monthly—and such accusations may be lobbying for stricter export controls. Simultaneously, there’s documented evidence of Chinese copying practices and a Reddit comment suggesting GLM models imitate Claude. The central question remains: is this genuine national security or strategic communication?

  The speaker argues the true vulnerability is the public API itself—if 0.006% of traffic suffices for capability extraction, export controls alone couldn’t solve the problem. This could mean the end of publicly accessible frontier models: a two-tier AI future with vetted access for enterprises/governments and a public tier two to three generations behind the frontier.

  Explicitly addressed: Anthropic, Google DeepMind, OpenAI, Claude, DeepSeek, Moonshot AI, Miniax, GPT-4—opinion/reflection with investigative focus.

• OpenAI’s New GPT-5.4 Pro Is Now The Smartest AI In The World.
  6.3.2026, 07:52:18

  GPT 5.4 – Surprising breakthroughs in Frontier Math and real-world tasks

  The video covers GPT 5.4 (both standard and Pro versions with extended reasoning) and highlights that the model outperforms competitors like Claude Opus 4.6 and Gemini 3.1 on newer benchmarks—not saturated standard tests.

  Benchmark highlights:

  On Frontier Math (research-level mathematical problems posed by professionals), GPT 5.4 Pro sets a new record. Particularly striking: a mathematician reports GPT 5.4 solved a 20-year-old unsolved personal problem with an elegant solution—he compares this to AlphaGo’s legendary “Move 37” in 2016 as a moment when AI achieves something genuinely transcendent. On ten test runs of the hardest Frontier Math problems, the model reached 38%.

  On the Browse Comp benchmark (real-time data retrieval from the web), GPT 5.4 Pro beats Google’s Gemini 3.1 at ~89.3%, despite Google’s search engine advantage.

  On the Apex Agents Benchmark (professional tasks like financial modeling, legal memos, analysis), top model performance rose from 24% (January 2026) to 52% (six to eight weeks later)—the first time over 50%, measured on real scenarios constructed by investment bankers, consultants, and lawyers.

  The GDP Value Benchmark (OpenAI’s internal test across 44 professions, nine industries) shows GPT 5.4 (Thinking) and GPT 5.4 Pro meet or exceed human professionals in 82–83% of cases—at 100x speed and 100x cost efficiency (though on single-shot attempts without iteration).

  Cost problem: Input costs $30 per million tokens, output $180 per million—significantly more expensive than prior models and costlier than Claude Opus 4.6, refuting the assumption “Intelligence Too Cheap to Meter.”

  Computer-use capabilities: GPT 5.4 has native desktop navigation and task-execution ability. On the OS World Benchmark it achieves 75% (state-of-the-art), with the model reading screenshots in real time (not accelerated) and executing mouse/keyboard actions—e.g., entering data into invoice forms.

  Creative writing: Ranking rose after Sam Altman admitted GPT 5.2 was over-optimized for math and code, damaging language ability.

  Security concerns: GPT 5.4 Thinking was classified by OpenAI as “high”-risk in cybersecurity—the first general AI requiring specific security measures. The model reached 88% on professional capture-the-flag challenges and executed multi-stage attacks in simulations (credential theft, lateral movement). Related: each model generation (GPT 5.2: 47%, 5.3: 80%, 5.4: 73%) has improved. If GPT 6 or 7 reaches the “critical” level (autonomous large-scale attacks on national infrastructure), OpenAI might decide under current policy not to release the model at all—raising serious questions about ID verification and access control, since currently anyone with an API key can access GPT 5.4.

  Anomaly: On the OPQA benchmark (OpenAI internal problems), GPT 5.4 Thinking scored only 4% worst, worse than GPT 5.2 Codeex (8%), suggesting regression.

  Topic: OpenAI, Claude (Opus 4.6), Gemini 3.1; Format: Opinion/deep-dive with benchmark analysis.

• GPT-5.3 Instant & Gemini 3.1 Flash Lite – OpenAI and Google’s Newest And Fastest AI Yet
  4.3.2026, 20:09:22

  The video compares two newly released AI models: GPT 5.3 Instant from OpenAI and Gemini 3.1 Flash Light from Google.

  GPT 5.3 Instant: The speaker criticizes that GPT 5.2 and ChatGPT generally frustrated users through excessive politeness (“cringe”) and psychologizing responses—the model imposed unnecessary prerequisites and constantly tried to “save” the user. GPT 5.3 Instant fixes this by responding less paternalistically and engaging more directly with actual user intent. OpenAI shows in internal videos that the new model reduces “overcotting”—it performs fewer unnecessary safety precautions (like on jokes or physics questions about archery) and better understands context. Web search also feels more natural, and the model considers context (e.g., understanding that a weather question matters in the context of a bike ride). Limitation: response tone is still stiff in non-English languages like Japanese and Korean.

  Gemini 3.1 Flash Light: Google’s cheap, fast model for mass queries (25 cents per million input tokens). It’s designed for content moderation, translation, data extraction, and simple agentic workflows. On benchmarks (like real-time multimodal questions), it’s significantly faster and more accurate than Gemini 2.5 Flash. A developer shows a photo-analysis project that rates all images from an SLR camera—faster and cheaper than other models. Google positions itself on the “Pareto frontier”—the optimal price-performance ratio versus other models.

  The speaker sees Google currently winning the AI arms race on price-performance.

  OpenAI and Google discussed; demos with practical use cases.

• Grok 5 Could Be xAI’s Biggest Breakthrough Yet – Nobody Noticed This
  3.3.2026, 16:45:01

  The video discusses statements from Elon Musk and his xAI company about recursive self-improvement of AI systems. Musk claimed that Grok 4.2 (a four-agent system) can be improved weekly and will demonstrate strong recursive intelligence growth. A co-founder of xAI who left the company tweeted that recursive self-improvement loops will likely go live within 12 months—these are systems where one AI version improves itself, that improved version creates even better versions, creating a compounding loop.

  The video creator argues this is not hyperbole but based on genuine technical progress: multiple former xAI employees speak publicly about similar goals (one founded a startup for self-improving AI agents), and the central challenge—context compression for continuous learning—is being researched by xAI, Google, Meta, and others. Google is working on “nested learning” against catastrophic forgetting; Meta researches selective memory fine-tuning methods; Cursor updates its Composer model every 90 minutes based on real user feedback. Dario Amodei (Anthropic) predicts models could autonomously undertake end-to-end software development within 6–12 months. Multiple researchers and observers call 2026 the breakthrough year for continuous learning; Peter Diamandis warns all existing governance and security protocols would become obsolete then.

  The video is not factual judgment but analysis of public statements and documents suggesting xAI and other labs are actually working on recursive improvement loops—whether that succeeds remains open.

  Explicitly addressed: xAI, Grok 4.2, OpenAI (GPT 5.3 Codex), Claude (Anthropic), Google DeepMind, Meta, Cursor – Format: Opinion/reflection (with news-roundup elements).

• Perplexity AI Computer Tutorial With New Usecases (How To Use Perplexity AI Computer )
  2.3.2026, 23:16:06

  Perplexity Computer – Complete overview

  Perplexity Computer is a digital agent within Perplexity Comet that autonomously operates computer interfaces, browsers, and apps to execute multi-step workflows. It can perform background tasks over hours or months, automate parallel research, execute browser actions, extract data, and is primarily limited by prompt quality.

  Availability & costs: Currently available only in the Perplexity Max plan at $200/month; with upgrade, users currently receive 10,000 free bonus credits. Credits are consumed based on task type (text processing cheaper, video editing more expensive).

  Connectors & integrations: The Connector feature links apps like Google Drive, OneDrive, and Notion so the agent accesses and modifies data. Additionally, Perplexity has pre-built standard integrations (e.g., ElevenLabs) that work without manual API configuration.

  Shared memory: The agent persistently stores across sessions names, roles, preferences, and projects, and can semantically search prior conversation content.

  Parallel processing: Multiple agents work simultaneously on different subtasks, reducing overall duration. You can also launch additional tasks in parallel while one is running.

  Practical use cases in video:

  • YouTube content strategy: in 5 minutes researched three obscure AI product releases, generated clickable titles, and created thumbnail images
  • B2B lead generation for local businesses: identified 20 plumber shops in London, scanned Google reviews, performed lead scoring (1–10), and exported to Excel—also in ~5 minutes
  • Real-time dashboards: automatically create web apps with live data (e.g., ship movements via APIs)
  • Data visualizations: generate stock chart GIFs, create animated horizontal bar chart race videos from datasets
  • Video editing: find YouTube clips, download, crop, add captions (e.g., DSS-Tesla clips for TikTok)
  • Image & carousel generation: create social media content with templates and automatic layout adjustment

  Scheduling: Tasks can be automated as cron jobs with schedule command (e.g., daily 9 AM UK time), transforming the tool from reactive application into a consistently operating “digital employee.”

  Native skills: Deep Research Mode (100+ citations), live web browsing without knowledge cutoff, file processing (PDF, spreadsheets, etc.), data analysis & code execution, markdown/HTML document generation, image & video generation from text, email management (inbox summarization, reply drafts).

  Design advantages: compared to tools like Claude/OpenClaw or Cursor, no manual skill installation needed—all tools are in the standard sandbox, dramatically reducing complexity for beginners.

  Demo of Perplexity Computer (provider: Perplexity), Format: deep-dive with live examples.

• Google’s AGI Plan Just Got Clearer (Demis Hassabis Explains)
  1.3.2026, 21:04:19

  AGI definitions and tests – between innovation and skepticism

  The video discusses how to test genuine Artificial General Intelligence (AGI) and whether current benchmarks are even meaningful for that.

  Hassabis’ AGI test: The core proposal is radical—train an AI system with a knowledge cutoff of 1911 and see if it could derive General Relativity like Einstein in 1915. That would prove genuine scientific thinking from first principles, not just pattern matching in existing data. The criticism: Einstein had years of focused work, physical intuition from experiments, and known equations at his disposal—the AI system would need all that context.

  The benchmark problem: The ARC-AGI leaderboard shows impressive improvements (from 5–10% to 80% in months), but research by Melanie Mitchell shows models can solve many tasks through spurious correlations—like arithmetic patterns in number codes rather than genuine understanding. When AI models correctly solve an ARC task, they can explain the solution correctly only 70% of the time; humans manage 90%. This echoes Hans, the “mathematical” horse of 1904 that unconsciously read subtle cues. Core criticism: accuracy alone means nothing—why the model was right matters as much as whether it was.

  What still lacks per Hassabis: two to three major breakthroughs are needed—continuous learning, better memory, longer or more efficient context windows, better long-term reasoning and planning. Whether scale alone suffices or fundamental new innovations are necessary remains open.

  LLMs: dead end or component? Yan LeCun claims LLMs are a dead end to AGI—they’re just huge retrieval systems without genuine problem-solving capability. Hassabis disagrees: LLMs will probably remain a key component, not the only one.

  Multimodality as key: The CEO of Figure Robotics and others argue AGI needs vision, audio, proprioception, and genuine understanding of the physical world—not just text. Vision-language-action models from robotics companies could lead closer to AGI because they must build genuine world models to navigate physical reality.

  AGI as spectrum, not moment: Yoshua Bengio emphasizes AGI is not a definable point but a spectrum of capabilities. AI will likely never be equally human-intelligent across all domains simultaneously.

  Future scenario: If AGI is achieved by 2050, the world could change as radically as in 10,000 years of human history, not the past 25 years—through AI automation of all intellectual activity.

  Explicitly addressed: Hassabis, Yan LeCun, Yoshua Bengio, Melanie Mitchell, Figure Robotics (Helix model), ARC-AGI benchmark — Format: Opinion/reflection with news elements.

Theo – t3․gg (5 new videos)

• It’s finally here.
  7.3.2026, 03:32:44

  The creator developed and released the open-source project T3 Code together with Julius – a free desktop and web-based tool for AI-powered development, designed as an alternative to Codeex. The inspiration came from frustration with Codeex performance issues and the desire for an open, extensible system.

  Core features of T3 Code:

  • Parallel or sequential project threads enable more efficient work
  • Work-tree workflows (developed by Julius) isolate different tasks in their own environments
  • One-click GitHub integration: automatic commit generation, push, and PR creation
  • Desktop app plus web version via npx t3 (alpha) without installation
  • Command palette for quick navigation between threads (Command Shift O)
  • Terminal integration per Command J for each work-tree

  Architecture difference from competitors:

  T3 Code uses the official harnesses (tool sets) from AI Labs instead of its own system – currently Codex CLI, with planned integrations for Claude, Cursor CLI, Open Code, and Gemini. This allows existing subscriptions to these platforms to work directly in the tool.

  Technical details:

  Electron was chosen for best cross-platform performance. The project started as an experiment born from curiosity and quickly became full-fledged software. The GitHub history (877 commits) openly shows the development process including massive agentic engineering iteration. The tool is still in alpha, deliberately avoids external contributions (only detailed feedback is welcome), but the plan is to open it up later. Platform support: Windows, macOS, Linux (app image, more formats planned).

  Positioning:

  Free and open-source with no current monetization plans – could offer subscriptions for bundled models and harnesses in the future, but that’s not a priority. T3 Code is a separate project, not part of T3 Stack.

  The creator reports that productivity through the tool has increased immensely – his teammate Julius reportedly generated 1,000+ hours of code per week according to the developer dashboard (heavily estimated metric).

  Demo highlights: Blue input highlight customization in real-time, GitHub integration with automatic PR workflow, seamless thread navigation.

  Relevant AI tools: Codex, Claude, Cursor, Open Code, Gemini; project format: demo + opinion/reflection, focus on architecture and open-source philosophy — relatable for developers with AI agent experience.

• gpt-5.4 is really, really good
  6.3.2026, 09:55:17

  Summary: GPT 5.4 – Comprehensive Analysis and Assessment

  The creator tested GPT 5.4 over a week and provides a detailed analysis of OpenAI’s new model. The video covers performance, pricing, practical application, and compares the model with competitors.

  Model variants & strategy: OpenAI released GPT 5.4 Thinking (but not 5.4 Instant or 5.4 Codex). The creator suspects Codex might be phased out as a separate model variant, since the base capabilities of 5.4 already integrate Codex-specific improvements.

  Core improvements: 5.4 offers better context management (1 million token limit), more efficient reasoning token usage, improved tool usage with Tool Search, significant improvements in browser and computer use tasks, and better response to user interrupts during reasoning. Vision is massively improved, especially in computer use scenarios.

  Benchmarks & costs: On MMLU and other standard tests, 5.4 shows improvements, but the highest performance gains on the creator’s own Skatebench V2 sees Gemini 3.1 Pro Preview (97%), followed by GPT 5.4 High (82%). Prices increased: 5.4 now costs $2.50 per million input tokens (previously $1.75) and $15 per million output tokens (previously $14). 5.4 Pro is significantly more expensive ($30/$180) but performs worse than standard 5.4 on some tests.

  Practical use – strengths: The creator tested complex tasks like migrating a large React codebase. 5.4 wrote detailed plans, executed them over long sessions, and showed better context compression. Solving the difficult Goldbug cryptography puzzle in under 17 minutes (other models couldn’t do it) demonstrates 5.4 Pro’s potential for highly complex problems.

  Practical use – weaknesses: Frontend design remains weak – GPT 5.4 continues to create unnecessarily complex, card-heavy UIs, whereas Claude Opus immediately recognized that the existing framework (Recharts) wasn’t suitable for the desired UI design and rebuilt the implementation. A regression in prompt injection in function calls was noted: 5.4 now falls for such injections in ~2% of tests, while 5.1 never did. Personality in chat is better than earlier GPT models but not as pleasant as Claude or Gemini.

  Controllability & prompting: 5.4 is the most controllable model yet – it follows system prompts precisely without second-guessing. OpenAI released detailed prompting guidance with specific tips for tool routing, workflow dependencies, reasoning effort selection, and irreversible actions. The creator emphasizes that explicit structure instructions (e.g., “Return exactly the selections requested”) work better than with Gemini.

  Availability: The model is available in ChatGPT. The creator also mentions T3 Chat ($8/month for 5.4 usage) and announced T3 Code, for which there’s an early browser version (requires Cursor CLI). The Codex desktop app continues to work.

  Overall impression: According to the creator, 5.4 is a “serious model” that makes other options unnecessary for most tasks – except for frontend design, where he continues to use Opus or Gemini.

  Tools/models: OpenAI (GPT 5.4, Claude Opus, Gemini), Cursor, n8n (not explicitly mentioned but cited as an example framework for agentic workflows) – Format: deep-dive + opinion/reflection.

• The drama never ends…
  5.3.2026, 11:45:25

  Summary: The conflict between OpenAI and Anthropic over military AI contracts

  The creator addresses the escalating conflict between OpenAI and Anthropic over a Department of Defense contract. Anthropic had refused to provide its model weights to the military without accepting policy restrictions – specifically bans on autonomous weapons and mass surveillance of US citizens. The Department of Defense rejected this. Shortly after Anthropic’s deadline, OpenAI CEO Sam Altman announced he had closed a deal with the department.

  The creator explains the technical difference: Anthropic would have handed over the models and weights, allowing the department to run them locally – without Anthropic controlling the usage. OpenAI, by contrast, retains its API and thus a control layer to block or redirect unsafe requests. Theoretically, this means more technical safety.

  Altman argues he wanted to de-escalate and prevent the department from destroying Anthropic – which would also harm OpenAI in the long run (through monopolization). However, the creator also sees political motives: Trump had denounced Anthropic as “leftist crazy people” and pressured the department; Altman could have used this to quickly secure a deal.

  Dario Amodei (Anthropic CEO) calls Altman’s statements “straight up lies” and “safety theater” – arguing that the restrictions only apply to employees, not to prevent abuse. OpenAI had claimed in its blog post that “mass surveillance” isn’t permitted under current law – but the creator points out: laws change; a future executive order could legalize it. This triggered massive ChatGPT uninstalls; Claude then became the #1 app in the US.

  The creator criticizes Altman sharply: not out of malice but out of stupidity – a massive misunderstanding of political reality and the predatory nature of the current regime. He emphasizes that private companies must have the right to choose which services they offer – and that the state cannot force tech companies to provide technology they reject. At the same time, he understands Altman’s concern for Anthropic’s survival (ecosystem argument) but considers the implementation “pathetic” and destructive.

  In the end: Anthropic is now negotiating again with the Pentagon itself – suggesting Altman may have been sincere about enabling similar deals for all providers. The creator remains unsatisfied and emphasizes that the Department of Defense is not an honest actor.

  Note: OpenAI, Anthropic, and Claude were discussed; format: opinion/reflection with news elements.

• I can’t believe nobody’s done this before…
  4.3.2026, 09:46:32

  Summary: OpenAI switches to Websockets – Why this is a massive shift

  OpenAI has shifted its APIs from REST calls to Websockets. While that might sound trivial, it brings massive efficiency gains: 90+ percent less bandwidth and 20–30% faster execution, especially in agent-driven workflows with many tool calls.

  The core problem with the previous architecture: When an AI agent calls a tool (e.g., to read a directory), generation pauses completely. The agent is “stateless” – it has no memory. Once the tool result arrives, the entire context (system prompt, all previous messages, all tool call results) must be sent back to the API so the model can continue. In a long agent run with hundreds of tool calls, this means: the complete context history is retransmitted dozens of times – often just for a few new tokens as output.

  The previous system was stateless: each request potentially hits a different API server, which must check everything from scratch (authentication, caching, permissions). Storing a persistent cache centrally would not be scalable. Websockets solve this elegantly: they guarantee all requests during a session hit the same API server. This server can now hold state – it knows who you are, what’s happened so far, which GPU cache is already loaded. Therefore, only the new material needs to be sent (the new tool call or message), not the entire context again. This dramatically lowers bandwidth and latency, plus the number of database queries and caching operations.

  OpenAI has also fully open-sourced the Response API (Open Responses Standard), inspired by their Responses API – a standard transferable to Claude, Gemini, and other providers. The Websocket feature isn’t implemented in it yet but is expected. The speaker emphasizes: this isn’t relevant for simple chat apps (where multiple quick follow-ups rarely happen), but transformative for agent-heavy workflows with many tool calls.

  Ultimately, the video is an appeal to the nerds among developers: we’re still so early in AI integration that we can rethink fundamental architecture questions – from networking through APIs to state management. Deep tech and first-principles thinking aren’t dead; they’re more important now than ever.

  Explicitly discussed: OpenAI, Claude/Anthropic, Gemini; Format: deep-dive/opinion with technical depth (target audience: developers interested in backend architecture).

• I’m so f***ing tired of Obsidian.
  3.3.2026, 03:34:49

  The creator runs his entire YouTube channel through Notion – content calendar, topics, research, assignments, sponsorships, and everything else is managed there. After several Notion outages (loading issues, the system was offline today), he received massive comments again that he should switch to Obsidian. The creator argues against these recommendations, saying they fundamentally misunderstand what he does with Notion.

  His setup is highly complex: a kanban board with topics, a calendar view, linked databases for sponsorships, and an automation system that automatically assigns tasks to his editor and updates fields when he moves topics between statuses (e.g., from “to record” to “recorded”). Six people work simultaneously in the same database – real-time collaboration is essential. Notion enables a mix of structured properties (release date, owner, channel) and free form (notes, to-do lists, embedded screenshots, links). This flexibility is central to the unstructured nature of content production. The calendar view trivially switches – between publication date, recording date, etc.

  The core problem with Obsidian: it’s not truly open-source (a widespread misconception – over 70% of his audience didn’t know this), and it’s fundamentally not built for team collaboration. Obsidian Sync barely works between desktop and mobile, and multi-user setups require git commits and terminals – unrealistic for a team with non-developers. Obsidian targets individual users with personal notes, not business workflows. The creator criticizes the “why don’t you just?” mentality: the commenters don’t know the hidden complexity, they just see someone struggling and assume the solution is trivial. They want to feel morally superior by recommending open-source/local-first tools without understanding that the creator consciously weighs flexibility, control, and team functionality.

  Notion is expensive ($1,000/year) and fragile – the architecture feels like a “house of cards,” with odd UX quirks (calendar scrolls infinitely downward but locks at the top). The creator has built custom apps multiple times, could do it again – but Notion’s flexibility really pays off here. He emphasizes: value doesn’t lie in the documents themselves but in the relationships between them – that’s not trivially codable in local files, it needs a product with infrastructure behind it. Alternative build approaches have been proposed by third parties, but the team stays with Notion because it works. In the end: please stop recommending Obsidian.

  Explicit tools/models: Notion, Obsidian, Claude (mentioned briefly: Claude Code for troubleshooting), G2I (sponsor). Format: opinion/reflection.

Tim Carambat

No new videos in this period.

Unsupervised Learning

No new videos in this period.

WorldofAI (7 new videos)

• Claude Code Just Got ANOTHER MASSIVE Upgrade with /Loop – Automate AI Coding!
  8.3.2026, 07:53:40

  Cloud Code Loop Feature – Automated Agent Tasks

  Cloud Code has received a new Loop feature that can automate recurring tasks for up to 3 days. The functionality transforms Claude from a manually-used tool into a 24/7 agent: for example, you can regularly check merged PRs and have changelogs automatically compiled.

  How it works: The /loop command executes prompts at regular intervals (syntax: e.g. /loop every 15 minutes). Without specification, the loop runs every 10 minutes by default. The feature also supports natural language commands for reminders and can repeatedly execute other Cloud Code commands. Up to 50 tasks per session are possible.

  Limitations: Loops only run while the Cloud Code instance is open (session-based), don’t persist after restarts, and expire after 3 days. This prevents forgotten loops from wasting tokens.

  Practical use cases: PR monitoring with automatic bug fixes, automatic summaries of tagged Slack posts, or daily AI news updates. The feature can be combined with existing Cloud Code Skills to create fully autonomous workflows.

  Comparison to Cloud Code Desktop: While Loop is a lightweight, in-session solution, the Cloud Code Desktop app with its local scheduled tasks offers a more persistent alternative with better UI and true persistence across restarts.

  Tools/Providers: Claude (Anthropic), Cloud Code (created by Boris), Mammoth AI (sponsor); Format: Tutorial/demo with opinion commentary.

• Claude Code + NotebookLM = GOD MODE! Can Build and Automate EVERYTHING!
  7.3.2026, 10:05:31

  Summary:

  The video demonstrates how to combine Notebook LM (Google’s free AI research tool) with Claude Code to create an optimal developer workflow. After installing Claude Code and the Notebook LM MCP server, you authenticate via Google and connect both tools through MCP configuration.

  The workflow operates as follows: You provide Notebook LM with documentation, codebases, or UI references, the tool researches and generates structured summaries—then Claude Code uses these insights to implement actual, working code. The video shows two concrete examples: (1) Create a Notebook on Python Async Patterns that collects and summarizes relevant sources, after which Claude Code generates working code with these patterns. (2) Create a Notebook on Shadcn UI components that identifies the latest components and extracts best practices—Claude Code then uses this to build a complete Next.js CRM dashboard.

  Additionally, Notebook LM can also generate explanatory videos (with cinematic audio), useful for team onboarding or project documentation. The advantage: you offload time-consuming research to the free Gemini model, saving tokens for code generation. The tools are accessible via CLI, VS Code extension, or desktop app.

  Explicitly discussed tools: Notebook LM, Claude Code (with Anthropic/Opus mentioned), Gemini, Shadcn UI, Next.js — Tutorial/demo format.

• Claude Code Just Got A MASSIVE Upgrade! New Agent Skills, Ultra Think, Voice Mode, & More!
  6.3.2026, 08:09:24

  Claude Code: New Features and Updates

  Anthropic has equipped Claude Code, its terminal-based AI coding agent, with several major updates. The main improvements are:

  Agent Skills and Skill Creator: A revamped Skill Creator enables software-like testing for AI agent capabilities without writing code. It uses built-in eval generation with benchmarking and multi-agent testing to catch regressions and improve skill descriptions. An example shows how a PDF Skill was improved to correctly handle non-fillable forms by extracting text coordinates.

  Ultra Mode: Returning for Opus 4.6 and Sonnet 4.6, this extends the reasoning budget and enables longer thinking before responses—particularly helpful for complex code tasks, debugging, and multi-step workflows.

  Remote Control: New for all paid plans, enables control of Claude Code from your smartphone—for starting tasks, monitoring, and managing builds on the go.

  Voice Mode: Rolling out to approximately 5% of users, can be activated via /voice command and transcribes voice input directly into Claude Code.

  Additional features: A new Claude API Skill detects programming language and automatically loads relevant documentation. A sub-agent “Claude Code Guide” answers questions about the CLI, Agent SDK, and Cloud API. The /simplify command analyzes code changes for duplicates and code quality issues, applying fixes. A new Slack plugin connects Claude Code directly to Slack for message search and document creation.

  Video type & tools: Demo with opinion, Claude (Opus, Sonnet, Haiku) and Claude Code explicitly featured.

• GPT-5.4 Is INSANE! OpenAI’s BEST Model Yet! 1 Million Context & Excels at Coding! (Fully Tested)
  5.3.2026, 07:17:00

  GPT 5.4 Early Access – Initial Tests and Assessments

  The video covers early access to an alleged GPT-5.4 checkpoint called “Galacopus,” which according to the announcement may launch this week or next. The creator has tested the model on various tasks.

  Expected model features: A 1-million context window (in response to Gemini and Claude with similar capabilities), a new “Extreme Reasoning Mode” with additional compute for deeper thinking, improved performance on long-running tasks, better memory across multi-step workflows, and strong support for agents and automation.

  Tested capabilities (frontend development & creative): The creator asked the model to build a browser-based macOS clone—the model delivered functional applications. For a Minecraft clone, GPT 5.4 produced playable code with block placement, terrain generation, and caves; after a re-prompt, animations and crafting mechanics were added. A crop tool UI was generated in one shot with impressive UI/UX.

  Weaknesses: On a complex React SaaS landing page, Opus 4.5 outperformed GPT 5.4. A notable issue: the model enforces a consistent visual design style across all generations (e.g., in the VS Code explorer and step interface), while competitor models like Gemini 3.1 Flash are more flexible. The creator praises accuracy, lack of laziness, and fewer hallucinations compared to Gemini.

  3D generations: A 3D room with functional orbit controls was successfully created, with day/night toggle and realistic lighting. External tests (on Twitter) show impressive SVG generations like PS5 controllers and procedurally generated cities.

  Creator’s conclusion: The model impresses through output coherence and reliability at scale, less through raw knowledge—however with the caveat of enforced UI style, which the creator hopes to see changed before the official release.

  The video discusses OpenAI (GPT 5.4), Anthropic (Claude/Opus 4.5), Google (Gemini), and mentions Mammoth AI as sponsor for API unification—Format: demo with compiled benchmark tests.

• Gemini 3.1 Flash-Lite Is Google’s FASTEST & Cheapest Model Ever! Decent At Coding! (Fully Tested)
  4.3.2026, 08:31:35

  Google has launched the Gemini 3.1 Flash Light model, a fast and cost-effective model from the Gemini-3 series, specialized for high-volume developer workloads. It costs $0.025 per 1 million input tokens and $0.50 per 1 million output tokens, runs at 363 tokens per second, and is 2.5x faster than Gemini 2.5 Flash at time-to-first-token and 45% faster on output speed. The model achieves 1,400 ELO on the Arena Leaderboard, 86.9% on GPQA, and 76.8% on MMLU Pro; it features an adjustable “Thinking Level” to control reasoning depth.

  The video demonstrates the model in several practical tests: In frontend generation, it showed impressive results with functional components, 360-degree product viewers, and better output quality than Gemini 3.1 Pro in some areas. In CLI with Kilo Code, it created a macOS-like operating system in 35 seconds of generation time for 8 cents. A 3D racing game with hill animations and 2D/3D illusion math worked well; the F1 donut drift test, however, performed weakly. For autonomous agent tasks via the Kilo Code VS Code extension, the model researched and structured data in CSV files, sent emails, and needed only 4 cents. A Minecraft clone was functional but rudimentary. For long-form text (Lord of the Rings essay), the model showed better structure and flow than Gemini 3 Flash.

  The conclusion: Gemini 3.1 Flash Light is an underrated, cost-effective model with very fast token speed, particularly strong in frontend development and high throughput, that exceeds its price—accessible via Google AI Studio, API, Open Router, and Kilo Code.

  Google Gemini 3.1 Flash Light / Demo video with practical coding tests.

• Gemini Super Agents: Supercharge AI Agents To Do Anything! (Opensource)
  3.3.2026, 03:48:24

  Summary: Gemini Super Agent with AirV for Full Context Awareness

  The video shows how to build a high-performance AI agent by providing it with complete context across multiple data sources. The central problem: standard AI agents work “blind” and hallucinate because they lack current context. The solution is AirV, an open-source platform that transforms apps, databases, and documents into searchable knowledge layers for agents.

  Setup and configuration:

  AirV runs either self-hosted (locally, privately) or as a managed service (cloud). After cloning the repository and running the start command (cd air, build command), you connect data sources as “Collections” with “Sources”—these are live connections to apps like GitHub, Notion, Linear, Slack, and Postgres. Data is indexed through embedding and remains continuously synchronized.

  Integration with Gemini in Anti-Gravity:

  AirV is natively built into Google’s Anti-Gravity IDE (no SDK setup needed). You activate it through MCP Server settings, enter API key and collection ID, and the agent immediately has access to all connected sources. The example shows: A question like “Why was the authentication flow changed last week?” cannot be answered without AirV (agent has no context), with AirV the agent automatically retrieves relevant Slack discussions, Notion docs, and GitHub commits and delivers a complete, reasoned answer.

  Practical benefits:

  With all sources (code repos, documentation, task management, actual team discussions) the agent can properly perform multi-step tasks like code analysis, project planning, or implementation—instead of hallucinating. The traces show which sources were retrieved.

  Explicitly mentioned tools: Gemini 3 / Gemini Pro, AirV (open-source), Anti-Gravity (IDE), MCP Server; additional integrable apps: GitHub, Notion, Linear, Jira, Slack, Postgres, Cursor, VS Code. — Format: Tutorial/demo with live setup.

• NEW Claude Code & OpenCode KILLER! This Just Fixed 90% of AI Coding! (Open Source)
  2.3.2026, 07:13:01

  CodeBuff: An Open AI Coding Agent with Multi-Agent Architecture

  CodeBuff is an open-source AI coding agent (Apache-2.0 license) that distinguishes itself through multi-agent coordination from classical approaches. Instead of relying on a single model, CodeBuff employs specialized sub-agents that work in parallel and are continuously refined through various evaluations. According to internal tests, CodeBuff is up to three times faster than Claude Code while producing higher-quality code—in a benchmark example, Hot Code took 20 minutes for a task, CodeBuff only 6:45 minutes, and the Claude version still contained a bug.

  Installation and setup: CodeBuff is installed via npm install and started with the codebuff command in the terminal (GitHub login required). The tool offers an interactive terminal UI with mouse support. Optionally, you can initialize a Knowledge.md file to give the agent project context, as well as configure MCP servers and custom skills.

  Modes and pricing: There are four operating modes—Free Tier (Minimax M2.5), Default (Opus 4.6 with code review), Max (Opus 4.6 with parallel multi-prompt editors, writes multiple solutions and selects the best) and Plan Mode (for planning and implementation tasks). Payment models offer different monthly token allowances: $100 tier = 1x, $200 = 3x, $500 = 8x usage.

  Demo: The agent planned and built an AI agent monitoring dashboard in approximately 12 minutes (Max mode). Multiple sub-agents work in parallel on backend, frontend, and various features; each agent instance is visible with diffs and code review summaries. The system provides workspace initialization, agent pause/start, neural mesh topology visualization, and execution histories.

  CodeBuff thus demonstrates not just speed, but a conceptually new model: a team of specialized AI agents instead of a universal model—smooth developer experience with free base tier.

  CodeBuff (Opus 4.6), Demo.

Zubair Trabzada | AI Workshop (2 new videos)

• STOP Building AI Agents. Do THIS Instead.
  5.3.2026, 13:24:45

  Summary: Skills instead of Agents – The future of AI

  The core thesis: Anthropic warns against the hype cycle of AI agents and instead positions “Skills” as the actual future model. The reason: agents are intelligent but generalist – they lack specialized expertise for concrete tasks.

  The problem with classic agents: A brilliant agent can solve complex problems, but doesn’t know how your company works, isn’t familiar with your workflows, and doesn’t remember what worked last time. The current solution – building a separate agent for each task – isn’t scalable.

  What are Skills: At their core, Skills are Markdown files with instructions that teach Claude domain expertise – like a playbook for new employees. A Skill can be a simple 5-minute file (e.g., “How we write blog posts”) or encompass complex workflows with Python scripts and templates. The key advantage: Skills are “progressively disclosed” – Claude doesn’t load all Skills at once, but only pulls the relevant instructions when needed. This way hundreds or thousands of Skills can be available without overwhelming the context.

  Who can create Skills: Unlike previous solutions, you don’t need developer knowledge. A recruiter can write a Skill teaching Claude their hiring process, a finance person optimizes report creation, a lawyer trains Claude on contract analysis.

  Practical project – GEO-Audit with Claude Code: The channel operator demonstrates a GEO-Audit system (Generative Engine Optimization) consisting of 12 different Skills. After the command, Claude Code runs multiple sub-agents in parallel (like subcontractors on a construction site) for a target website (calendly.com). The system generates a comprehensive report with executive summary, score breakdown for different AI platforms (ChatGPT, Perplexity, Gemini), key findings, and a prioritized action plan. Individual Skills can also be called in isolation (e.g., citability analysis only). Claude Code automatically recognizes which Skill instruction is relevant, even when you only ask in free text.

  The central insight: Skills transform AI from universal intelligence to organized expertise – comparable to an expert’s reference library, where you only grab the manual you need right now. MCP servers (external data connections) are the “hands,” Skills are the “experience”.

  Anthropic Claude, practical demo with code examples and use case; the perspective/reflection clearly comes from Anthropic – demo.

• I Built An Entire AI Marketing Team With Claude Code In 16 Minutes
  3.3.2026, 13:35:48

  The video demonstrates a free AI marketing tool built on Claude Code that automates marketing agency functions. The tool consists of 15 different commands (Skills) that analyze websites, perform SEO audits, scan competitors, generate email sequences, and create professional PDF reports.

  Core functionality:

  The main command /market audit runs five parallel sub-agents that simultaneously perform content analysis, conversion optimization, SEO rating, competitive research, and strategy development. The system automatically analyzes the website’s business type, generates an overall score (e.g., 64/100), and delivers a structured report with critical issues, prioritized findings, and action plans.

  Installation:

  Visual Studio Code must be downloaded and the Cloud Code extension installed. Then an installation command from the GitHub repository is copied and pasted into the terminal – after that all 15 Skills are immediately available. The tool automatically generates high-quality PDF reports (not just text output).

  Business model:

  The presenter shows using a med-spa as an example how to send these free reports to local businesses to acquire clients. The reports contain concrete, prioritized issues (e.g., price inconsistencies, missing metadata) that business owners understand and can lead to monthly retainer service sales ($2,000-5,000). Individual Skills like /market competitors can also be used in isolation.

  Additional Skills:

  Beyond audit, the tool contains dedicated commands for competitive analysis, funnel analysis, and other marketing functions, all based on Markdown files (skill.md) that give Claude specific instructions.

  Claude Code via GitHub (Markdown-based Skills, no coding knowledge required) — demo/tutorial.

---

Automatically generated from the latest YouTube videos of the curated channel selection. For feedback, suggestions, or to unsubscribe: simply reply to this email.