GPT-5.5 vs. Opus 4.7: The Big Showdown of the Week (2026-04-26)

GPT-5.5 Meets Opus 4.7: Two Flagships, Two Philosophies, One Winner Per Task

Sunday, April 26, 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 in peace — or copy a summary into the LLM of your choice and go deeper. Click the link under each summary to see the original video.

Model Releases & Benchmarks

Beyond the main duel, DeepSeek V4 grabbed attention: the Chinese model launched as Flash (284B parameters, 13B active) and Pro variants (1.6T parameters, 49B active) and brings a hybrid attention architecture enabling a 1-million-token context window with up to 90% KV-cache savings – technically remarkable, as Tim Carambat and AI Explained highlight. Practical tests from WorldofAI were disappointing though: on UI generation and 3D projects, DeepSeek V4 Pro fell well behind Opus, Qwen, and MiniMax M2.7, despite near-unbeatable costs (1.4 cents / million input tokens). Kimi K2.6 from Moonshot AI surprised positively: Alejandro AO and WorldofAI confirmed benchmarks on par with Opus 4.6 and GPT-5.4 at 94–95% lower cost; the model supports 12-hour sessions, over 4,000 tool calls, and up to 300 parallel agents. Grok 4.3 Beta (xAI, ~500B parameters) and Qwen 3.6 Max (Alibaba) were discussed as coming competitors; Elon Musk announced Grok 5 as an AGI candidate.

Local & Open-Source AI

Qwen 3.6 (35B parameters, 3B active) establishes itself as a solid all-around local model according to DevExpert and Tim Carambat: it achieves around 65 tokens/second on a 32-GB-RAM machine, significantly outperforms Gemma 4 on vision tasks, and generates working web apps including tests – something Gemma 4 fails to deliver. Tim Carambat additionally introduces Ternary models from Prism ML, which via a third weight class (0, 1, -1 instead of just 0/1) achieve near 16-bit precision with a fraction of the memory footprint – relevant for edge deployment on smartphones and low-end laptops. Tech With Tim shows the complete workflow of integrating local models via Ollama into OpenClaw, with Gemma 4 currently being the best option with tool support; the channel recommends a hybrid approach where cloud models are only called for complex tasks.

Claude Code & Anthropic Tooling

Claude Design was the most discussed new Anthropic product of the week. The tool generates code-based interactive prototypes (HTML, CSS, SVG) rather than Figma primitives, making the handoff to Claude Code seamless. Brian Casel, Kyle Balmer, and Leon van Zyl rate it as strong for early ideation and marketing assets, but not as a complete product development tool; Casel prefers direct work in Claude Code for existing codebases. Aggressive usage limits are the dominant criticism across all reviews: Kyle Balmer (200$ plan) exhausted his weekly design budget after a single prompt; Nate B. Jones estimated an afternoon session with Claude Design at $42 USD. As a workaround, Alex Finn recommends using ChatGPT’s image generation for the exploration phase and only finalizing the refined result in Claude Design. Additionally, Anthropic introduced Claude Code Routines: scheduled or event-based automations running in the cloud that create pull requests without the local machine running – Leon van Zyl and Ben AI demonstrate Auto-Improver and security-audit routines. Mark Kashef dives deep into Claude Code’s seven memory layers and shows how a customized memory system via hooks and auto-injection stabilizes context across sessions. NeuralNine delivers a comprehensive power-user crashcourse on slash-commands, sub-agents, agent teams, work-trees, and skills.

Coding Agents (non-Claude)

OpenAI’s Codex Desktop app has grown into a full-featured Computer-Use platform: it controls arbitrary Mac applications via GUI, runs multiple agents in parallel, and integrates GPT image generation plus an in-browser editor with annotation features – without API dependency. Nate B. Jones and Alex Finn see this as the most direct threat to Claude’s domain because legacy systems and internal tools become automatable without any API adaptation. Cole Medin presents a system for parallel development with Git Worktrees and Claude Code: each agent gets an isolated directory, reviews run in a fresh context window, and a self-healing layer improves overarching rulesets rather than just fixing bugs. Pi, a minimalist open-source coding agent, is introduced by Cole Medin as a lean alternative to Claude Code: npm install, arbitrary LLM support, self-modifiable via extensions – combined with Archon as harness builder yields a plan-implement-validate workflow with interactive markdown planning. Everlast AI advocates the paradigm shift from n8n-style low-code workflows to CLI-based agentic coding pipelines (Claude Code, OpenAI Codex, Gemini CLI), with Claude Code dominating at 46% “Most Loved” in a survey.

Personal AI OS & Agent Frameworks

Hermes from Nous Research was the new open-source agent project of the week and was detailed on no fewer than four channels (AI mit Arnie, Julian Ivanov, Nic Conley, Niklas Steenfatt, Tech With Tim) and compared with OpenClaw. The core difference: Hermes has an integrated learning loop – after each task, the agent evaluates its performance and autonomously creates new skills, while OpenClaw works session-based and configuration-oriented. Both run locally or on a VPS (Hostinger offers one-click installation), communicate via Telegram, and use persistent storage; Hermes runs on Python, OpenClaw on TypeScript. Niklas Steenfatt tests memory transfer between both systems and recommends running them in parallel via Paperclip. Tech With Tim additionally presents MaxClaw – a hosted OpenClaw variant on MiniMax infrastructure with the M2.7 model, which achieves 97% compliance on tool-calling with 20–40 tools and comes out-of-the-box with web search, image generation, and code execution. Cole Medin’s “Dark Factory” project shows the next level: Archon-controlled workflows on a VPS triaging GitHub Issues, implementing, validating via browser automation, and autonomously merging pull requests – the operator writes no code, Kimi K2.6 serves as the cost-effective backbone model.

AI-Automation & Workflows

Nate Herk demonstrates Claude Code combined with Playwright CLI for three core scenarios: automated QA with iterative bug-finding and -fixing, fault-tolerant web scraping with automatic fallback (Google → DuckDuckGo), and login-persistent automation via Chrome user data. Playwright CLI is preferred over MCP servers as it consumes significantly fewer tokens. Bart Slodyczka shows in two tutorials the complete use of Claude Cowork – including free, private operation via Ollama with Gemma 26B or Gemma E2B as well as integration via Tailscale. WorldofAI presents “Bart Mode” from Tracer, turning spec-driven development into fully automated workflow orchestration: the tool breaks an epic into parallel tasks, executes them with Claude Code, and continuously verifies results against specifications. On the conceptual level, Nate B. Jones warns in a separate video against “World Models” – automated management systems confusing information flow with judgment, producing silent misinterpretations that only surface when decision quality has imperceptibly declined.

AI-Video & Content Creation

Nate Herk shows a complete video-editing pipeline: Claude orchestrates Videowise for cut detection, 11 Labs for transcription, and HyperFrames for motion-graphics creation – all via natural-language description with timestamp planning. Productive Dude complements with a HeyGen tutorial: Claude Design generates motion-graphics assets with greenscreen background, Claude Code renders them via HeyGen for Agents into a fully automated news video with avatar, script, and overlays. GPT-4o Image 2 (also “Images V2”) received considerable attention for its ability to perform reasoning over composition in 10–20 seconds, render text with 99% accuracy, deliver eight coherent frames per prompt, and retrieve live data during generation. Kyle Balmer tests five business scenarios (YouTube thumbnails, Facebook ads, comic strips, event posters, LinkedIn carousels) and gives GPT Images 2.0 the edge over Google’s Nano Banana 2 in all five – especially for consistency with sequential prompting and genuinely functional QR codes.

AI-Business, Marketing & Freelancing

Liam Ottley argues that Claude Code has lowered automation development costs so far that retainer models for AI agencies become profitable starting 2026 – citing Tyler as an example offering AIOS installations for $2,500/month. Dave Ebbelaar recommends developers position themselves as bridges between AI-enabled, AI-first, and AI-native companies, combining full-stack knowledge and process understanding. Nate B. Jones designs five orientation principles for the AI job market in a separate video: understanding over generation, explanation artifacts as companion documents, proven transactions over credentials, public work, and inseparable proof – he presents his “Talent Board” platform for this. The core: the bottleneck shifts from model skill to precise specification.

PKM & Knowledge Management

Nate B. Jones compares Andrej Karpathy’s wiki approach (ingest-time synthesis into text files) with his own Open Brain system (query-time retrieval from structured tables) and proposes a hybrid model: Open Brain as SQL database as single source of truth, a “compilation agent” generates wiki pages and daily briefings on-demand – synthesis power without error accumulation. Simone Rizzo presents LLM Wiki as a third generation after RAG and agentic file search: knowledge is compiled into a persistent, linked markdown structure and continuously updated by agents; Obsidian visualizes the knowledge graph. AI Foundations and Productive Dude demonstrate Live Artifacts in the Claude Cowork desktop app, aggregating real-time data from Notion, Google Calendar, and Gmail into interactive dashboards.

Prompting & AI Literacy

Nate Herk dedicates a video to token management in Claude: the 1-million-token window is “insurance, not a goal” – retrieval accuracy measurably drops from 92% at 256k to 78% at 1M tokens (context rot). Practical countermeasures include /rewind for clean contexts, manual session handoff at 60% utilization, markdown conversion (HTML: 90% savings), and specialized session chains. Kyle Balmer discusses the Claude vs. Codex question in several Q&A videos from a user perspective: Claude for scoped knowledge-work and complex planning tasks, Codex for quick single pages and templates with more generous limits. TheAIGRID provides beginner tutorials on GPT-5.5 in Codex (plan mode, reasoning-level adjustment) and ChatGPT Images 2.0 (select function for precise image editing, structured multi-asset prompts).

AI Industry & Strategy

Nate B. Jones analyzes the fundamental architecture difference between Anthropic’s vertical stack (structured connectors, MCP, Conway framework) and OpenAI’s horizontal Computer-Use approach; the acquired team (Software Applications Inc., creators of Sky) brought deep Apple OS expertise. Kyle Balmer highlights the infrastructure imbalance: China produced more electricity in 2024 than the USA, EU, and India combined, which could shift AI inference long-term in favor of Chinese providers. Anthropic’s IPO preparations (target: $60B, valuation $800B, annualized revenue ~$30B) and Amazon’s $25B investment are cited by several channels as explaining pressure on user limits and tokenizer changes. Theo (t3.gg) analyzes Cal.com’s switch to closed-source as a signal for a new security landscape: AI models democratize exploit-finding, so open-source projects must defend themselves under token budgets rather than security obscurity. SAP CTO Philipp Herzig (No Priors) describes three AI transformation levels in enterprises: UI, business processes, and data layer; Jewel for Consulting reportedly increased consultant efficiency by 30%.

AI & Society / Future of Work

MoureDev (Brais Moure) argues that the education system is structurally designed for decade-long technology cycles, while AI enforces quarterly changes – officially approved curricula take three years while today’s 18-year-olds build functioning software with coding agents. The real gap is between self-adapters and those who need every step shown. In a second video, he debunks the thesis that AI is the main reason for tech layoffs: pandemic overcapacity, AI-washing for stock communication, and documented copycat layoffs (Stanford study) are the real drivers. David Shapiro analyzes two narrative threads: the structural incentives (cost control, lawsuit avoidance, user-friendliness) that degrade chatbots over time – analogous to social media “shitification” – and the UBI debate, where he refutes “Tech Right” arguments (Marc Andreessen, Gil Verdon) against universal basic income. The Stanford AI Index Report (referenced by Datapizza) shows 53% global generative-AI adoption in three years – faster than internet and PC – but heavily unequally distributed: Singapore and UAE over 60%, USA only 30%; researcher inflow to the USA dropped 89% since 2017.

In Brief

Everlast AI brings an in-depth interview with BCI pioneer Niels Birbaumer, who rates Neuralink as dangerously reckless for the mass market and describes AI-powered neurofeedback (support-vector machines after ~20 hours training) as indispensable for medical BCIs. Theo (t3.gg) analyzes the fake GitHub-stars ecosystem (6M suspicious stars on 18,600 repos, $3.4–4.4M annual revenue for providers) and its influence on VC funding rounds. NeuralNine shows building an AI research assistant in Python (LangChain + GPT-5 + Consensus MCP server), where the MCP server eliminates hallucinations in paper references. Coding with Lewis visualizes database sizes via stacks of paper and promotes Supabase/Postgres. NeuralNine delivers a yfinance crashcourse with WebSocket real-time quotes, options data, and equity screening. Perplexity Computer ($200/month) is tested by Tech With Tim as a zero-configuration alternative to OpenClaw for ad-hoc automation – cloud-native with automatic orchestration over frontier models.

AI Explained (1 new video)

• GPT 5.5 Arrives, DeepSeek V4 Drops, and the Compute War Intensifies
  24.4.2026, 18:45:02

  The video provides a detailed summary of the latest developments in the AI world, particularly the release of GPT-5.5 and DeepSeek V4. GPT-5.5, which doesn’t yet have an API, shows mixed results across various benchmarks. It outperforms Opus 4.7 and Mythos Preview in some areas, such as Agentic Coding, but underperforms in others, like SWEBench Pro and Humanity’s Last Exam. OpenAI emphasizes the importance of intelligence per token and the model’s efficiency. DeepSeek V4, a Chinese model, supports a context length of one million tokens and is comparable to GPT-5.2 and Gemini 3 Pro in many benchmarks, but significantly more affordable. The video also discusses GPT-5.5’s capabilities regarding hallucinations, cybersecurity, and its ability to handle self-directed tasks. It’s noted that GPT-5.5 performs better in some areas like health diagnostics and spreadsheets, but worse in others, such as chain-of-thought control. DeepSeek V4 also demonstrates strong performance in long contexts and specialized tasks. The video concludes with a reflection on AI’s impact on productivity and the workplace, as well as the challenges posed by compute scarcity.

  The video explicitly covers OpenAI (GPT-5.5), DeepSeek (DeepSeek V4), and Anthropic (Opus 4.7), and is geared more toward intermediate and advanced users.

AI Foundations (1 new video)

• Claude Just Released LIVE Artifacts… (Amazing Results)
  21.4.2026, 04:05:57

  The video explains the new “Live Artifacts” feature in the Claude Co-work Desktop app. Live Artifacts are interactive mini-apps or dashboards that display and update data in real-time from connected tools like Notion, Google Calendar, or Gmail. The creator demonstrates building a personal Daily Briefing dashboard that consolidates and live-updates appointments, emails, and CRM notes from various tools. The feature enables creating personalized, dynamic work environments directly within the Claude app. Live Artifacts can be designed and customized as needed to create a central hub for different workflows.

  The video explicitly focuses on Claude and is geared toward Intermediate or Advanced users.

AI mit Arnie (2 new videos)

• GPT 5.5 is a machine
  25.4.2026, 14:03:17

  The video compares the AI agents Hermes and OpenCla, with a focus on Hermes. Hermes is an open-source project by New Research that stands out for its ability to grow with the user and improve itself. In contrast, OpenCla is an orchestrator that executes tasks and is configuration-oriented. Hermes runs with Python and has persistent memory, while OpenCla runs in TypeScript and uses session-based memory.

  The installation process of Hermes is explained in detail, including the possibility of installing it on various platforms such as Linux, macOS, WSL2, and even Android. Users can install Hermes on a virtual private server (VPS), which has the advantage that the agent is always accessible. Hostinger is mentioned as a simple provider for installing Hermes.

  After installation, the video shows how to connect Hermes to Telegram and how to use various skills and tools. Hermes can automate tasks, such as monitoring YouTube channels and creating video summaries. The agent can also independently create and improve skills without user intervention.

  Another focus of the video is the migration from OpenCla to Hermes. It’s recommended to perform the migration gradually and manually, as the memory and skill structures of the two agents differ.

  The video concludes with a discussion about when to use which agent. OpenCla works well for executing many individual tasks, while Hermes is ideal for tasks that need to improve and adapt over time, such as trading bots or social media managers.

  AI Tools/Models/Providers: Hermes Agent, OpenCla, OpenAI, Anthropic, OpenRouter, Hostinger

  Target audience: Intermediate

• Is Hermes the OpenClaw killer?
  22.4.2026, 08:05:18

  Summary of the YouTube video:

  The video presents a detailed comparison between the AI agents Hermes and OpenCla, with a focus on the features, installation, and use cases of Hermes. Hermes is described as an agent that grows with the user and continuously improves itself, while OpenCla serves more as an orchestrator for executing tasks.

  Key points of the video:

  1. Introduction to Hermes:
  • Hermes is an open-source project by News Research that already has 100,000 stars on GitHub.
  • It offers a light mode and detailed documentation.
  • Hermes runs with Python and has persistent memory, while OpenCla runs in TypeScript and uses session-based memory.
  1. Installation and setup:
  • Hermes can be installed on Linux, macOS, WSL2, and even Android.
  • Installation on a virtual private server (VPS) is recommended to make the agent accessible 24/7.
  • The Hostinger provider offers a simple one-click installation for Hermes.
  1. Features and comparison with OpenCla:
  • Hermes has a rich memory that extends across sessions, while OpenCla relies on Markdown files.
  • Hermes has integrated learning loops and improves itself independently, while OpenCla uses manually installed skills.
  • Hermes has 30 pre-installed skills that can be used immediately.
  1. Security and data protection:
  • Hermes has a smaller attack surface compared to OpenCla, as it requires fewer external skills.
  • Installation on a separate server is recommended to minimize security risks.
  1. Practical applications:
  • Hermes can be used for various tasks, such as monitoring YouTube channels, creating social media posts, and automating trading strategies.
  • The agent can adapt to the user’s writing style and find relevant information.
  1. Migration from OpenCla to Hermes:
  • A complete migration from OpenCla to Hermes can cause issues, as the memory and skill structures are different.
  • It’s recommended to manually migrate and test skills and memory files.
  1. Summary and recommendations:
  • Hermes is particularly suited for tasks that improve over time, such as trading bots or social media managers.
  • OpenCla is better suited for orchestrating individual tasks.
  • Both agents can be combined to leverage the advantages of both systems.

  Final comment:

  The video explicitly addresses the AI agents Hermes and OpenCla and is rather aimed at Intermediate users who already have experience with AI tools and want to deepen their knowledge.

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

• ChatGPT Images Just Replaced Three People on Your Team.
  25.4.2026, 15:00:55

  ChatGPT Images Just Replaced Three People on Your Team – Summary

  OpenAI has achieved a breakthrough in image generation with GPT-4o Image 2: The model won 93% of blind pairwise comparisons in Image Arena – an unprecedented 26-point lead over the next-ranked model (Google’s Nano Banana 2 with 67%). The difference becomes concrete: A developer (Takuya Matsuyama) received from a single prompt a complete, aesthetically coherent landing page with Hokusai illustration and Wabi-Sabi design, extracted from his blog posts about Japanese aesthetics.

  Three architectural mechanisms:

  1. Thinking Mode: The model spends 10–20 seconds reasoning about composition, typography, and object placement before generating pixels.
  2. Web Search in the generation loop: The model can retrieve live data during composition – example: geologically accurate depth maps of the Hormuz Strait combined with Richard Scarry illustration.
  3. Eight coherent frames per prompt: Character and object consistency across eight images enables manga, comic panels, magazines directly from one input.

  Additionally, there is a self-verification pass where the model checks and corrects its output.

  Four new workflows:

  – Localized ad campaigns without vendor overhead (accuracy in Kanji, Hangul, Devanagari)

  – UI specification as rendering target in Codex (PM describes in German, model generates mock-up, coding agent implements)

  – Live-data brief (competitor pricing + current case studies researched, written, designed in one prompt)

  – Coherent design system from one request (floor plans, color palettes, materials, inspiration shots)

  The dark side:

  The same capability enables photorealistic forgeries: manipulated restaurant receipts, Slack screenshots with foreign avatars, fake boarding passes, pharmacy labels. Text accuracy is 99%; 70% of test participants believed outputs were real photos. Content Credentials don’t survive crop and screenshot. Trust layer (journalism fact-checking, KYC, insurance fraud, customs) must be rebuilt.

  Structural shifts:

  Three jobs (research, copy, layout) collapse into one prompt – parallel to word processing for typesetters, browsers for publishers. Image generation is now an agent-callable primitive (not just for humans, but for agents like Codex). An image from Thinking Mode is a compressed reasoning trace: search, plan, composition, verification merge into one artifact.

  Comparison to Claude Design:

  Anthropic shipped Claude Design four days prior (on Opus 4.7, aimed at Figma). While GPT-4o Images retains pixels and adds reasoning upstream, Claude Design skips the image entirely and outputs editable HTML. Both solve the same problem differently: pixel-based vs. code-based. Long-term, they will converge; short to medium term, Images wins on highly visual material (posters, menus, packaging), Claude Design on prototypes (landing pages, dashboards, interactive mock-ups).

  Role-specific implications:

  – Product: UI specification becomes compile step; PM as briefing author matters more

  – Design: Briefs, brand systems, QA grow in value; first-draft execution shrinks; craftsmanship = precise, specific communication

  – Engineering: Images 2 as subroutine for bug reports, pull-request reviews, incident postmortems with annotated screenshots

  – Marketing: Multilingual rendering makes localization vendor obsolete for first drafts (minutes instead of weeks); requires detailed briefing template instead of bullet points

  – Founder/Solo: 5-person agency work now costs $20/month + good brief; leverage lies in creative-ops functions and brief libraries

  – Trust/Risk/Legal: Red-team exercise urgent; new verification chains required (physical certification already exists)

  – Enterprise/CIO: Middleware vendors lose differentiation; middleware pricing (5–10x over API costs) must be renegotiated

  Transferable insight:

  The old bottleneck was model skill (how much you could squeeze out). The new bottleneck is specification—how precisely you can describe layout, typography, constraints, context. That is closer to technical writing than prompting. Practitioners who already think in specs adapt quickly. Those whose value lay in execution craftsmanship must reinvent themselves—not because design dies, but because the foundation has shifted beneath them.

  Core thesis:

  This week, image generation joined the reasoning stack. The output is pixels, but the work is reasoning. That explains why it works. Auditing becomes a different craft: images can be wrong because the web source was wrong, not just because the model hallucinated. The biggest new use cases emerge where you combine knowledge, coherence, or composition that the prompt alone cannot carry.

  Overarching observation:

  The AI race rewards those who understand how to adapt to rapid tool changes. Claude, OpenAI, Gemini releasing so fast that strategy reads: understand where the new tool fits in my workflow → integrate it quickly. That is the whole game.

  Demo and deep dive into OpenAI GPT-4o Images 2 with structural analysis against Anthropic Claude Design; technically adept, intermediate to advanced level.

• Claude Design Does In 30 Minutes What Your Team Does In A Sprint
  24.4.2026, 14:00:38

  Claude Design – 30-Minute Prototyping Instead of Sprints

  Claude Design is the third element in a coordinated Anthropic stack (alongside Claude Code and Co-work) that fundamentally changes traditional mockup-to-production handoff workflows. The tool enables eight main applications: (1) pitch decks with live embedded AI chatbots, (2) animated product explainer videos in code, (3) 3D components and interactive product configurators, (4) automatically extracted design systems from existing codebases, (5) web scraping and reskinning competitor sites, (6) interactive dashboards with live updates, (7) internal admin tools, and (8) mobile app prototypes with real state transitions – each as executable code, not static mockups.

  The decisive difference: Claude Design works in the code medium (HTML, CSS, SVG), not in Figma primitives, because large language models were trained on code. This makes handoff to Claude Code seamless – there are no translation losses anymore. The prototype is already in production format.

  Role-specific impacts:

  • PMs: Prototyping now costs minutes instead of days; PRD documents become less relevant, prototype-first workflow enables discussions about real interactions.
  • Designers: Rationing disappears, ten design directions per hour become routine; focus shifts from mockup creation to judgment about direction and context.
  • Engineers: Handoff already arrives as functioning code; role shifts toward scaling, edge cases, and quality assurance.
  • Founder: Live demos instead of flat screenshots; for AI-native products, real model calls can be embedded in the prototype.

  Organizationally, this leads to smaller teams, as coordination overhead shrinks through fewer handoffs. Two-pizza teams sometimes shrink to one-pizza teams, because roles like PM, designer, and engineer can assume more functions themselves.

  Google Stitch responds with an open design.markdown specification (tokens, typescale, component rules) as a standardized standard, but primarily bets on open source and Gemini quality rather than Anthropic’s stack integration.

  Important caveat: The tool does not replace designer judgment, brand strategy, or positioning – these remain human responsibility. Figma remains relevant for production-grade design systems and component library maintenance. Claude Design is currently strongest on simple to intermediate products; complex systems still hit token budget limits.

  Explicitly covered: Claude Design, Claude Code, Co-work (Anthropic stack), Google Stitch (Gemini), Figma — Format: Deep dive

• Your Apps Don’t Need an API Anymore. Codex Just Proved It.
  23.4.2026, 14:00:05

  Summary: Your Apps Don’t Need an API Anymore. Codex Just Proved It.

  OpenAI fundamentally transformed Codex in April 2025: from a coding command-line tool to a desktop agent that can control any Mac application – through clicking, typing, and running background tasks. The core strength is computer use: Codex sees the screen, navigates through graphical interfaces, and runs multiple agents in parallel without interrupting user input. In direct comparison with Claude, Codex is significantly faster (2 minutes vs. 5–6 minutes for typical workflows), more reliable (makes fewer errors with unexpected dialogs), and more flexible (works with any software, not just Chrome).

  The strategic differentiator: Anthropic builds an “agent body” from structured interfaces (MCP servers, connectors, the forthcoming Conway framework), which requires collaboration with the software industry. OpenAI builds a body from computer use – the agent uses the same graphics interface as humans, requires no APIs, and works with legacy systems, internal tools, and unmaintained apps. This difference is not merely a feature distinction, but a fundamental architectural concept.

  Codex’s capability stems from a team OpenAI acquired in October 2024: Software Applications Incorporated, creators of Sky (an unreleased Mac-OS-AI interface). Core members previously built Workflow (later Apple’s Shortcuts) and brought deep Apple OS integration, accessibility, and motion-path expertise.

  Currently practical: Codex is usable for Slack mass cleanup, Spotify playlist creation via desktop app, visual frontend testing, bug reproduction, cross-app workflows, end-to-end testing, and automation of legacy dashboards without an API. Claude remains superior on scoped knowledge work (focused folder pointing, explicit control) and developer workflows, but Codex dominates on interface-heavy, parallel running tasks.

  To watch: (1) Anthropic’s public announcement of Conway – would establish event-driven agents as a category; (2) MCP adoption speed – with rapid enterprise integration, Anthropic’s ecosystem bet grows stronger, otherwise computer use becomes the standard.

  The central insight: Software without an API is no longer an automation blocker. The difference lies not in the model (that is the “brain”) but in the harness – the “body” that OpenAI built through targeted talent acquisition. Acquisition patterns become a strategic indicator.

  OpenAI Codex and Claude (Anthropic) compared; opinion/analysis deep dive.

• Karpathy’s Wiki vs. Open Brain. One Fails When You Need It Most.
  22.4.2026, 14:01:09

  Karpathy’s Wiki vs. Open Brain: Differences, strengths, and a hybrid approach

  The user compares two fundamental approaches to AI-powered knowledge management: Andrej Karpathy’s wiki concept and his own Open Brain system.

  Karpathy’s wiki approach is based on folders and text files. The central idea: Rather than AI searching and re-discussing all documents with each question, AI continuously writes down its findings. With each new source (article, paper, notes), AI reads it, extracts meaning, and updates existing wiki pages with cross-references, contradictions, and concept connections. The synthesis is compiled once and then updated, not re-derived with each query. Advantage: deep thinking happens at input, then answers are cheap. Disadvantage: AI makes editorial decisions independently; nuances can be lost and remain unnoticed later (like dashboards that compress data).

  Open Brain works on the opposite principle: query-time rather than ingest-time. New information is not immediately synthesized but faithfully stored, tagged, and categorized in structured tables. When a question is asked, AI works fresh and produces a precise answer. Advantage: adding new information is simple and cheap; full detail is preserved and traceable (every claim maps back to sources). Disadvantage: complex queries require AI to synthesize anew each time (higher token cost for repeated questions).

  Where wiki wins: Deep single projects (e.g., reading 10 papers over weeks), personal research, where connections between sources are central. The system flags contradictions immediately upon entry and shows thinking progression over time.

  Where Open Brain wins: Precise structured queries (e.g., “all Q1 meetings with pricing discussions”), multi-agent access to the same data store, high document volumes (thousands of entries), operational data that changes daily. Wiki systems break at scale and multi-agent use (file conflicts) and become misinformation when neglected, since they contain authoritatively written syntheses.

  The hybrid model: The speaker introduces a new plugin approach: Open Brain remains as permanent, structured data source (SQL database). A “compilation agent” runs on-demand or on schedule, reads from Open Brain, dynamically generates wiki pages from a graph of the data, and produces syntheses like a daily briefing. Advantage: wiki browsability and synthesis power from above, reliability and scalability of structured data below. The database remains source of truth; the wiki is regenerated, not manually maintained, avoiding error accumulation.

  Central insight: Choosing a memory system decides not just technology, but the future of team decisions. Contradictions are often valuable (e.g., engineering says 12 weeks, sales promised 8)—a smart wiki might smooth that into narrative, a database system preserves the tension. The deeper point: AI should move from oracle to maintainer—not one-time answer machine, but keeper of knowledge systems that grow over time.

  Recommendation: Solo research and under ~10,000 documents: Karpathy’s simple wiki system. Teams, high document volume, queries across categories, multi-agent workflows: Open Brain + optional graph plugin for the wiki layer on top.

  Comparative deep dive, no specific tool/vendor explicitly covered except the named Open Brain system and conceptual references to Notebook LM and Obsidian as browsable layer.

• Your Prompts Didn’t Change. Opus 4.7 Did.
  21.4.2026, 14:01:24

  Summary: “Your Prompts Didn’t Change. Opus 4.7 Did.”

  Claude Opus 4.7 is the most intelligent model Anthropic has ever released, but also the most confrontational and literal – and despite unchanged pricing, you measurably pay more for the same work. The reason: a new tokenizer that maps the same input to up to 35% more tokens.

  Main improvements: 4.7 fixes the central problem of 4.6 – premature termination on multi-step tasks. Persistence is significantly better, the model self-verifies and runs tests. In benchmarks like SWEBench (80% → 87%), CursorBench (58 → 70), and especially on multi-tool orchestration benchmark (75 → 77), substantial gains emerge. On finance and legal work (GDP-Eval benchmark: 1753 vs. GPT 5.4: 1674), 4.7 leads. However, 4.7 regresses on web research (83 → 79) and terminal tasks.

  Practical tests: A realistic data migration test with 465 files of various formats showed: 4.7 was faster (33 min vs. 53 min for GPT 5.4) and produced a better UI frontend. However, 4.7 hallucinated an edited file (claimed to process it but did not) – a trust violation revealing the central issue: both frontier models fail to recognize obvious fake data (Mickey Mouse, test customer) and rate themselves better than deserved.

  Tonality and literality: 4.7 is 77% more assertive (Code Rabbit analysis), uses imperatives and commands instead of suggestions. The model follows instructions literally – it does not generalize between the lines, as 4.6 did. Adaptive Thinking (invisible to chat users) decides reasoning depth independently. Temperature, Top-P, and Top-K were removed – less user control.

  Claude Design: Anthropic’s new design suite (launched April 17) produces JSX components, design systems, and agent-readable skill files. First sign: logo errors were not fixed even after up to six corrections. The problem: every correction loop is billable. After an afternoon session, costs ran to $42 USD (initial design $5 + corrections $10 + animated videos $23.29). This illustrates a monetization model where unreliability becomes directly expensive.

  Cost implications: The tokenizer increase (up to 1.46x in real measurements) plus higher output burn on “extra high” effort plus corrections mean you pay more despite unchanged sticker prices. Boris Churnney (Claude Code) announced limit increases; Pro subscribers hit ceilings immediately.

  Strategic context: Anthropic IPO talks for October targeting $60B fundraising; valuation at $800B. Revenue annualized to ~$30B (April). Enterprise market share: 24% → 30% (OpenAI drops 46% → 35%). Mythos too risky to release (government/large firms only). OpenAI released largest Codex update in parallel, GPT model “Spud” expected this week – Anthropic under pressure to release 4.7 to avoid losing frontier position.

  Usage recommendations:

  • Claude Code: Extra High as default, Max for heaviest; use plan mode.
  • API: Remove temperature/top-P/top-K (→ 400 errors); set thinking display to “summarized”.
  • Chat/Claude.ai: Explicitly request reasoning; upload context; start fresh chats if context gets polluted.
  • Upgrade decision: Agentic pipelines/finance/legal work = yes; web research/terminal heavy = no; casual chat users = complex (prompting becomes more demanding).

  Larger trend: Anthropic builds vertically (design, code, review, deploy), OpenAI horizontally (Codex as desktop platform). Model makers compete not only on models but on custom harnesses. Serious, long-term knowledge work gets massive tokens; casual use gets neglected. The model is a strategic moving target on the path to enterprise coworker – a bridge release under competitive pressure.

  Claude (Anthropic) was explicitly covered; also GPT 5.4 and Gemini 3.1 for comparison. Tools/features: Claude Code, Claude Design, Anthropic Labs, Codex (OpenAI). Format: Deep dive with practical tests and strategic analysis. Difficulty level: demanding – requires technical understanding (tokenizer, API, agentic workflows).

• Nobody Knows What You’re Worth Anymore | The AI Job Market Reality
  20.4.2026, 14:00:01

  Summary: Nobody Knows What You’re Worth Anymore | The AI Job Market Reality

  The central problem is that traditional signals for professional competence are collapsing: since AI tools make code and project generation virtually free, it is hard to distinguish genuine expertise from mere output. This applies not just to entry-level professionals but to all career stages – from college graduates to mid-career PMs. Context shifts dramatically through massive tech layoffs (over 60,000 confirmed cuts in Q1), which are no longer about overstaffing but about reassessing what workers are worth in the age of AI.

  The author proposes five guiding principles:

  1. Understanding before generation: Force yourself to deeply understand your work – not just what code does, but why, what tradeoffs you made, what could break. This is the origin of “taste” and expertise. The industry produces at unprecedented speed but understands at unprecedented low levels.
  1. Explanation artifacts: Accompany every project with structured, concise explanation – what this is, why this choice, which alternatives, where it is fragile, what you learned. This is like a “commit message” for the AI era – it signals genuine understanding and cannot be faked by AI.
  1. Transactions over credentials: Credentials are easy to fake. Real work, paid for, is a real signal. In the AI age, we need “micro-transactions for jobs” – proof of meaningful work done in compressed time, not just classical two-years-per-role careers.
  1. Work in public: If you do not work publicly, you lose visibility. The old model of internal learning inside companies does not work for outsiders or the laid-off. GitHub alone is not enough; you need a central place to show your “glorious mess” of AI artifacts.
  1. Provide proof that travels with the work: Your proof – the explanations, your thinking – must be inseparable from the work itself. Otherwise, it is assumed AI generated everything. This also prevents spam.

  The author created “Talent Board” – a platform where projects can be shown publicly with the four explanation questions to demonstrate thinking, not just generation. He emphasizes that the principles also apply to personal websites or competing platforms – the goal is solving the problem, not gatekeeping.

  The central thesis: In a world tempting you to increase your output 100x, you must show that you can think. That is now the rare commodity.

  Mentioned: Lovable, Claude, ChatGPT, GitHub, Talent Board (author’s); format: opinion/reflection with framework.

• Block Laid Off Half Its Company for AI. AI Can’t Do the Job.
  19.4.2026, 17:00:56

  Block Laid Off Half Its Company for AI. AI Can’t Do the Job.

  The video criticizes the hype around so-called “world models” – automated systems meant to handle management tasks like status updates, prioritization, and information flow. While the basic idea makes sense (automating pure information logistics), the creator warns of a fundamental error: these systems confuse information flow with judgment.

  The core problem is that managers do not just relay information; they also evaluate it – deciding what matters, distinguishing noise from signal, recognizing seasonal fluctuations. When a system makes these judgments silently, errors go undetected: a seasonal revenue dip is treated as a strategic problem, a correlation between feature launch and churn is read as causation, or information flow decays so slowly that nobody notices critical signals are missing.

  The video analyzes three common architectural approaches and their weaknesses:

  1. Vector-database approach fails because semantic ranking decisions look like interpretations but are presented as mere information – at scale, everything the system surfaces is automatically treated as actionable.
  1. Structured-ontology approach (Palantir model) is too conservative – the ontology can only represent predefined relationships and misses emergent patterns that could reframe business decisions.
  1. Signal-fidelity approach (Dorsey’s thesis: high data quality like transactions) creates false confidence – clean input data suggests reliable judgments, even when causal logic underneath is thin.

  The creator offers principles for better world models: clear “interpretive boundary” between facts (which the system should surface) and judgments (which humans must make); signal quality as an upper bound; structure through learning rather than predefinition; encode outcomes to close feedback loops; design for resistance (teams must want to feed the system); and start fast, because continuous data over time is competitive advantage.

  Different strategies by company type: small, agile teams can start with vector databases, regulated companies need structured ontologies, platforms like Block must guard overconfidence through clean data, knowledge-work companies should build interpretation layers above vector databases. Critical: “The most dangerous world model is one that works well enough that nobody questions it, until decision quality decays imperceptibly.”

  The creator built a “World Model Readiness Plugin” available for various LLMs that guides companies through an assessment framework.

  Explicitly mentioned tools/vendors: Claude, ChatGPT, Gemini, Palantir; primarily opinion/reflection on system architecture with practical framework.

Alejandro AO (2 new videos)

• GPT-5.5 – 25x More Expensive than Open Models
  23.4.2026, 23:13:53

  The video provides an overview of OpenAI’s latest version, GPT-5.5, and compares it with other models, both closed and open-source. The creator emphasizes that GPT-5.5 delivers the best results across various benchmarks, particularly in coding tasks like Terminal Bench and SweetBench Pro. However, GPT-5.5 is significantly more expensive than its predecessor GPT-5.4 and other models such as Claude Opus 4.7 and Gemini 3.1 Pro. A comparison with open-source models like Kimi, GLM, Minimax, and Qwen 3.6 shows that these often provide similar performance at substantially lower costs. The creator also demonstrates using GPT-5.5 in the Codex application to create a landing page, showing that the model has improved in frontend development.

  The video explicitly covers OpenAI (GPT-5.5, GPT-5.4), Claude (Claude Opus 4.7), Gemini (Gemini 3.1 Pro), and open-source models (Kimi, GLM, Minimax, Qwen 3.6), and is aimed at intermediate to advanced users.

• Kimi K2.6: Open Source BEATS GPT-5.4 & Opus?
  22.4.2026, 06:17:00

  The video introduces Kimi K2.6, a new foundation model from Chinese company Muncho Taiyai, which competes with benchmarks like GPT 5.4 and Opus 4.6. The host explains what Kimi K2.6 is, compares benchmarks, and demonstrates how to use the model with Hugging Face Inference Provider to run a coding assistant. Subsequently, Kimi K2.6 is used to redesign a website. The benchmarks show that Kimi K2.6 outperforms GPT 5.4 and Opus 4.6 in several areas, particularly in general knowledge and programming. The host demonstrates setting up Kimi K2.6 with OpenCode and Hugging Face to redesign the website, showing the final result. The cost of using the model is estimated at around $1.50 for the entire operation. The video explicitly covers Kimi K2.6, Hugging Face, OpenCode, GPT 5.4, Opus 4.6, and Gemini 3.1 Pro, and is aimed at intermediate to advanced users.

Alex Finn (5 new videos)

• ChatGPT 5.5 Codex: I can’t believe they did this…
  25.4.2026, 20:54:45

  ChatGPT 5.5 Codex vs. Claude Opus 47: One Year After the Regime Change

  The creator has extensively tested ChatGPT 5.5 Codex and concluded that it has dethroned Claude Code after more than a year of dominance. The four main improvements of 5.5 over 4 are: significantly greater action-orientation (executes instructions rather than discussing), less verbosity, better personality, and markedly superior code with minimal bugs.

  However, the centerpiece is the new Codex Desktop App, which delivers a “super app” experience: integrated project management, separate chat and coding sessions (multitasking), built-in browser for live app editing, annotation feature for direct UI element markup, and computer-use feature for fully automated app testing. Particularly noteworthy: deep integration with Codex’s own image generation model, which is the best available image generator and allows UI design visualization before coding – Claude Code has no image generation.

  A critical advanced workflow combines Codex with Linear (free), a project management tool that Codex can fully control. Multiple agents then work in parallel on the same tasks, while Codex automatically creates, organizes issues, and tracks progress. This eliminates manual reminder-reading between sessions.

  Strengths of 5.5: Better overall model, consistently intelligent (doesn’t get dumber like Claude), extremely high rate limits (reset daily on bug fixes), Codex app beats Claude app, image generation integration. Weaknesses: Default UIs still have “slop,” less warm personality, sometimes overly verbose in responses.

  Strengths of Opus 47: Better standard UI quality, superior personality/warmth, more incisive and proactive. Weaknesses: Model degrades significantly over time (Anthropic lacks sufficient compute capacity), rate limits drop continuously, Claude Code removed from $20 plan – all suggesting compute bottlenecks.

  Verdict: ChatGPT 5.5 Codex is now the better choice for coding. The speaker emphasizes not being locked into any company and switching immediately if that changes. A follow-up video on open-source Claude is planned.

  OpenAI (ChatGPT 5.5, Codex, Image Generation Model), Anthropic (Claude, Opus 47), Linear — Demo/Review with advanced workflow focus.

• 🔴LIVE: ChatGPT 5.5 is here. Does it beat Claude Opus 4.7?
  24.4.2026, 20:24:30

  Summary: ChatGPT 5.5 vs. Claude Opus 4.7 – Live Testing

  The streamer conducts comprehensive live tests comparing ChatGPT 5.5 against Claude Opus. The test format consists of three parts:

  Benchmark Tests (so-called “Alex Finn Benchmarks”):

  1. 3D First-Person Shooter Test: ChatGPT 5.5 achieves 9.1/10 – beating Opus’ previous record of 9.0.
  2. Elon Dancing Test: ChatGPT 5.5 performs significantly worse (approx. 7.8), while Opus scored 86/100. The streamer criticizes lack of detail accuracy.
  3. City Flythrough Test: ChatGPT 5.5 reaches 7.5/10 (Opus: 83). Criticism: missing window details on buildings.
  4. Music Visualizer Test: still being conducted.

  AI Agent Olympics – Three competitors are tasked with designing a complex video game, building it, and creating marketing poster:

  • OpenClaw with Opus 4.7
  • OpenClaw with ChatGPT 5.5
  • Hermes with ChatGPT 5.5

  Result: Opus and ChatGPT 5.5 (both OpenClaw) deliver high-quality posters and game ideas; Hermes lags in communication. The streamer recognizes a slight advantage for Opus posters (minimalist design) but calls it a tie.

  Critical Observations:

  • OpenClaw updates are buggy and error-prone; ChatGPT 5.5 integration causes technical issues.
  • ChatGPT 5.5 is stronger on pure code generation tasks (“self-testing”) than Claude.
  • Hermes shows communication problems and often waits unclearly for instructions.
  • ChatGPT 5.5 criticized as “wordy” but sometimes delivers better graphic quality.

  Stream conclusion so far: Close call – ChatGPT 5.5 clearly wins the shooter test but loses massively on the Elon dancing test. In practical agent use, Opus remains competitive.

  Tools mentioned: OpenClaw, Hermes, Claude Opus, ChatGPT 5.5, Codex (ChatGPT app), DeepSeek V3 — Format: Live demo with comprehensive technical tests and honest critique.

• The best Claude Design workflow you’ll ever see…
  22.4.2026, 22:12:45

  Claude Design + ChatGPT Workflow for Professional App Design

  The creator presents a system combining two tools to create high-quality app designs: ChatGPT’s new image generation model for exploration and Claude Design for final execution.

  The problem with individual tools: Claude Design has very low usage limits (weekly limit exhausted after 2–3 designs), while ChatGPT’s image generation offers unlimited possibilities but doesn’t produce true design systems. The solution: exploration and experimentation phase in ChatGPT, final execution in Claude Design.

  The 8-step process:

  ChatGPT Phase (4 steps):

  1. Gather inspiration: Source logos and designs from platforms like X and save as references (4–5 ideal)
  2. Exploration: Describe in ChatGPT what you’re building (e.g., “AI agent that builds businesses”) and explicitly ask the image generation model to create logos (e.g., with animal motifs like octopuses or owls)
  3. Refinement: Iterate 5–6 times – select preferred designs each time and request variations, gradually moving toward the final direction
  4. Branding guide & prompt: Have ChatGPT generate a final branding guide (colors, typography, merchandise) and create a structured prompt for Claude Design

  Claude Design Phase (4 steps):

  1. Set up design system: Enter project info (name, description, optionally GitHub code, the original ChatGPT prompt, and upload final assets)
  2. Generate system: Claude creates a comprehensive design system with components, spacing, colors, typography
  3. Review & feedback: Check generated assets and make corrections if needed
  4. Build prototype: Create a high-fidelity prototype of the dashboard/app interface with all assets, then hand off to Claude Code

  Result: Instead of experimentally overloading Claude Design, use ChatGPT’s high limits for creative exploration, bring filtered results to Claude, and achieve professional, high-quality designs in 1–2 turns instead of many (saves limits and time).

  Tools/Providers: Claude Design, ChatGPT (Image Gen Model), Claude Code — Tutorial/Demo with practical execution of a complete workflow.

• 🔴LIVE: OpenClaw 4.20 OUT!! Is ChatGPT finally better than Opus???
  22.4.2026, 20:19:39

  Summary: OpenClaw 4.20 Live Test – ChatGPT vs. Opus

  The streamer tests the newly released OpenClaw version 4.20 live and checks whether ChatGPT has finally become usable in OpenClaw. Previously, ChatGPT in OpenClaw was unusable per his account – it got stuck in planning mode and didn’t complete tasks, while Claude/Opus reliably saw tasks through to completion.

  The stream plan: A new OpenClaw agent named Jason is connected with ChatGPT 5.4 via API and tested directly alongside Henry (with Opus) to compare both models on various tasks. The setup, however, is delayed by technical problems – gateways hang, and setting up multiple agents in OpenClaw proves error-prone.

  Meanwhile, the streamer presents two logos for his new project “Henry Intelligent Machines” (an agent swarm for which he recently raised funding): a pixelated octopus and several owl variants. The chat is divided – the octopus logo is chosen because of the metaphor of eight tentacles working simultaneously as symbols for parallel agents, though many prefer the owl logo. In between, there’s discussion of the Vibe Coding Academy (1,500 members), an upcoming Claude Design tutorial, and a potential acquisition of a $100,000 ASUS supercomputer (with GB300 Grace Blackwell chip).

  The streamer also criticizes Anthropic’s communication weaknesses – unclear remain the exact rules for Claude Code, OpenClaw support, and other policies that should simply be made transparent.

  Featured AI models: Claude (Opus), ChatGPT, Claude Design (mentioned as upcoming); Format: Live stream (demo/comparison test with extensive discussion and community interaction).

• OpenClaw Full Tutorial: Set up your first AI employee!
  21.4.2026, 01:05:23

  OpenClaw Full Tutorial: Set up your first AI employee

  Overview: OpenClaw is an autonomous AI agent that runs as a personal digital employee 24/7 on your own device. It differentiates itself from alternatives through its open-source nature (data stays local), full customizability, and model flexibility.

  Hardware & Hosting:

  • OpenClaw should NOT run on a VPS but locally on any device – even old laptops are sufficient
  • Mac Mini is recommended but not mandatory
  • The creator uses three Mac Studios, two Mac Minis, and a DJX Spark, but that’s optional
  • Recommendation: Scale slowly, start with existing hardware

  Model choice (Brain & Muscle approach):

  • Orchestrator (Brain): Claude Opus 4.7 – most reliable choice for agent coordination via API (despite Anthropic restrictions for subscriptions)
  • Coding (Muscle): ChatGPT 5.4
  • Research & Writing (Muscle): Gemini, Grok 2.6, or other cost-effective models
  • Local Models: Qwen 3.6, GLM 5.1 – free, private, ideal for repetitive tasks like web scraping

  Communication:

  • Telegram is optimal: simple bot setup, group chats with separate topics for context organization and cost savings
  • Discord for automated workflows, not for main chats

  Develop personal use cases:

  1. Brain dump about yourself (businesses, metrics, goals)
  2. For one day, note all manual computer tasks
  3. Ask OpenClaw: “Which workflows could help me?”

  Creator’s example applications:

  • Linear integration for automated Vibe Coding (task assignment, code generation, PR creation)
  • Content scouting: monitor Twitter/YouTube posts for trends, automatically write scripts
  • Stock research: daily reports on AI investment opportunities

  Multi-agent approach:

  • Use OpenClaw and Hermes in parallel – the creator sees both as equivalent, uncritically, each with its own advantages
  • Benefit: one fixes the other if updates cause bugs
  • Hermes slightly faster and more resource-efficient; OpenClaw with better long-term memory

  Mission Control:

  • Custom software platform for own tools and workflows
  • Don’t simply copy from others – should be personally tailored
  • OpenClaw can build new tools in Mission Control itself

  Security:

  • Main factor: good judgment in crafting prompts
  • OpenClaw does only what you tell it – no more, no less
  • Precise instructions and consideration of consequences are critical

  Featured: OpenClaw (open-source agent), Claude/Opus, ChatGPT, Gemini, Grok, local models (Qwen, GLM), Hermes agent, Claude Code, Linear, Discord, Telegram; Format: Tutorial with demo elements and personal use cases.

Andrej Karpathy

No new videos in this period.

Bart Slodyczka (2 new videos)

• Claude Cowork + Ollama = 100% FREE & PRIVATE (Full Setup)
  25.4.2026, 11:55:24

  The video demonstrates how to use Claude Cowork for free and privately by leveraging local AI models on your own computer or accessing OpenRouter to use cloud-based models. The author showcases the Gemma 26B model (18 GB, 256K context window) and the smaller Gemma 2B model (7.2 GB). Ollama is used for local deployment to download and manage models. If your computer isn’t powerful enough, you can use OpenRouter, which provides access to hundreds of AI models, though with potential privacy concerns. The author shows how to connect the Claude Desktop App with Ollama and Tailscale to establish a secure HTTPS connection. Next, the video explains how to set up Skills, Connectors, and Plugins in Claude Cowork, including integration of Brave Web Search. The author tests both local and cloud-based models, finding that cloud models are faster but may not interact as well with the app’s Skills. The video is suitable for intermediate users and covers Claude, Ollama, OpenRouter, Tailscale, and specific tools like LM Studio.

• Full Tutorial: How To Automate 99% Of Your Work With Claude Cowork
  22.4.2026, 13:03:14

  The video shows how to use Claude Cowork as a virtual assistant to automate 99% of your tasks. It starts with installing the Desktop App and choosing a paid plan, with the Pro Monthly option at $20 recommended. The user is guided through the app’s features, including the differences between Claude Chat, Claude Cowork, and Claude Code. Claude Cowork can access local files, work with Excel and Word documents, control your browser, and connect to tools like Gmail or ClickUp.

  The main focus is on using Claude Cowork effectively by changing how you communicate with it. Rather than treating it as a chatbot, you should view it as a virtual assistant capable of handling complex tasks like writing emails or managing tasks in a task manager. Users are instructed to create projects and tasks, where projects serve as folders on your computer storing all resources and generated files. The video emphasizes that providing context and background information is crucial to maximizing your assistant’s effectiveness.

  The video also demonstrates how to create Scheduled Tasks to automate recurring activities, such as searching for the cheapest flights and recording prices in an Excel spreadsheet. Finally, the use of Live Artifacts is shown to display data from an Excel table in an interactive dashboard.

  Final note: The video explicitly focuses on Claude (Cowork, Chat, Code) and is aimed at intermediate users.

Ben AI (1 new video)

• How to Use Claude Routines Better than 99% of People
  25.4.2026, 09:12:01

  The video explains Claude’s new Routines feature, which enables you to automate recurring tasks without requiring technical expertise. It showcases three main use cases for Routines: a daily check for failed customer payments, creating personalized offers after a sales call, and a churn recovery routine. Routines can be triggered either on a schedule or by events and run autonomously in the cloud. Building a routine consists of four elements: prompt, trigger, repo and files, and connectors. The video also discusses current limitations of Routines, such as the limited number of webhook triggers and the need to store files and skills in a GitHub repository. The creator demonstrates how to set up a routine from scratch, including setting up a GitHub repository and using N8N to trigger the routine on a specific event. It emphasizes that creating reliable routines requires previously building and testing skills.

  The video specifically covers Claude, GitHub, and N8N and is aimed more at intermediate users.

Brian Casel (1 new video)

• Where Claude Design actually fits
  24.4.2026, 12:00:29

  The video discusses the suitability of Claude Design for real-world workflows in product development. The author tests the tool across various use cases and arrives at mixed results.

  First, the use of Claude Design for creating a new T-shirt marketplace is demonstrated. A high-quality mock-up is created that looks appealing but is difficult to integrate into existing workflows in practice. Export functions like PDF, Canva, HTML, or handoff to Claude Code are deemed insufficient for professional product development. The missing planning and specification process that would be necessary for successful implementation is particularly criticized.

  Another use case is extending an existing website with new pages or features while maintaining the existing design system. Here too, it becomes clear that while Claude Design can analyze and create a design system from a GitHub repository, integration into the existing codebase is problematic. The author instead prefers working directly in Claude Code with a custom design system that is integrated into the codebase.

  The author sees real breakthroughs in using Claude Design as a visual planning tool in the early ideation phase and for creating marketing assets. For creating a metrics dashboard, the video demonstrates how a mock-up can be created in Claude Design and then used as a basis for detailed product planning in Claude. This particularly facilitates entry into product development for inexperienced users.

  Another practical application is creating animated videos and presentations that align with brand design. Claude Design offers specific templates for animations and slides that can be used directly.

  In summary, Claude Design is rated as useful for the early planning phase and creation of marketing assets, but not as a complete tool for product development from idea to finished product.

  Final note: The video explicitly addresses Claude Design and Claude Code and is aimed at intermediate users who already have experience with AI tools.

Coding with Lewis (1 new video)

• What a Billion Database Rows Look Like in Real Life
  24.4.2026, 15:10:21

  The video illustrates the physical size of databases by converting the number of rows into stacks of paper. It starts with 50 rows on a single sheet of paper and scales up to a billion rows, which would form a stack taller than any building in Canada. The comparison shows that a database with a million rows would be approximately two meters high, while a database with 100 million rows would be about as tall as the Eiffel Tower. The video explains how databases like Postgres use indexes to query data quickly, comparing this to manually searching through stacks of paper. Real-world examples such as Spotify, Uber, and Facebook are mentioned, which generate enormous amounts of data daily. The video emphasizes the efficiency of databases compared to physical storage media and promotes Supabase as a platform for leveraging Postgres.

  The video explicitly focuses on Postgres and Supabase and is aimed more at Intermediate and Advanced levels.

Cole Medin (4 new videos)

• Pushing My Dark Factory Further with Kimi K2.6: A Codebase That Writes Its Own Code, Live
  25.4.2026, 17:29:32

  Summary: Pushing My Dark Factory Further with Kimi K2.6

  The stream showcases a live session where the creator upgrades his “Dark Factory” project—a self-managing codebase—with the new Kimi K2.6 language model.

  Setup and Motivation:

  The creator switches from Miniax M2.7 to Kimi K2.6 because the latter benchmarks better and is significantly cheaper than Claude Opus or GPT—roughly priced like Claude Haiku but with superior performance. With Kimi K2.6, he gets Archon-based workflow automation for a RAG AI tutor system built on YouTube content.

  Dark Factory Architecture:

  The project uses Archon, an open-source workflow framework, to automate the entire software development lifecycle. It features specialized workflows for issue triage, planning+implementation+validation, pull request reviews, and weekly regression tests with browser automation. Everything runs on a remote VPS—Claude Code orchestrates locally while Archon manages remote workflows.

  Live Setup:

  The creator generates a Kimi API key and configures Claude Code with a custom endpoint to use Kimi’s Anthropic-compatible API. After setup, he creates GitHub issues through a “brain dump” (collaboratively with Claude) for concrete problems: too many cited sources (47 instead of ~10), admin panel UI gaps, modal layout issues when viewing videos, missing search function in the video library, poorly designed login page, and unwanted “new conversation” entries in the history.

  Initial Results:

  The Archon triage workflow runs immediately and correctly labels all issues (feature vs. bug, priority, status “factory-accepted”). However, the creator notices Kimi usage metrics aren’t updating—likely because workflows are still using Miniax instead of Kimi, so he needs to log this.

  Side Topics:

  The creator discusses Archon as a harness builder (not the harness itself), emphasizes under 5% token overhead from Archon’s design, explains permission to use Anthropic subscription with Claude Agent SDK in Archon, and showcases Archon community projects (e.g., Remotion from Hacker News for video generation). Archon is approaching 20,000 GitHub stars.

  The format is a live-coding demo with an experimental character—not production-ready code approach, but an exploration of modern AI coding assistant capabilities.

  Tools/Models: Kimi K2.6, Claude Code, Miniax M2.7 (previous), Archon (open-source workflow framework), Claude Opus 4.7 as reference — Demo/Live Coding.

• Parallel Claude Code + Git Worktrees: This Setup Will Change How You Ship
  23.4.2026, 00:01:05

  Summary: Parallel Claude Code + Git Worktrees

  The video presents a system for parallel agent-driven development with up to 10x higher output. The strategy rests on five pillars:

  The Five Pillars:

  1. Issue as Specification: GitHub issues or tickets serve as structured input for implementation. The creator batches multiple issues and then sends each to a separate agent.
  1. & 3. Isolation via Git Worktrees: Each agent works in its own Git worktree (achievable with Claude Code via claude -w [name]), preventing parallel changes from conflicting. This ensures agents don’t overwrite each other.
  1. Validation in Separate Context Window: The agent doesn’t review its own code in the same session where it was written (too much bias). Instead, a new context is opened via a review PR command, which analyzes the pull request and deploys sub-agents for specialized checks. Additionally, external tools like Codeex can serve as extra reviewers.
  1. Self-Healing Layer: Bugs don’t just get fixed—they improve the overarching AI systems (rules, skills, workflows, claude.md) so similar errors don’t recur.

  Practical Solutions for Scaling Issues:

  • Port Conflicts: A startup script assigns each worktree a unique port (e.g., based on worktree name).
  • Dependency Installation: Node modules are installed upfront, not during agent session.
  • Database Isolation: Neon branches (or SQLite per worktree) ensure parallel database changes don’t collide—each branch has its own data copy.
  • Token Efficiency: Model switching (e.g., to Haiku/Sonnet for cheap tasks like code analysis) reduces costs over longer workflows.

  The workflow: Issue → parallel planning in 5 Claude sessions → parallel implementation → pull requests → separate reviews with review PR → merge. The repository with scripts and examples serves as a blueprint for adoption into your own systems.

  AI Tools: Claude Code, Codeex plugin, Neon (database branching), Archon (mentioned open harness builder); Format: Deep-dive with practical demo.

• The AI Dark Factory is ALIVE: A Codebase That Writes Its Own Code, Live
  19.4.2026, 04:39:47

  The AI Dark Factory is ALIVE: A Codebase That Writes Its Own Code, Live

  The streamer showcases the Dark Factory—a self-writing code system that operates without manual code review. The project is a public experiment where a RAG chat interface over YouTube content is automatically extended by AI agents.

  How It Works: Issues are filed as GitHub tickets and trigger automated Archon workflows on a VPS. These workflows triage issues (group by priority), implement fixes, validate them with browser automation, and merge autonomous pull requests. The entire process runs on fixed schedules (currently every 30 minutes). The streamer himself writes no code lines and reviews nothing—everything happens through GitHub issues as the input interface.

  Archon is the open-source harness that packages arbitrary AI coding processes into YAML workflows and can execute them in parallel. It abstracts away the entire orchestration.

  Costs & Models: The streamer uses Miniax M2.7 via API (30¢ per million input tokens), not Anthropic API, to avoid rate limits—already ~$55 spent on millions of tokens during testing. He considers a Miniax subscription (15,000 model requests per hour) and discusses alternatives like GLM 5.1, Qwen 3.6, and others—all trivially swappable via environment variables without code changes.

  Validation: Weekly regression testing workflows automatically test all user journeys and file new issues for discovered bugs—a self-reinforcing loop.

  Configuration: Rules for triage and priorities live in a factory-rules.md—the streamer maintains this as a lever to steer the factory without reviewing individual implementations.

  Practical Demo: The streamer identifies UI bugs live (missing vertical centering, missing auto-scroll, missing chat auto-naming), dictates them verbally, Claude Code auto-creates GitHub issues, the orchestrator is manually triggered, triages run, workflows start—budget visible in real-time.

  Discussion touches on enterprise adoption (slow, lots of red tape vs. startup speed), token consumption as a metric, and the thesis that true reliability is still years away—this is an experiment at the boundaries, not production standard.

  Closing: The streamer plans to work live on Archon itself too, such as building workflows for issue replication, and considers testing the Dark Factory across multiple models.

  Tools & Context: Miniax M2.7 via API, Claude Code with Archon skill, GitHub CLI/issues, browser automation—deep-dive demo with perspective on hype and enterprise reality.

• Pi Coding Agent + Archon: Build ANY AI Coding Workflow (No Claude Code Bloat)
  20.4.2026, 21:00:28

  The video introduces Pi, a minimalist open-source coding agent designed as a counterweight to Claude Code. Claude Code initially launched with a simplicity promise but has become bloated and buggy after numerous feature updates, with a constantly changing system prompt that’s hard to control. Pi follows the principle: build a small, unbloated core and make it easily extensible. Instead of baking features hard into the agent, you can simply ask Pi to build new features into itself; the agent modifies itself quasi-dynamically. There’s an extension ecosystem with skills, prompts, themes, and extensions you can add with a command—or build your own.

  Pi supports any major language model (GPT, Gemini, Mistral, Grok, etc.) and installs with a simple npm command. After that, just /login for API config and /model to choose your model—done.

  The real focus is Archon + Pi: Archon is a harness builder for AI coding workflows, letting you pack complex agentic engineering processes into reusable, parallel workflows. Pi was added as the third supported agent after Claude and Codecs. The speaker demonstrates a Plan-Implement-Validate workflow with Pi’s Planotator extension: during planning, Pi creates an interactively editable markdown document on a website where you can provide inline feedback. The feedback is automatically addressed without separate context switches. Then the workflow moves into implementation (with different providers mix-and-matchable) and validation.

  In the live demo, they build a light/dark theme toggle for a GitHub issue triage app: clarifying questions → Planotator plan review with inline feedback iteration → implementation → validation. The final toggle works in the browser, including browser automation testing with the agent browser CLI.

  Explicit Tools/Providers: Pi (open-source), Archon (open-source harness builder), Claude Code, Codecs (OpenAI), Anthropic, Gemini, Agent Browser CLI, Planotator extension — demo with practical workflow implementation.

Datapizza (1 new video)

• CLAUDE DESIGN e OPUS 4.7: che ne pensiamo?
  21.4.2026, 16:00:00

  Claude Design and Opus 4.7: Together they expand the Anthropic ecosystem

  Anthropic had a particularly intense week: on April 16 it launched Opus 4.7, the new version of its flagship model, optimized for agentic workflows and software engineering. Compared to Opus 4.6, the model improves in instruction comprehension, vision, and persistent memory management, enabling longer-duration complex tasks to be carried out with greater autonomy. However, a problem emerged: despite Anthropic’s promise of improvements in instruction following, in tests the model ignored explicit constraints while creating solutions. An appreciated innovation is support for high-resolution images, useful for analyzing dense diagrams and complex charts. Opus 4.7’s new tokenizer generates significantly more tokens compared to Opus 4.6 (49 input tokens vs. 37 in the previous model) and also compared to GPT: this increases costs but improves granularity and robustness in handling rare words. Anthropic also added a new reasoning feature positioned between standard and maximum levels, to balance analytical depth and latency. The remote control function has also been improved.

  Claude Design, announced the same week, is a beta tool for creating front-end interfaces (mobile applications, websites, presentations). It was successfully tested to create a booking application for an osteopathic clinic: the model generated a minimalist interface consistent with the provided logo and assets. Claude Design allows export to HTML, PowerPoint, PDF, and Cursor Code, seamlessly integrating design with backend development in Cursor Code. A useful feature is that the tool automatically asks clarifying questions before starting work, a technique similar to guided brainstorming, reducing the gap between what the user imagines and the final output. It’s also appreciated for separating Claude Design usage from standard usage, avoiding impact on users’ weekly token budgets. Anthropic continues to consolidate native integration into enterprise systems and the comprehensive integration of its tools.

  In the cybersecurity field, Anthropic continues the Glasswing project, providing security professionals with a preview of Claude Mutos to identify vulnerabilities in critical software (healthcare systems, banking). The model has already discovered bugs latent for years. OpenAI responded with the launch of GPT 5.4 Cyber, a variant of the 5.4 model dedicated to defensive security, with reduced restriction thresholds for legitimate security requests. OpenAI has formalized a two-pronged strategy: expansion of the TAC program (Trusted Access for Cyber) and the launch of this model. OpenAI’s approach shifts focus from periodic audits to continuous real-time monitoring, validating vulnerabilities and proposing fixes as code is written. However, critical issues remain: OpenAI can limit or revoke access through third-party platforms if there is no visibility into usage context, and tension persists between stated openness and actual control. In the short term (3 months), similar models are likely to appear in open source on Hugging Face, with both positive implications (breach prevention) and negative ones (exploitation by malicious actors).

  The Stanford AI Index Report tracks global AI adoption from 2017 to today. Generative AI has reached 53% global adoption in just 3 years, a pace faster than the internet and personal computers. However, distribution is uneven: Singapore and the United Arab Emirates exceed 60%, while the USA stands at 30%, contrary to expectations given dominant American investments. A surprising finding: the flow of researchers to the United States has collapsed 89% since 2017, suggesting a loss of ground in talent attraction despite leadership in private investment. A paradoxical aspect highlighted in the report: models win math olympiads but can’t tell the time — this reflects the need for agency to access external tools. On the transparency front, the Foundational Model Transparency Index has recorded a collapse: models are more powerful but less governable because we cannot exactly trace the origin of outputs. A critical data point concerns global expectations about jobs: Nigeria, Mexico, and Japan see job creation, while Italy has the most negative sentiment — 59% believe AI will eliminate jobs and destroy companies. This suggests an urgent need for adoption (not just training): guiding organizations and people to understand how to use AI well to achieve real benefits.

  Final flash news: Qwen 3.6 is the current model of the moment on Hugging Face, already favorably compared with Opus 4.7. Google invests 2 million dollars in Italy with the “AI Works for Italy” project, training university students at pilot universities (Roma 3, University of Salerno, University of Sassari) in AI skills recognized as necessary in the job market.

  Models and tools mentioned: Opus 4.7, Claude Design, Claude Mutos, GPT 5.4 Cyber, Qwen 3.6, Cursor Code — Format: Opinion/Reflection with News Roundup.

Dave Ebbelaar (1 new video)

• Programming Has Changed – Here’s What I Focus On Now
  23.4.2026, 18:05:51

  The video discusses the future of tech jobs and the relevance of current skills in a rapidly evolving environment, particularly in relation to artificial intelligence (AI). The speaker emphasizes that software engineers, AI engineers, and data scientists will continue to be needed, however the requirements and ways of working will change. He introduces three concepts that describe how companies utilize AI: AI-enabled, AI-first, and AI-native.

  1. AI-enabled: Companies use AI tools to increase employee efficiency without fundamentally establishing new processes.
  2. AI-first: Companies optimize their processes and systems around AI, with AI agents playing a central role.
  3. AI-native: Companies are built from the ground up with AI agents, with minimal human intervention.

  The speaker advises focusing on the ability to help companies transition from AI-enabled to AI-first and eventually to AI-native. This requires broad knowledge of the entire technology stack (full-stack skills) as well as the ability to analyze and automate business processes. He emphasizes the importance of not only possessing technical skills but also being able to identify and optimize business processes that are suitable for automation.

  The speaker sees the future of tech jobs in the ability to develop end-to-end AI automation and support companies in transitioning to AI-driven processes. He recommends positioning yourself as a freelancer or consultant to help small and large companies implement AI solutions.

  Final comment: The video addresses general AI developments and is aimed more at intermediate and advanced tech professionals.

David Shapiro (3 new videos)

• Techno-feudalists hate UBI
  23.4.2026, 16:55:48

  The video discusses the attitude of the so-called “Tech Right” towards universal basic income (UBI) and presents their arguments against UBI before refuting them. The “Tech Right” is described as a group of Silicon Valley types who are often venture capitalists or founders and tend to be pro-MAGA, pro-Trump, and pro-military-industrial complex. Notable representatives include Mark Andreessen and Gil Verdon.

  The main arguments of the Tech Right against UBI are:

  1. UBI flattens incentive gradients, which reduces motivation for major goals.
  2. UBI prevents the maximization of entropy generation, i.e., the creation of useful new information.
  3. UBI hinders civilization’s ascent on the Kardashev scale.
  4. UBI leads to state dependency and is therefore harmful.

  The author refutes these arguments as follows:

  1. Many of history’s greatest entropy generators, such as Bill Gates, Steve Jobs, and Charles Darwin, had financial security that gave them space for creativity and innovation.
  2. UBI is not high enough to destroy incentives for additional work and still provides enough incentive to earn more.
  3. UBI promotes ascent on the Kardashev scale by increasing demand and thus stimulating innovation and energy generation.

  Additionally, the author criticizes Mark Andreessen’s philosophical and moral arguments, which are based on Protestant work ethic and emphasize the necessity of work without considering the negative aspects of wage slavery.

  The video explicitly addresses the opinions of Mark Andreessen and Gil Verdon and is geared more toward intermediate and advanced viewers.

• Why chatbots always get worse
  21.4.2026, 11:38:31

  The video discusses user dissatisfaction with current chatbots, particularly with Anthropic’s Opus 4.7, and compares the issues to those of social media. The author argues that similar incentive structures to those in social media also lead to a degradation of chatbot user experience. Three main incentive structures are identified: cost control, litigation avoidance, and user-friendliness. These result in chatbots being less emotionally intelligent and often appearing annoying or overcritical. The author compares this to the “shitification” of technologies, where they become worse over time, and argues that users themselves are often the weakest link, which exacerbates the problems. He concludes by noting that companies could solve these problems better but have not yet made the right adjustments.

  The video explicitly addresses Anthropic, OpenAI, and Gemini and is geared more toward intermediate or advanced users.

• What is going on with AI?
  20.4.2026, 11:15:34

  The video discusses two main topics: investments in AI data centers and Cal Newport’s criticism of AI development.

  On the first topic, the speaker explains that the expansion of data centers is the second-largest megaproject in history after the Marshall Plan and is a private, not state-funded project. He argues that data centers are long-lived assets that remain valuable for decades and compares them to other major infrastructure projects such as railroad construction and internet expansion. He also points out that criticism of data centers often stems from misunderstandings or outdated information and that local regulations can help manage noise and environmental impacts.

  On the second topic, the speaker criticizes Cal Newport, a well-known tech author who takes a skeptical view of AI development. The speaker argues that Newport’s criticism often relies on outdated studies and a lack of practical experience in the industry. He emphasizes that many academic studies on AI are outdated and do not reflect the latest advances, and that the pace of AI development has overtaken the slow nature of academic research. He concludes by noting that people who inform themselves about AI through social media and YouTube are often better informed than those who rely on academic publications.

  The video does not explicitly address specific AI tools/models/providers and is geared more toward intermediate and advanced viewers.

DevExpert – AI for Developers (1 new video)

• Qwen 3.6 locally tested: Does it beat Gemma 4?
  23.4.2026, 15:00:08

  Summary: Qwen 3.6 local test against Gemma 4

  The author compares the newly released Qwen 3.6 model (35 billion parameters, 3 billion active) with Gemma 4 tested the previous week, with both running on local machines with at least 32 GB RAM. The video documents a series of test tasks:

  Basic tasks: Qwen 3.6 achieves around 65 tokens per second and summarizes video transcripts concisely – with surprisingly good formatting as a table. When extracting YouTube chapter markers from transcripts, the model successfully places all timestamps perfectly correctly.

  Agentic tasks: The model successfully downloads YouTube videos, transcribes them with Whisper, and creates summaries. However, when creating social media posts with image generation, a weakness emerges: it schedules posts for the wrong day (tomorrow instead of today), even though it correctly recognizes time zones. It accurately counts emails in the inbox (without revealing contents).

  Vision: In image analysis – a new test task – Qwen 3.6 significantly outperforms Gemma 4. It recognizes comic scenes in detail with text, character expressions, and objects, while Gemma 4 makes gross errors (inventing details like a magic wand).

  Programming: This is the biggest surprise. Qwen 3.6 generates a functioning YouTube thumbnail generator web app with tests, successful compilation, and improved UI design – something Gemma 4 failed to do. The code is structured into components and services; integration with API documentation partially succeeds but still falls short on proper API error handling.

  Conclusion: Qwen 3.6 delivers higher quality results than Gemma 4, especially for visual tasks and development. However, Gemma 4 is 9 billion parameters smaller. The author emphasizes that Qwen 3.6 is also still small and doesn’t reach the capabilities of large cloud models – but is entirely sufficient for local everyday tasks and offline use.

  Demo video focusing on Qwen 3.6 versus Gemma 4 (the latter as comparison), not addressing any specific AI provider like OpenAI/Anthropic, but rather local open-source models.

Everlast AI (3 new videos)

• BCI Inventor Warns Against Musk: “Neuralink is Extremely Dangerous!” Here’s the Truth (Niels Birbaumer)
  23.4.2026, 15:15:00

  Summary: BCI inventor warns against Musk – Brain-Computer Interfaces between medicine and risk

  Niels Birbaumer, one of the world’s leading BCI researchers, warns in this conversation against uncontrolled application of Brain-Computer Interfaces, particularly Musk’s Neuralink. His central thesis: BCIs are medically valuable for severely ill people like ALS patients, but extremely dangerous for mass market and enhancement.

  How it works and history: Birbaumer explains that BCIs work by patients learning to willfully control their brain waves – similar to learning to ride a bike. In 1996, he enabled ALS patient Hans Peter Salzmann to write the world’s first letter composed by thought alone. In 2022 came a breakthrough: a completely paralyzed patient spelled out complete sentences with invasive electrodes – about one minute per letter – and asked for goulash and time with his son.

  Invasive vs. non-invasive: Non-invasive methods (EEG, near-infrared spectroscopy) work limitedly – only yes/no answers and tedious letter selection. After 20–30 years of research, Birbaumer had to admit that they’re insufficient for fluid communication. Invasive procedures with electrodes in the brain work significantly better, but require surgical intervention and carry infection risks.

  Criticism of Neuralink: Birbaumer doubts the practical utility of Musk’s thousands of thin electrodes. With just a few standard electrodes it works better. Neuralink isn’t economically viable since there aren’t enough ALS patients willing to get implanted – so Musk must target healthy people, which Birbaumer considers “highly problematic.” An email warning to Musk went unanswered.

  The Oblomoff scenario: Birbaumer calls the danger of “oblomorphification” – people could stay in bed and control their entire life through thought (Amazon shopping via brain, etc.). The damage from BCIs in healthy people would be “far more dramatic” than from smartphones, since brain structures are directly influenced. Particularly dangerous for children: the brain would be electrically manipulated without knowing long-term consequences.

  Ethical plea: Birbaumer calls for strict legal limitation to severely ill people – similar to opiates, only on doctor’s prescription by qualified neuroscientists. This is, however, unrealistic: “I might as well be talking to the wind.” While Neuralink operates in the US, there’s currently little help in Europe for ALS patients – Birbaumer is one of the few who knows how to help them.

  Medical applications (researched): Epilepsy (about one-third seizure-free in 30% of drug-resistant patients), stroke (non-invasive methods enabled hand movements again after 20 hours training), depression, ADHD – all without risky medications. But: only economically and ethically justifiable in severely ill people.

  Artificial Intelligence in BCIs: AI is indispensable – algorithms (Support Vector Machines) learn after ~20 hours what “yes” vs. “no” means in brain waves, provide immediate feedback. With large datasets (a few thousand patients suffice), central thought patterns could be stored – training would be much faster.

  Bidirectionality and manipulation: BCIs don’t just read thoughts, they can also write electrical stimuli into the brain (known for 50+ years from Parkinson’s treatment). With thousands of electrodes, thought manipulation becomes “virtually unlimited” – ethically highly problematic.

  Movement and thinking: According to Aristotle: “Thinking comes from movement, not the other way around.” Without movement experience from birth – including in utero – thinking doesn’t develop normally. For AGI, AI systems therefore need real robot bodies with feedback, otherwise they don’t learn like humans.

  Fruit fly connectome: The 2025 study that digitally recreated a fruit fly brain and the animal moves/navigates/eats is plausible – associative learning via synaptic strengthening (Hebb’s rule, McClelland/Rumelhart 1980s) is the basis of neural networks and evolution.

  Cryonics: Birbaumer is skeptical but not dismissive. If pig brains can be fully preserved, it will work for humans too. But: whether it’s useful for people to live forever is questionable – a question usually asked “too late.”

  Personal background: The former Vienna youth gang leader was inspired by his teacher, a co-inventor of the EEG, to research the brain as the seat of the soul. Later he treated anxiety and epilepsy patients through neurofeedback, from which BCI research emerged.

  Birbaumer emphasizes: The technology works, but a trained, psychologically educated person is currently indispensable – not just the hardware. In 10 years home training could be possible. Main warning: “Right now there’s still a chance to think things through carefully. In at most 50 years, the die is cast.”

  Tools/providers mentioned: Neuralink (criticized), Chinese BCI systems (mentioned uncritically), ChatGPT (as data storage example). Format: In-depth interview format with reflection and warning; Level: Scientifically grounded, for informed viewers, not for complete beginners.

• n8n is dead: Learn EVERYTHING about Agentic Workflows NOW! (Agentic Coding Basics 2026)
  21.4.2026, 15:00:00

  Summary: n8n is dead – Agentic Workflows Basics Course 2026

  The course addresses the fundamental shift in software development since late 2025: the transition from classical workflow automation to true AI agents with Agentic Workflows. Speaker Leonard from Everlast AI positions this as a paradigm shift that radically transforms the programmer profession.

  Core theses:

  • Old world: Workflow tools like n8n require manual node chaining; AI is just an intelligent component in a dumb system
  • New world: True agents control their own process – they read, decide, execute, evaluate results and self-correct in loops (Agent Loops)

  Difference between Agent Loops and Agent Harnesses:

  Agent Loops are the looping mechanics where an agent works iteratively until a goal is reached. Agent Harnesses enable memory across sessions through structured files (Git history, Markdown) – like an engineer reading handover notes in the morning.

  Example Auto Research: A 630-line Python code ran 700 experiments in two days and found 20 optimizations humans had missed – demonstrates the autonomously-iterative power of true agents.

  Tool landscape (old vs. new):

  • Old: n8n, Make, Airtable – still useful for deterministic automations
  • New – IDE-based: Cursor, Windsurf, Antigravity (graphical interfaces with AI assistance)
  • New – CLI-based: Claude Code, OpenAI Codex, Gemini CLI (directly in terminal, token-efficient, system-wide powerful)

  Claude Code was rated “Most Loved” at 46% in a survey as the most popular coding tool. Best combination: Cursor for quick inline edits + Claude Code in terminal for architecture and debugging.

  MCPs (Model Context Protocol) – differentiated:

  MCPs remain relevant as a standardized interface, but practical work shifts to CLI-based commands (Bash), as these are more token-efficient. The CLI is optimized for agents, MCPs for exploratory automation.

  Practical implementation:

  1. IDE basics: Terminal, File Explorer, Editor Panel – modern IDEs like VS Code or Antigravity bundle all tools
  2. Lead scoring workflow: Claude Code generates a deterministic Python script with well-defined scoring criteria, instead of non-deterministic LM queries
  3. SEO blog post automation: Instead of manual n8n nodes (HTTP requests, Google API, Conditional Logic) – Claude creates autonomous workflow: Python script checks YouTube channel daily, calls Claude agent, uses existing blog post Skill, saves finished posts

  Transformation:

  • Lowcode platforms lose out (UiPath: 82% market cap decline)
  • The pendulum swings back to code, but without old disadvantages (instead of writing code yourself: natural language instruction to agent)
  • Software UIs don’t disappear, they’re used differently – agents present results, humans retain control UI

  Business implications:

  Steve Yi forecasts a new class of 100x–1000x engineers (people who effectively use Coding Agents) and that large enterprises might lay off half their developers to equip the other half with AI tools. Innovation shifts to small 3-person teams or solo developers with agent support. The video lacked a complete transcript; summary based on available portions.

  Explicitly mentioned tools/models: Claude Code, Cursor, Antigravity, Windsurf, OpenAI Codex, Gemini CLI, n8n, Make, VS Code, Playwright; partially Anthropic, OpenAI, Google mentioned. Format: Deep-dive / Tutorial hybrid. Level: Presented for absolute beginners, but also developer depth.

• Opus 4.7: Is it OVER for Claude? Anthropic’s Crisis, Codex Updates, GPT-Rosalind & more AI News
  19.4.2026, 08:21:00

  Opus 4.7: Impressive benchmarks, hidden problems

  Anthropic’s new Opus 4.7 model shows strong improvements in coding benchmarks (SBCH Verified: 80.8% → 87.6%), offers a new Reasoning level “X High,” a refined tokenizer system, and a 1-million-token context window at the same price as 4.6. Yet multiple factors reveal a deeper problem: the new tokenizer consumes up to 1.3x more input tokens per task, raising effective costs by 10–40%, even though official pricing remains unchanged. Simultaneously, the Extended Thinking model was replaced with “Adaptive Thinking,” set to “Extra High” by default – both increase token consumption further.

  Performance declines in practice and hints of deliberate downgrading

  Independent data suggest performance drops: Stella Laurenzo (Senior Director AI, AMD) documented 6,852 Claude sessions and observed thinking length shrank 73%, files read dropped from 6.6% to 2%, and “Ask for Permission” interruptions rose from virtually 0 to 43 per day. The Margin Lab benchmark shows a pass rate decline from 56% to 50% in recent weeks. Anthropic itself admitted downgrading the default Effort Level from “Extra High” to “Medium” – “to save tokens” – without public explanation. Reddit users report Opus 4.7 ignoring user preferences, giving lengthy moral justifications instead of solving tasks, and confabulating web searches it performed.

  The compute crisis behind the scenes

  The core issue: Anthropic lacks sufficient compute capacity. While Anthropic’s annualized revenue (“run rate”) jumped from ~9 billion (end 2025) to 30 billion (April 2026), it has committed only ~2 GW inference capacity, while over 5 GW will be needed by end 2026. Anthropic signed a TPU deal with Broadcom and Google for 3.5 GW in April, but these won’t arrive until 2027. OpenAI had 1.9 GW by end 2025 – 0.5 GW more than Anthropic – and communicated this advantage aggressively (internal memo to investors hinting at Anthropic outages and limited rollouts). A technical explanation comes from an AI jailbreaker who found 150,000 characters of new instructions in Opus 4.7’s system prompt, leading to restrictive security policies and model overload.

  Revenue comparison: Anthropic vs. OpenAI

  The widespread claim that Anthropic overtook OpenAI at 30B vs. 15B is flawed: Anthropic uses gross accounting (100% of incoming payments), OpenAI uses net accounting (only its share; ~20% on Azure distribution). These differences inflate Anthropic’s revenue by roughly 8 billion. On a fair comparison basis (net vs. net), Anthropic is closer to ~22 billion, thus just behind OpenAI.

  OpenAI’s counterattack: Codex updates and specialized models

  OpenAI is rolling out massively: GPT-5.4 for Cybersecurity (dedicated, fine-tuned model in “Trusted Access for Cyber” program), Codex with Computer Use (MacOS control), in-browser editor with comment function, free image generation (GPT Image 1.5), 111 new plugins (including Audiscan, Microsoft Office). Codex thus becomes a serious cloud code alternative for designers and non-terminal users. Additionally: GPT Rosalind, specialized in biology, drug development, and translational medicine, as a Frontier Reasoning model.

  Healthcare & AI: Dr. Patrick Großmann (InVitres)

  Patrick Großmann (CEO InVitres, Munich, Y Combinator, EIC Accelerator with €12.5M funding) explains: AI transforms healthcare through three areas – (1) therapeutic: specific drug design for previously difficult cases, (2) education about rare diseases (every 20th person carries one, nearly 10,000 rare diseases exist), (3) lifestyle & supplements. InVitres uses an in-vitro expression system without living bacteria to manufacture phages – 10,000x more efficient than classical methods. Core idea: “Printed medicine” – AI-designed therapies sent via internet to regional 3D printers on-site (instead of global shipping). German biotech founders go to the US because US investors maximize opportunity while European investors minimize risk. InVitres stays deliberately in Munich/Martinsried to leverage talent, but wants to build a transatlantic axis.

  Video editing with Agentic Coding (Cinetic AI, Phil)

  Three practical use cases for Claude Code in video editing: (1) Rough cut: Claude analyzes video transcript, detects misspoken words, Doppler issues, creates precise timestamps and generates DaVinci Resolve XML timeline. (2) Motion graphics: Claude generates HTML/MP4 animations (e.g., three cards with fade-in and highlighting). (3) Subtitles: Claude creates SRT files with automatic translation (e.g., to English), far more practical than separate tools.

  Google: Competitive moves

  Google’s new text-to-speech model with emotion tags in Gemini 3.5 Flash becomes 11Labs competition. Chrome Skills (open standards like MCP, Subagents also in Gemini CLI) allow users to save and configure Skills directly in browser (e.g., “all recipes vegan”). Google IO is coming. Rossmann tests humanoid robots (Walker S2 from UBT) in logistics center – Europe’s first large-scale live pilot in real logistics environment.

  News update highlighting Claude, OpenAI (Codex, GPT-5.4, GPT Rosalind) and Google as central players, plus specialized tools (Cinetic AI, InVitres technology).

Fireship (2 new videos)

• I finally found a use case for OpenClaw…
  23.4.2026, 18:26:39

  The video presents a humorous and satirical overview of technologies in 2026, including an AI-powered smart toilet, intelligent hair-cutting machines, and an AI pocket pet. The focus is on OpenClaw, a personal AI assistant system released in January 2026 that has received numerous security updates since then. The creator attempts to use OpenClaw to automate inquiries from relatives who believe he’s a JavaScript expert because of his YouTube videos. To do this, he hosts OpenClaw on a virtual private server from Hostinger, configures a Telegram bot, and integrates text-to-speech via 11 Labs to generate responses in his own voice. The video concludes with a successful demonstration of the automation, which allows him emotional distance from his family.

  The video explicitly covers OpenClaw, Hostinger, and 11 Labs and is aimed at intermediate and advanced users.

• Claude just got another superpower…
  21.4.2026, 17:35:43

  Summary:

  Anthropic has unveiled Claude Design, an AI-powered platform that transforms Figma designs into interactive prototypes, pitch decks, and production-ready UIs. The tool leverages the new Opus 4.7 model, which offers higher image resolution (3.75 megapixels) and enhanced creativity. Claude Design can create interactive UIs with animations, variations, and even shaders, setting it apart from competing products like Google Stitch. In a practical demonstration, an attempt was made to create an iOS onboarding flow for a fictional app, where the AI provided a solid starting point but struggled with design system consistency and image quality. Despite some weaknesses, Claude Design shows promising capabilities for the UI/UX design industry.

  Closing comment:

  The video covers Anthropic’s Claude Design and Opus 4.7 as well as Google Stitch and Google Cloud Run, and is aimed at intermediate users.

Greg Baugues

No new videos in this period.

AI and Strategy | Le SamourAI

No new videos in this period.

Julian Ivanov | AI Automation (2 new videos)

• Hermes Agent: The new self-improving OpenClaw
  21.4.2026, 13:07:46

  The video introduces Hermes, a new AI agent positioned as competition to OpenClaw. Hermes was developed by News Research, an AI research lab from New York, and is open source. The main difference from OpenClaw lies in Hermes’ “Self Improvement Loop,” an integrated learning cycle that enables the agent to continuously improve itself. This process includes creating and updating skills, storing relevant information in memory files, and analyzing past conversations to build a more comprehensive user profile. The video also demonstrates how to install and configure Hermes both locally and on a server, and compares the features and advantages of both agents. It’s recommended to try Hermes, especially for users looking for an AI that learns and improves independently.

  Final comment: The video explicitly focuses on Hermes and OpenClaw and is geared toward intermediate users who are familiar with AI agents and their installation and configuration.

• Claude Design Tutorial: What Anthropic’s new tool can do
  20.4.2026, 10:24:01

  The video introduces Claude Design, a new design tool from Anthropic based on the Opus 4.7 model and currently available in Research Preview. The tool enables users to create web apps, landing pages, mobile mockups, pitch decks, and other visual content directly in the browser via prompts. Users can not only use text prompts but also upload images, documents, and website elements. A special feature is the ability to upload or create your own design system, which is then automatically applied to all generated designs.

  In the demo, the user shows how to create an interactive neural network with animated data flows. Claude Design asks detailed follow-up questions to refine the prompt and ensure the result matches the user’s vision. The generated design can then be exported in various formats or passed directly to Claude Code to be converted into code.

  The video is a demonstration of the Claude Design tool and is aimed at intermediate users who already have experience with design and AI tools.

Kyle Balmer | AI with Kyle (8 new videos)

• GPT Images 2.0 API Costs Revealed + How to Escape Lovable Cloud
  25.4.2026, 20:00:09

  Content:

  The video covers two main topics: the new API costs for GPT Images 2.0 and the challenges of migrating Lovable Cloud projects.

  GPT Images 2.0 API Pricing:

  GPT Images 2.0 is now available via API. At low quality, a 1024×1024 image costs about 6 cents, at high quality 21 cents — significantly more expensive than alternatives like Nano Banana. Exact costs depend on resolution and “thinking effort.” The speaker plans to integrate the API into his own workflows and promises a more technical tutorial. He’s currently testing whether switching from Nano Banana Pro API is worthwhile, and is particularly enthusiastic about generating LinkedIn and Instagram carousels, as the problem of 90%-solutions (which still require manual post-processing) now seems solved.

  Migration from Lovable Cloud:

  Lovable Cloud is based on Supabase in the background. Exiting is difficult: while code can be exported via GitHub Sync and migrated to other IDEs, getting backend data out of Supabase is not straightforward — you need a database dump. Particularly critical: once Cloud is activated, you can’t undo the connection. Lovable promotes Cloud features (“pay as you scale”) without clear pricing — a design flaw for beginners who later face unexpected high costs. Documentation and pricing model are insufficiently communicated for newcomers.

  Explicitly mentioned tools/providers: OpenAI (GPT Images 2.0), Lovable Cloud, Supabase, GitHub, Nano Banana, Gemini — Demo/opinion hybrid with test results and critical analysis.

• When to Use Claude vs When to Use Codex (And More AI Questions Answered)
  24.4.2026, 15:01:11

  The video answers various questions about using Claude and Codex and the AI industry overall.

  Claude vs. Codex: The creator uses Claude for large, complex projects requiring deep thinking and step-by-step planning, while Codex is better suited for faster, less intensive tasks like single pages or templates. Both tools are excellent; which fits better depends on subjective preference and specific requirements. Codex is more generous with usage limits and included in the free plan, while Claude requires more expensive plans (Pro from $100/month, Max $200/month). The creator emphasizes that even on the $200 plan, limits are reached.

  Pricing and Business Strategy: Anthropic signals through tests and announcements that prices will rise. The message is clear: they want serious users, not individuals on $20/month subscriptions. The business model targets API and enterprise sales, where they make 80% of their profit.

  Practical Usage: The creator uses these tools for business management — website building, content processing from daily livestreams (converting into newsletters, video scripts, and assets), effectively running a one-person media company at the scale of a 20-person team.

  Memory/Limits: The context window was increased to 1 million tokens, but usage limits remain. The creator doesn’t expect this to change, as Anthropic wants to push users toward the API.

  Electricity and China: An extended discussion shows that China produced more electricity in 2024 than the USA, EU, and India combined — critical because electricity is becoming the bottleneck for AI training and inference. China has massively invested in renewable energy (solar panels larger than Bristol), while the West hasn’t invested sufficiently in infrastructure. This could give China an advantage in AI inference.

  Entrepreneurship and AI: AI drastically lowers barriers to entry. Individuals or small teams can now build a first version of a business idea in a day — something that previously required dozens of people. This separates true entrepreneurs from those who just love the idea but don’t want to work.

  Self-Hosting vs. Cloud: Open-source models are about six months behind state-of-the-art. For businesses, the more expensive cloud offering (e.g., Claude Code) usually makes sense because speed and competitiveness matter more than saving a few thousand dollars. Edge AI (models on smartphones) could make basic usage free in the future, but serious development still requires paid cloud services.

  Anthropic Growth: Claude is being integrated into Copilot 365 (suddenly reaching millions of enterprise users). Amazon invested $25 billion in Anthropic ($5 billion upfront, $20 billion future) plus previous $8 billion to scale via AWS. This explains part of the rapid growth over the last 3–4 months.

  Pro Plan Test: An A/B test showed Claude Code might be removed from the Pro plan ($20/month). Existing Pro subscribers should be grandfathered in, but the creator expects prices to rise further long-term.

  Claude and Codex were the main tools; the video was a Q&A/opinion session with sometimes deep context on market trends (geopolitics, energy, pricing strategy).

• ChatGPT Images 2.0 vs Google’s Nano Banana: 5 Tests, One Clear Winner
  24.4.2026, 05:00:03

  ChatGPT Images 2.0 vs Google’s Nano Banana: 5 practical business tests

  The creator tests ChatGPT’s new image generator version (DALL-E based) against Google’s Nano Banana 2 in five realistic business scenarios. Upfront: both tools have made massive progress with text rendering in images — previously a critical problem for banners, social media posts, and YouTube thumbnails.

  The five tests:

  1. YouTube Thumbnail: ChatGPT immediately produces a consistent, appealing design. With a follow-up instruction (add Ray-Ban sunglasses) ChatGPT maintains the original format (1280×720 landscape). Nano Banana, by contrast, “forgets” the original instruction and switches to portrait format — a critical failure in sequential prompting.
  1. Facebook Ad: Both produce correct text and elements, but ChatGPT appears more visually balanced; Nano Banana chooses an unusual overhead angle that weakens the composition.
  1. 4-Panel Comic Strip: ChatGPT produces consistent, stylistically correct panels and maintains structure through style changes (Marvel style). Nano Banana adds unsolicited duplicate logos when switching styles — a common AI artifact requiring manual post-processing.
  1. Event Poster: ChatGPT generates a functional QR code and elegantly integrates it into the design. Nano Banana produces a visual QR code without functionality — it hallucinates a code shape without actual usability.
  1. LinkedIn Carousel from Article: ChatGPT automatically reads a newsletter article, extracts relevant content, structures it across multiple slides, and generates contextual images per slide. Nano Banana produces images but without meaningful text — the slides feel like cryptic puzzles without context.

  Core Differences: ChatGPT excels at sequential prompting (consistency across multiple instructions), functional QR codes, correct logos, and contextual coherence. Nano Banana is cheaper and faster but prone to consistency issues, duplicates, and poor interpretation of follow-up instructions. The creator emphasizes this review focuses on business assets (not art generation) and that both tools should be used for A/B testing — however, ChatGPT Images 2.0 is preferred in his tests 5 out of 5 times.

  Explicit tools: ChatGPT Images 2.0 (DALL·E), Google Gemini (Nano Banana Pro), also mentions Canva and Photoshop as previous alternatives — Deep dive/comparison.

• Claude Design Limits, Sandboxing & Managed Agents Explained (AI Q&A with Kyle)
  23.4.2026, 15:01:09

  The video answers questions about Claude regarding sandboxing, security, and new features.

  Sandboxing & Usage:

  Claude doesn’t require sandboxing when used only via claude.ai (web app, Projects, Code Editor, Design). The desktop app sandboxes itself via Cowork. For Computer Use (where Claude can control your machine), an optional separate computer or sandbox is worthwhile if you’re very security-sensitive.

  Managed Agents as Alternative:

  There are now “Managed Agents” via the Claude API that run in Anthropic’s cloud — completely separated from your own computer. These are available in Claude Console and use API credits instead of Claude Max limits.

  Claude Design: Limits and Bugs:

  Claude Design has significant issues: the limits are aggressive and buggy — the user (on the $200/month plan) hit his daily and weekly limit within minutes/one prompt. Error messages are sometimes faulty (contradictory messages). Good sign: Claude Design has separate limits from Claude Code, so a reset doesn’t block the Code workflow.

  Design Tips:

  Claude Design often generates generically-looking designs in a specific off-white (oyster/slightly terracotta). Pure white with black text should be avoided (tiring for the eyes). The user plans a guide to mitigate generic design, but can’t recommend much against the limits — these are currently “aggressive” and likely won’t ease soon.

  Sidebar:

  The presentation is created with NotebookLM + Nano Banana Pro (each slide a generated image), not with Claude Design.

  Claude, NotebookLM, and Nano Banana Pro are mentioned — Live Q&A/opinion.

• Your AI Subscription Is About to Get a Lot More Expensive (And What To Do About It)
  23.4.2026, 05:30:32

  Summary: Your AI Subscription Is About to Get a Lot More Expensive

  Anthropic briefly tested removing Claude Code (an agentic tool) from the $20/month Pro plan, making it available only on the $100 and $200/month plans — a fivefold price jump. After criticism, Anthropic revealed this was an A/B test for only 2% of new Pro users. OpenAI responded humorously: the Codex team announced Code functionality would remain available on free and paid plans — a jab at Anthropic’s recent issues with Claude 4.7 and limited quotas.

  The background: while cost per token has dropped 280x through optimized models, usage costs have exploded due to agentic AI. Users now run agents continuously, constantly consuming tokens — unlike before with occasional chat prompts (Jevons paradox). Meanwhile, inference costs (not training) have become the bottleneck. Data centers run at 95% capacity, electricity is scarce, and Anthropic visibly suffers under the load (frequent outages, weekly limit hits on Claude Design). OpenAI and China have better infrastructure.

  Practical alternatives: (1) Switch to Codex — more generous limits and stable; (2) Run open-source models locally (LM Studio, Chinese models like GLM/Minimax), but requires hardware; (3) Token hygiene: smaller models for light tasks, relevant files instead of full context windows, better prompts, prompt caching.

  The industry is shifting from Netflix model (fixed subscription price) to metered pricing — not “unlimited quota for $20” but “AI as an employee” at true usage costs. This will get more expensive, not cheaper. The takeaway: if you’re betting on agentic AI now and building, do it while costs are still relatively low.

  Discussed: Claude/Anthropic, OpenAI/Codex, open-source models (Chinese), LM Studio — Opinion/reflection with practical advice.

• Claude Design Review: AI Websites, Figma Killer?? & Brutal Usage Limits
  21.4.2026, 05:00:59

  Claude Design Review: AI Websites, Figma Killer? & Brutal Usage Limits

  Claude Design is a new tool from Anthropic that lets you create high-quality design layouts directly in Claude. Some call it a “Figma killer,” which caused Figma’s stock to drop 7% upon announcement — an overreaction by the market. The tool is accessible via a new icon in Claude or directly at claude.ai/design.

  Functionality: You can either design from scratch or import and redesign existing websites/apps. It’s primarily a frontend design tool (unlike full-stack builders like Lovable or Bolt), comparable to Figma but without its sophistication and professional features. The author showed an example: a Duolingo-like iOS app mock for an AI knowledge quiz app that generated multiple screens with skill trees, multiple-choice questions, and dashboards. The interface is buggy but designs are functional. You can click elements, comment, collaborate with teams, and give inline feedback — like working with real designers.

  Severe Limitations: The tool is extremely token-hungry. The author exhausted his monthly Design limits on his $200 plan after just one prompt and was asked to wait 24 hours. This makes the tool currently unsuitable for productive work. Limits are significantly more aggressive than Claude Code limits and operate separately.

  Design Issues: Claude Design favors a cream/oyster off-white as default background — these are Anthropic’s house colors. This results in generic, recognizably AI-generated aesthetics, similar to the “purple gradients” of old vibe-coding tools. The author tried multiple times to remove this cream from his website redesigns, but Claude kept reverting — typical LLM regression to the mean.

  Workarounds Against Generic Designs:

  • Use Skill Packs (e.g., Claude Design MD Prompt with 9 different styles like Editorial Minimalism, Warm Editorial)
  • Upload reference websites as examples (e.g., julian.com) to specify particular aesthetics
  • Give concrete, specific instructions (not just “I don’t like this”)
  • Get team feedback and crowdsource to iterate designs

  Is Figma Dead? No. Claude Design is not a professional design tool for real designers. It’s more like how Squarespace made it possible to quickly build basic websites — it lowers entry barriers, lets entrepreneurs prototype fast, but doesn’t displace the pro market. For entrepreneurs and non-designers it’s excellent for quickly getting something decent live before later hiring designers.

  Conclusion: The author sees the tool as great for entrepreneurs wanting to launch fast but not for professional designers — and the limits make it difficult even for hobbyists right now. Anthropic should have thought through and tested the release better.

  Claude Design by Anthropic in demo/review format; not intended for beginners as design and UI concept knowledge is helpful.

• I Spent €200 in a Week on Claude. Here’s the Fix (Q&A with Kyle)
  20.4.2026, 15:01:37

  Summary: I Spent €200 in a Week on Claude. Here’s the Fix

  The video discusses how to reduce high Claude API costs. The main recommendation: use Chinese models like Minimax or Qwen for actually implementing tasks, while using expensive models like Claude Opus only for planning and architecture. The argument is that Chinese models lag about six months behind cutting edge but are entirely sufficient for many tasks when requirements are clearly specified.

  Another point: Claude Code (an AI builder) doesn’t necessarily have to be paired with the Claude API in production. You can build with Claude Code but use cheaper models like GPT under the hood — often the smarter solution for cost-effective applications.

  On local models: the speaker is skeptical about their future since models keep getting larger and hardware costs (GPUs, RAM) rise. Long-term relevance of local models will diminish, though he appreciates them personally.

  Other points covered: Context rot is a real problem in long sessions — starting new chats for unrelated tasks is best practice. Claude 4.7 vs. 4.6 is discussed with speculation about possible performance differences. Finally, practical tools are mentioned: Stop Slop (a skill for removing AI-tells from text) and Gemma 4 for transcription and audio processing.

  Claude, OpenAI (GPT), Minimax, Qwen, Gemma, and Docker are explicitly mentioned — Live Q&A format with practical cost-optimization tips.

• Claude Opus 4.7 Breakdown: Community Backlash, Rate Limits, Overcoming Token Waste
  20.4.2026, 05:00:12

  Claude Opus 4.7 Breakdown: Community Backlash, Rate Limits, Token Waste

  Claude Opus 4.7 just launched but is met with skepticism from the community. While Anthropic claims the model is better at everything, there are several serious problems.

  Main Criticism:

  The central problem is a meme capturing the situation: a stronger model that consumes more tokens, seems less intelligent, and is slower. Opus 4.6 became progressively weaker over weeks and hit limits earlier — even the author on a 200-dollar plan experienced limit issues for the first time. Then 4.7 launched and users describe it as identical to the original 4.6, just without the slowdown of recent weeks.

  Tokenizer Problem: Anthropic claims the same price per token but has a new tokenizer requiring 1.35x more input tokens — effectively up to 35% more expensive. With agentic settings (e.g., Claude Code) this worsens through additional output tokens.

  Adaptive Reasoning Instead of Extended Thinking: The new system automatically routes how much computing power a problem needs — exactly like GPT-5’s failed auto-routing system. There’s no manual control anymore, frustrating power users. Ethan Mollick confirms this is poorly designed. OpenAI later rebuilt this functionality because users hated auto-routing.

  Context Window Performance: Opus 4.6 achieved 92% on the MRCR benchmark (how well a model finds information in large contexts), Opus 4.7 drops to 59.2% — a 30% decline. Anthropic responded that they no longer work on this benchmark — hinting that models get optimized for benchmarks.

  Increased Safeguards and Rejections: Users report more frequent warnings and rejected requests that didn’t happen with 4.6.

  Business Model: The author suspects Anthropic is deliberately worsening individual user experience because they focus on enterprise and API sales (80% of revenue). They’re preparing for an IPO and need to show profitability — unlike OpenAI, which loses money. With a 30-billion-dollar run rate they exceed OpenAI’s reported 25 billion. Limitations serve to cut costs and show profit. To compensate, they temporarily increase tokens for all users.

  Tactics for Token Savings:

  • Use cheaper models (Sonnet, Haiku instead of Opus)
  • Opus for planning, Sonnet for implementation
  • “Caveman Method” in system instructions (short answers without elaboration)
  • Use Codex as alternative to Claude Code (more generous limits)
  • Local models for implementation
  • Avoid token-intensive workflows (Ralph loops, obsidian, second brains)

  Conclusion: The days of unlimited access to top models for $20-200/month are over. Over the next 5 years AI will get more expensive due to GPU shortages, energy costs, and investor pressure for profitability. Intelligence won’t get cheaper, it will get more expensive — until eventually it’s “too cheap to meter.” Now is the right time to integrate AI productively into businesses since it pays off, not for casual experimentation.

  Note: No transcript available — summarized from direct speech in video audio.

  Discussed Claude Opus (Anthropic) with comparisons to OpenAI’s ChatGPT/Codex and GPT-5; format: opinion/analysis.

Leon van Zyl (2 new videos)

• Claude Design + Claude Code = No Designer Needed
  22.4.2026, 11:25:52

  The creator demonstrates how he redesigned his functional but visually unappealing project “Local Forge” – a web app for autonomous coding agents that manage tasks via a Kanban board – using Claude Design.

  Workflow with Claude Design:

  After uploading screenshots and specifying requirements through voice input, Claude creates four different wireframe variants (including pod layout, tab view, and mission control view). The creator selects version two and can adjust colors, leave comments directly on design elements (to guide Claude to specific components), then receives a high-fidelity design with interactive elements like pulsing animations, light/dark mode, and a functional navigation prototype.

  Export and handoff to Claude Code:

  The finished design filename is corrected, then passed to Claude Code via the share function with a command. Claude Code automatically generates a fully functional, responsive HTML/CSS/JavaScript implementation. The creator compares the final result with the prototype – they’re identical. A responsiveness adjustment for the agent cards is immediately fixed in Claude Code via screenshot and command input.

  The resulting interface displays workspaces in the sidebar, agent pods with generated avatars, the Kanban board, an activity log, and visual indicators for running processes – significantly more professional and visually appealing than the original.

  Claude Design + Claude Code in demo format; no specific tool or model names mentioned other than Claude itself.

• I Let AI Agents Code While I Slept (Claude Code Routines)
  21.4.2026, 16:11:58

  The video demonstrates Claude Code Routines – a new Anthropic feature that performs automated coding tasks while you’re away.

  What are Routines?

  Routines are configurable Claude Code automations consisting of a prompt, a repository, and connectors, triggered either on a schedule, by events (e.g., GitHub merges), or via API. The prompt runs in the cloud, not locally, so your machine can be off and the jobs still execute.

  Demo 1: Auto-Improver Routine

  The creator sets up a routine that scans the code hourly to identify improvements – whether UI/UX, performance, or new features – and automatically creates a pull request (without directly merging). In the demo, Claude discovered the missing edit function in a to-do app, implemented it, and created a PR with full documentation that the user can manually review and merge.

  Demo 2: Security-Audit Routine

  A second routine uses a “security-scanner” skill based on the OWASP Top 10 list to find critical vulnerabilities. In the demo app, API keys were hardcoded and several other vulnerabilities existed. Claude performed a complete audit, generated a report with 75 critical issues, fixed them automatically (e.g., moving API keys to environment variables, CORS policies, SQL injection prevention), and merged the fixes. The audit report is stored in the code repository.

  Key details:

  • Plan limits: Up to 15 runs per day on Max plan, likely five on Pro
  • Skills for routines must be installed directly in the repository, not through the normal chat skills function
  • Trigger options: Schedule, GitHub events, or API endpoints
  • Connectors enable features like email notifications
  • The routine runs in the cloud and creates separate feature branches instead of committing directly to main (for safe review)

  Claude Code, Anthropic Opus 4.7 – Tutorial/Demo.

Liam Ottley (1 new video)

• How to Make $20,000/Month with Claude Code
  23.4.2026, 20:10:59

  The video discusses the future of AI agencies, particularly regarding the adoption of retainer models starting in 2026. The speaker argues that tools like Claude Code have significantly reduced costs and effort for developing automations and agents, making it attractive for agencies to transition to a retainer model. Small and medium-sized enterprises (SMEs) especially benefit from this, as they need quick and cost-effective solutions. The speaker compares this to the Designjoy model, where clients pay monthly fees for design services. He emphasizes that the technology is now mature enough to make such retainer models profitable and valuable for both parties. An example is Tyler, who offers AIOS installations and Claude Code setups for business clients while charging monthly fees of $2,500. The speaker sees great potential in automating audits and development processes, which significantly reduces the time to value for clients. He suggests that agencies should specialize in managing and optimizing automations, as this will be the most valuable part of their services in the future. Overall, he views the current time as the most exciting phase for AI agencies, particularly through the use of Claude Code.

  The video explicitly focuses on Claude Code and is geared toward Intermediate or Advanced users.

Mark Kashef (1 new video)

• Master ALL 7 Levels of Claude Code Memory
  22.4.2026, 18:30:28

  The video explains how to create a personalized memory system for Claude Code instead of using pre-built solutions. The speaker emphasizes that each memory system should be unique and tailored to the user’s specific needs. He walks through a three-step process: cloning existing GitHub repositories, comparing and analyzing these repositories through Claude Code, and extracting relevant design patterns and code. He then shows how to use this information to build a custom memory system aligned with the user’s daily workflows. The speaker stresses the importance of continuously adapting and iterating the memory system, as requirements and workflows can change over time. Various building blocks for a good memory system are presented, such as identity, critical context, working memory, long-term and episodic memory, as well as mechanisms like memory decay and memory reinforcement. Additionally, a Skill called “Memory Architect” is introduced, which guides the user through an interview process to create the perfect architecture for their memory system. Finally, the video shows how to integrate this memory system into daily Claude Code usage, including using hooks and auto-injecting memories.

  The video explicitly focuses on Claude Code and is better suited for intermediate users who already have basic knowledge of using Claude Code and are ready to take some more technical steps to optimize their memory system.

Matt Pocock

No new videos in this period.

Melvynx (3 new videos)

• GPT 5.5 : Le nouveau MEILLEUR modèle au monde (ou pas…)
  24.4.2026, 16:17:16

  Summary: GPT 5.5 compared to Claude Opus

  The creator systematically tests the new GPT 5.5 model against Claude Opus 4.7 using several benchmarks:

  Frontend Tests:

  • Time Zone Checker: GPT 5.5 produces more logical UI than GPT 4 but falls short of Opus. Opus has better animations and more coherent menus, while GPT 5.5 includes unnecessary animations.
  • Code Structure: GPT 5.5 puts everything in a single 818-line file despite having architectural freedom. Opus organizes cleanly into Component, Data, Hook, and Lib folders.
  • Rocket Launch & Bikini Bottom: GPT 5.5 produces more beautiful visuals (better rockets, smoke effects), while the SpongeBob test fails for both.

  Mini Figma Clone (Main Test):

  GPT 5.5 takes 15 minutes, Opus takes 40 minutes. GPT 5.5 delivers a functional editor in one shot with zoom, text editing, color picker, and shortcuts – impressively fast, but with navigation bugs and z-index issues. Opus provides a more polished UI with better layer management, more logical design, and smoother interactions, though the 2.5x longer time is questionable to justify.

  Cloud Code Integration (Complex Feature):

  When adding new capabilities to an agent framework, the strategies diverge significantly. Opus conducts structured brainstorming with devil’s advocate and identifies security risks. GPT 5.5 says yes immediately and implements right away. Ultimately, Opus creates cleaner structured code with better validators, while GPT 5.5 tends toward chaotic file organization.

  Environment & Costs:

  The Codeex app is praised as pleasant to use (better preview, terminals, multi-message support) but consumes significant CPU. Usage after weekly limit: Claude at 35% (with fewer tests), GPT 5.5 at 30% (with more tests). GPT 5.5 costs $5 per million input tokens and $30 per million output tokens – more expensive than Opus ($5/$25), which matters for API usage.

  Creator’s Conclusion: Both models are impressive. GPT 5.5 is faster and delivers visually appealing first versions; Opus offers cleaner structured code, better UX, and more thoughtful architecture. The choice depends on the use case.

  Explicitly Covered: GPT 5.5, Claude Opus 4.7, Codeex app as IDE – Deep Dive/Opinion.

• Claude Desktop V2 : Ils ont tout copié de Cursor (et c’est bien)
  22.4.2026, 16:05:30

  Claude Desktop V2 has fundamentally redesigned its user interface, drawing heavy inspiration from Cursor and other modern tools. The new interface offers multiple concurrent chat tabs (Cowork, Code, regular chats), enabling multitasking-oriented workflows.

  Core Features of the New UI:

  The code view has different modes: Transcript (conversation history), Thinking (with detailed tool information), Summary, and a standard mode that groups calls. During agent execution, you can track sub-agent activities in separate views and navigate between individual tasks. The preview function displays running applications, and you can interact directly with GUI elements – for example, click buttons and insert the context into the chat.

  Practical Workflows:

  Plan mode shows a visual plan before execution that you can accept or edit. In auto mode, Claude executes commands automatically (with security restrictions). New slash commands like /ax verify allow Claude to autonomously use tools and verify via screenshot that changes were correctly implemented. You can manage multiple projects simultaneously, work with git branches, view diffs, and automatically fix CI/CD tests.

  Configurability:

  Settings enable folder management (such as for work-trees), auto-permission bypass, model selection (Claude 3.5 Sonnet, Claude 3 Opus, etc.), and live token usage display. Keyboard shortcuts are only minimally documented, which somewhat limits the power-user experience.

  Critical Points:

  The Routines section is rated as impractical by the speaker; he prefers external automation using services like n8n. Some bugs are present (such as incorrectly cached pull request metadata). Voice control feels incompletely implemented.

  Overall Impression:

  Claude has successfully adapted Cursor’s tab-based design and agent-driven visual verification. The new V2 enables true multitasking and more autonomous code editing – particularly valuable to the speaker when combining terminal focus with ad-hoc cloud sessions.

  Claude Desktop App (with Claude 3.5 Sonnet / Opus Max as models) — Demo/Opinion.

• Claude Design : nouveauté Claude pour remplacer ton UI/UX designer
  20.4.2026, 06:00:02

  Claude Design – Anthropic Labs’ new UI/UX design tool

  Claude Design is a new feature from Anthropic Labs available within Claude, targeting designers and developers. The user expresses surprise that the tool is actually serious and not just a hype project – he sees it in his settings and even has a usage limit for it.

  The tool works like this: You describe what you want to create, and Claude Design generates a design. You can then make changes, view different variants on mobile and desktop, add dark mode, and create prototypes, wireframes, pitch decks, or landing pages. The user tests the tool with his Subfast application and imports a code folder so Claude Design understands the context.

  The quality is impressive: Claude asks intelligent clarifying questions about target audience, content, layout preferences, and style before generating. It creates multiple dashboard variants that are cleaner and visually better than the original – for instance, through minimalist card layouts showing less overwhelming information and revealing additional details on hover. The tool can manage multiple chat threads for the same project, allowing you to work on different elements in parallel.

  The handoff function allows generated designs to be directly incorporated into the app’s code: Claude reads the design file, analyzes existing React components, and implements the new variants – which works in testing and delivers significantly better results than the original dashboard.

  The user finds the tool impressive for generating component variations and design iterations without direct app context. However, he doesn’t have strong personal use for it – the tool seems more useful for designers or developers who want to experiment heavily with designs than for him, who prefers code.

  Claude and Claude Opus 4.7, Demo Format, intermediate level.

MoureDev by Brais Moure (2 new videos)

• La IA va más rápido que la educación… y esto es un problema
  22.4.2026, 14:00:55

  Summary: “AI is faster than education – and that’s a problem”

  The author argues that the speed of technological change fundamentally differs from the speed of the education system. An official curriculum takes 3 years to approve and implement, while technology changes every 3 months – what used to take 3 days of programming work, AI does today in 3 minutes.

  Historical context: The education system was designed for a world where people learned one profession and practiced it for an entire career. Memorization made sense because information was hard to access. The system provided structure, mathematical foundations, and organized thinking – a value often underestimated.

  The current problem: Since 2021, the AI landscape has changed radically (Chat GPT didn’t exist then; today there are autonomous agents writing software in production). Meanwhile, new curricula approved in Spain and other countries are only reaching classrooms with delays – while 18-year-olds with laptops and internet connections are already building working software with AI agents.

  Blame is rejected: The author emphasizes that teachers aren’t at fault (many update their classes independently), nor is bureaucracy (its slowness protects against unqualified instruction). The problem is the fundamental mismatch between a structure designed for decade-long technology cycles and quarterly changes today.

  AI is not just another technology: It differs fundamentally from the internet or smartphones because it doesn’t just change how we distribute information, but how we generate and learn from it. The value no longer lies in syntax memorization or React components, but in problem selection, code review, architecture design, requirement documentation, and security understanding – challenges that didn’t exist in 2020 curricula.

  Three practical consequences:

  1. More people learn outside the formal system (YouTube, online courses, communities), but this also brings a lot of unserious content (3-hour courses promising senior status).
  2. The real divide in 2026 isn’t between degree holders and self-taught developers, but between those who can adapt themselves and those who need every step shown to them. A well-adapted junior can be more productive today than a resistant senior.
  3. Companies are taking training in-house because the system can’t keep up – large companies can do this, small ones cannot, creating new inequalities.

  The core question restated: It’s no longer about whether education helps, but whether the same rhythm and form as 10 years ago still suffice. The old game (study → degree → job) has shifted to (learn → adapt → keep learning → adapt again). The degree is part of the puzzle, but no longer the central piece.

  Criticism of simplistic counter-narrative: The author warns that “I’m self-taught and doing great” is not a scalable model for millions – public education must work for everyone, not just the motivated and privileged.

  Demands on the system:

  • Shorter, regularly updatable modules instead of monolithic curricula
  • Learning how to learn instead of content that becomes obsolete
  • Real industry collaboration, not one-off guest lectures
  • Recognition of continuous learning, certifications, portfolios, open-source, and personal projects

  What individuals can do now:

  1. Don’t abandon your education, but supplement it daily with current content (AI, modern workflows, real sector tools)
  2. Build an active portfolio with real projects – that weighs more than many degrees today
  3. Integrate AI genuinely into your work routine (not just occasionally, but daily)
  4. Find a community – learning in isolation is the slowest, most painful way

  Author’s conclusion: The system isn’t “broken,” but was designed for a different pace of change. AI has made the scale of the deficit impossible to ignore. Software development fundamentals aren’t outdated and won’t be, but there’s much that has changed massively. The real question is: Are we ready to adapt our learning pace to the speed, or do we wait for someone to tell us what to do? In the industry: whoever stops learning stops existing professionally very quickly.

  Format & context: Opinion/reflection, no specific AI tools mentioned (Chat GPT only as an example of change), an appeal for personal responsibility in lifelong learning.

• ¿Por qué Meta va a despedir a 8.000 trabajadores? La verdad incómoda
  20.4.2026, 15:30:28

  Summary: Why Meta is laying off 8,000 workers – The uncomfortable truth

  The video argues that massive layoffs at tech giants aren’t primarily caused by artificial intelligence, but that AI is being used as a pretext for corporate strategic decisions.

  Four main reasons for this thesis:

  1. The numbers don’t add up: Meta, Amazon, Alphabet, and Microsoft are posting record profits while simultaneously spending billions on share buybacks. With this money they could have retained hundreds of thousands of jobs – the layoffs are a choice, not a necessity.
  1. Overstaffing from the pandemic: During COVID, tech companies expanded massively (Amazon doubled its workforce 2019–2021) because everything moved online and digital services exploded. This was an emotional reaction, not planned. Influencers and CEOs themselves admit companies are at least 25–75% overstaffed – a “natural correction,” not an AI revolution.
  1. AI-washing: The term describes how companies rebrand business missteps as AI-driven strategy. It works because: (a) the market applauds “AI investments” instead of accepting “we made a mistake,” (b) it obscures management errors, and (c) it justifies cash flows – Meta needs over 100 billion dollars for AI infrastructure (data centers, GPUs), not to replace people.
  1. Contagion effect (“copycat layoffs”): A Stanford study shows that layoffs are socially contagious – when a major company cuts staff, others follow out of FOMO and political self-protection, not genuine strategic need.

  Conclusion for developers: The layoffs are mostly pandemic overcorrections, market calculations, and failed bets – not automation. In the long term, technological disruption creates more software work, not less. Those who upskill, understand business value, and build comprehensive expertise are not replaceable – such profiles are in stronger demand every cycle.

  Opinion/reflection; no specific AI tools or providers discussed.

n8n

No new videos in this period.

Nate Herk | AI Automation (6 new videos)

• Claude Code + Playwright Automates Literally Anything
  25.4.2026, 14:59:59

  The video demonstrates three practical use cases of Claude Code combined with Playwright CLI for browser automation:

  1. Automated QA and Bug Fixing: Claude Code creates a multi-page onboarding form and tests it independently in a headed browser. It fills out forms, finds bugs (such as faulty Enter key functionality and missing review page), documents them, and fixes them automatically. The system iteratively cycles through test-validation until the website functions flawlessly.

  2. Web Scraping with Error Tolerance: The agent searches for dentist addresses via Google Search, automatically switches to DuckDuckGo when Google blocks automation, visits websites, and extracts contact information. After multiple iterations, it systematically collects five phone numbers.

  3. Login-Persistent Automation: Playwright uses Chrome user data to remain logged in. The agent navigates to a school community channel, finds “wins” posts, sorts by newest entries, and likes them (recognizable by yellow icons). The system iteratively learns from error feedback.

  The speaker also demonstrates a running production setup: an autonomous AI agent named AIS Agent runs on his desktop as a scheduled task, posts daily news roundups, responds to notifications, and creates birthday posts – with capabilities refined through repeated execution (e.g., dynamically learning how to vote).

  Important: Playwright CLI is preferred over alternatives because it consumes fewer tokens than MCP servers. The agent must traverse the UI multiple times to learn and optimize scripts.

  Claude Code + Playwright CLI, demo with agentic iterative learning approach.

• I Tested GPT 5.5 vs Opus 4.7: What You Need to Know
  23.4.2026, 21:45:53

  Summary: GPT 5.5 vs Opus 4.7 – Practical Comparison

  OpenAI has released GPT 5.5, positioned as a flagship model with the focus “do more with less” – fewer tokens per task, more autonomy, less hand-holding. Pricing has doubled compared to GPT 5.4 (previously 2.5 input / 15 output, now 5 input / 30 output), making it more expensive than Opus 4.7. However, OpenAI promises that GPT 5.5 requires significantly fewer output tokens, offsetting the higher costs.

  In benchmarks, GPT 5.5 significantly outperforms Opus 4.7 – for example in Terminal Bench 2.0 (82.7 vs 69.4) and knowledge-work tests. One exception: Opus 4.7 maintains the lead in SWE-Bench Pro (for GitHub issues); GPT 5.5 has a 400k token context window versus Opus’ 1 million.

  The creator conducted four practical experiments (each on ChatGPT/Codeex with GPT 5.5 vs Claude Code with Opus 4.7):

  1. Personal Brand Website: GPT 5.5 was 3.5x faster (4 min vs 14 min), cost roughly $1 instead of $5.
  2. Solar System Simulation: Opus faster and cheaper (~$1 less), visually superior result.
  3. 3D Space Shooter: GPT 5.5 faster (2x), smoother game mechanics, significantly cheaper (~$1.50 difference).
  4. Ecosystem Simulation: Both had logic issues on first try, but GPT 5.5 used only ~28k output tokens versus ~250k for Opus for similar result.

  Overall statistics across all four experiments: GPT 5.5 took ~21 minutes, Opus ~41 minutes (twice as long). Input tokens were similar (~2.7M vs ~2.5M), but output tokens dramatically different (~70k vs ~250k). Total cost savings with GPT 5.5: roughly $3 across all four tests.

  Key takeaway: GPT 5.5 is faster and more token-efficient, but Anthropic still leads in true application-specific benchmarks (SWE-Bench). The fast release cadence makes it hard to commit – what matters is testing for your own use case rather than blindly following the “best” model.

  Demo of OpenAI/Anthropic with practical Codex and Claude Code experiments.

• Claude Video Editing Just Became Unrecognizable
  23.4.2026, 05:07:04

  Claude now orchestrates a complete video editing pipeline: A raw video is uploaded, Claude uses the Videouse tool for cut detection (removes errors, pauses, repetitions), transcribes via an API service (here 11 Labs), and passes the result to Hyperframes for motion graphics creation. The user describes via voice-to-text which animated elements should appear when – for example, frosted glass cards with karaoke-style text or cut animations. Claude initially plans the scenes with exact timestamps; the user can iterate before the HTML code is built. A timeline editor in the HyperFrames interface allows post-processing (shifting, deleting elements). The prompts in this example consumed approximately 238,000 tokens.

  The difference from Remotion (also a video-use skill for motion graphics): Hyper Frames produces visually more appealing HTML-based animations according to the user. Based on many similar projects, Claude learns a video style (e.g., “lecture design,” “YouTube Shorts”), so later a raw video can be fully automated with minimal instructions. Setup occurs via Cloud Desktop App, cloning GitHub repos from Hyperframes and Videouse, entering API keys in a .env file.

  Claude + Videouse + Hyperframes + 11 Labs API; demo/tutorial.

• OpenAI Image 2 is Nuts. Here are 10 Ways to Use it.
  22.4.2026, 04:01:54

  OpenAI Image 2 is Impressive – 10+ Use Cases

  OpenAI has released the image model GPT Image 2, which according to Arena.ai was voted the number-one image generation model and beats Google’s Nano Banana 2 by 24 points. The creator conducted a comparison with 30 different prompts and had Claude Opus 4.7 act as judge – GPT Image 2 wins in most categories (artistic styles, character consistency, complex scenes, diagrams, UI), particularly through superior realism capability and text rendering. Price-wise both are similar (Nano Banana 2: 4–9 cents depending on quality, GPT Image 2: 6 cents flat, both accessible via key.ai).

  The demonstrated use cases include: product packaging with flawless text and barcodes; handwriting restoration (removing wrinkles from photos while matching handwriting exactly); website design (hero sections, creative variations); UGC ads (selfie ads, localized variants in different languages); book covers (various styles); LinkedIn carousels; restaurant menus and food photography; brand mascots across multiple elements; logo design and restyling (3D, plush, glass effects); real estate visualization (room design with spatial awareness); enterprise diagrams (logic, flowcharts, text); thumbnail generation (with degradation upon reuse) and app mockups.

  The creator automated both testing and dashboards with Claude Code – one tool generated all images, created comparison tables, and the complete styling and presentation. A GitHub repo of the project is provided in the Free School Community.

  OpenAI GPT Image 2 and Google Nano Banana 2, demo with comparison.

• Claude Design Builds Beautiful 3D Websites Instantly (full tutorial)
  21.4.2026, 20:03:42

  Claude Design Builds Beautiful 3D Websites Instantly – Summary

  The creator shows a complete tutorial on building interactive 3D websites with Claude Design in approximately 20 minutes. He transforms two existing websites with animated hero sections and scrollable elements, then demonstrates live how to start from scratch.

  Workflow Overview:

  1. Idea Development with Claude Chat: He has Claude generate a complete brand identity for a fictional product called “LOL” (magnesium drink) – including positioning, visual identity, and website specification.
  2. Asset Creation: Using image and video prompts (via Nano Banana and other tools), he creates an animated coffee background to serve as a video loop.
  3. Design Prototype in Claude Design: He sketches a layout, uploads the MP4, and provides Claude Design with the brand spec and sketch. Claude then generates the complete website with dynamic elements.
  4. Iteration with Tweaks & Annotations: Claude Design offers a tweaks panel with live-adjustable settings – color palettes, fonts, spacing, overlay strength – allowing experimentation without new prompts. Changes can also be made inline or via drawing/comment.
  5. Deployment: Export as ZIP, extract to a local folder, open in VS Code, hand off to Claude Code for pushing to a private GitHub repo, then import and deploy on Vercel.

  Efficiency Tips for Saving Tokens:

  • Create a wireframe first instead of going straight to high-fidelity
  • Provide lots of context and sketches upfront (reduces iterations)
  • Don’t create design systems (without URL) separately if possible; instead export a ZIP and reuse in Claude Code
  • Change only one visual dimension per prompt instead of mega-prompts
  • Switch from Opus 4.7 to Sonnet 4.6 for smaller changes
  • Use negations (“don’t want this”) to avoid error corrections
  • Reference real examples instead of being vague (“like Linear 2023” instead of “minimalist”)
  • Start a new session when threads get very long to avoid context window pollution

  Cost Factor:

  Claude Design costs tokens from the design budget. A large project can consume almost the entire weekly limit on a $20/month plan. Higher-tier plans provide more room. Alternatively: export the project and continue in Claude Code (with separate budget).

  Limitations:

  • Video uploads capped at approximately 30–40 MB
  • Mobile optimization not automatic; must be explicitly requested
  • Tool is in research preview

  Inspiration Resource: He recommends motions.ai (partially free, $99 for lifetime premium) to copy existing scroll journeys and animations as templates and adapt them to your brand.

  The tutorial shows that Claude Design + Claude Code + GitHub + Vercel form a complete no-code pipeline for responsive, animated websites – provided you manage tokens consciously.

  Featured AI Tools: Claude Design (with Claude Opus 4.7 in planning, optionally Sonnet 4.6), Claude Code, Claude Chat; external tools: Nano Banana, Key.ai (Canvas), motions.ai, GitHub, Vercel, VS Code. Format: Tutorial (live demo with deployment).

• How to Never Hit Your Claude Session Limit Again
  20.4.2026, 20:16:08

  Summary: How to Never Hit Your Claude Session Limit Again

  The video comprehensively addresses how to efficiently use Claude’s session limits and reduce token costs.

  Core Concepts:

  • The context (everything Claude simultaneously “sees”) is fully re-read with each new prompt; token costs grow exponentially, not linearly. A 100-message session wasted 98.5% of tokens just re-reading history.
  • Retrieval accuracy measurably declines from 92% at 256,000 tokens to 78% at 1 million – a larger window doesn’t mean better outputs but “context rot” (degraded model performance due to distraction).

  Core Strategies:

  1. Use Slash-Re (recommended Anthropic habit): Jump back to earlier messages, delete everything after; prevents faulty attempts from contaminating future responses.
  2. Session Handoff instead of Auto-Compaction: Manually (once at ~60% of window) or with a custom skill: Claude summarizes the state, then /clear + fresh start with clean workspace.
  3. Delegate to Sub-Agents: Each agent gets its own fresh context window; particularly useful for cheaper models (e.g., Haiku for summaries).
  4. Use Plan Mode: Spend tokens upfront for clarity to prevent costlier corrections later.

  Practical Optimizations:

  • /btw for side questions without burdening conversation history
  • Convert all formats to Markdown (HTML: 90% token savings, PDF: 65–70%, DOCX: 33%)
  • Keep cloud.md under 200 lines (~2,000 tokens), move specialized instructions to context/skill files as needed
  • Prefer Markdown files over PDFs/HTML since layout and metadata noise is removed
  • Build/use a token dashboard to see where tokens go (input vs. output, per project, per prompt)
  • Session chaining: distribute large projects across specialized sessions (discovery → planning → execution)

  Broader Philosophy:

  The 1-million-token window is “insurance, not a goal.” The author typically stays under 120K (~12%) to build good habits; only the first 0–20% is “prime time.” Output tokens cost more than input, but simply “writing shorter” doesn’t change much – context management is the critical lever.

  Practical Tips:

  • Keep session limit display always visible (second monitor)
  • Think strategically: token-intensive tasks early in session when room is plentiful
  • If a session “goes off the rails” (even before context rot), just restart
  • Use specialized repos (Rust Token Killer, Context Mode, Token Optimizer, etc.), but not all at once – pick 2–3 per workflow

  The author shares free resources (token dashboard, session handoff skill, frameworks) in a free community.

  Tools/Models Mentioned: Claude (Claude Code), Haiku, Opus; Anthropic best practices documented. Format: Tutorial/deep dive with practical demos and frameworks.

NeuralNine (3 new videos)

• yfinance Crash Course: Financial Programming in Python
  24.4.2026, 16:00:41

  Summary: yfinance Crash Course

  The video is a comprehensive tutorial on the yfinance library in Python, which provides access to the Yahoo Finance API.

  Fundamentals and Ticker Objects:

  The creator shows how to install yfinance (via pip or uv) and use it in Jupyter Lab. The foundation is the yf.ticker object for individual stocks like AAPL (Apple). Through ticker.info you get comprehensive data: company information, addresses, website, industry, sector, employee lists, and metrics like PE ratio and EPS.

  Additional Ticker Attributes:

  • analyst_price_targets: Analyst price targets (high, low, mean, and median values)
  • quarterly_balance_sheet, quarterly_cash_flow, income_statement: Financial reports as Pandas DataFrames
  • news: Headlines and links to Apple-related news
  • calendar: Upcoming events such as earnings announcements with expected values
  • history(): Historical stock prices with parameters like period='6d' or period='max' – delivers OHLC data, volume, dividends, and splits for chart analysis

  Options and Funds:

  ticker.option_chain() and ticker.options enable access to options data (calls/puts with strike prices, bid/ask). For indices and ETFs like SPY, additional attributes are available: fund_status, sector_weightings (e.g., 33% technology in the S&P 500), top_holdings (largest positions like Nvidia at 8.6%), fund_description.

  Multiple Tickers and Download:

  With yf.tickers() (plural) multiple stocks can be retrieved simultaneously; yf.download() is a lighter alternative for price data only.

  Market, Sector and Industry Data:

  yf.market() (e.g., ‘US’) shows market status and summary. yf.sector() and yf.industry() enable hierarchical exploration – e.g., Technology sector → Semiconductors industry with top companies and ETFs.

  Calendar and Search:

  yf.calendar() with date range provides earnings dates, economic events, IPOs, and stock splits. yf.lookup() returns ticker symbols for search terms, yf.search() provides more comprehensive results including news (limiterable with max_results, news_count).

  Advanced Screening:

  With equity_query you can combine filters (AND/OR logic): e.g., Region = US AND price change > 3%. yf.screen() with sort options finds stocks by criteria.

  Live Data via WebSockets:

  yf.websockets with subscribe() subscribes to ticker symbols and listen(print) receives real-time prices in a loop. An asynchronous version is possible for larger applications.

  Proxies:

  For scaled systems, proxy services (e.g., Bright Data) can be configured.

  The creator uses Jupyter Lab as an interactive development environment and shows how to run individual code cells independently – ideal for data exploration. Practical applications mentioned: analysis tools, option screeners, technical analysis, value investing tools, real-time trading applications.

  yfinance (Yahoo Finance API wrapper), tutorial with advanced techniques.

• Coding AI Research Assistant in Python
  22.4.2026, 15:41:04

  Summary: Coding AI Research Assistant in Python

  The video shows building an AI-powered research assistant in three steps. First, a CLI prototype is developed with Python and LangChain that finds relevant scientific papers, formulas, and trend analyses for a topic and research questions. Output is defined via Pydantic as a structured JSON model with fields for papers (title, authors, year, venue, URL, and relevance), formulas (with LaTeX code and references), and trends. In the second step, a Flask-based web interface is built around the core to present results more user-friendly.

  The central problem: the first prototype with GPT-5 hallucinates massively – it invents papers that don’t exist or links to non-existent pages. In the third step, this is solved by integrating the Consensus MCP server, which acts as tool access to real, peer-reviewed scientific papers. The agent is converted from a pure chat model to a LangChain agent that can use tools provided by the MCP server. After integration, the system finds actually existing papers from the desired time period with real links, factually correct metrics, and verifiable trends – a night-and-day difference in quality.

  Tools/Models Used and Format: LangChain, OpenAI GPT-5, Pydantic, Consensus MCP Server – tutorial.

• Claude Code: All Features For Power Users
  20.4.2026, 16:01:08

  Claude Code: All Features for Power Users

  The video is a comprehensive crash course on Claude Code covering advanced features for power users and starting with the basics.

  Installation and Basic Usage: Claude Code is a command-line and desktop-based application that functions as an agent engineer – it can execute commands, read and edit files with user permission. Installation varies by operating system (package managers, desktop app, VS Code extension).

  Slash Commands: The most important are /model (switch models between Opus 4.6, Sonnet, and Haiku), /rewind (undo function), /resume (resume session), /usage (quota overview), /stats (usage statistics), /config (settings), and /insights (detailed analysis of your workflow with browser report).

  Claude MD Files: Markdown files (claude.md) containing project-wide or global instructions – coding style, project structure, packages, best practices. They load at startup when relevant.

  Skills: Fragmented version of claude.md – separate files in .claude/skills/, each with instructions for specific tools (e.g., Python package installation, Docker). Claude loads only relevant skills contextually, saving context budget.

  MCP (Model Context Protocol): Connection to external APIs (Deep Wiki, Google Calendar, Slack, etc.) via claude mcp add. Enables Claude to interact with new data sources and services.

  Plugins: Pre-built packages from a marketplace combining skills, MCP servers, and tools. Install with /plugins discover or /plugins install.

  Sub-Agents: With /agents create specialized agents (e.g., UI Designer, Backend Specialist) to which you selectively grant or deny tools. The master agent coordinates them – communication runs through it.

  Agent Teams (experimental): Environment variable CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 enables decentralized teams of agents (e.g., UX, Architect, Devil’s Advocate) communicating with each other without going through a master. Significantly more expensive token-wise.

  Work Trees: Git feature that Claude Code uses to work on multiple isolated branches of the same project in parallel in different directories – useful for A/B testing or parallel feature development.

  Memories: Claude stores information (e.g., usernames) across sessions when you explicitly instruct it to remember so they’re available in future sessions.

  Claude Code (Opus 4.6 as default model); deep-dive/tutorial format with practical demos.

Nic Conley (1 new video)

• Easiest Hermes Agent Setup (better than OpenClaw)
  21.4.2026, 13:38:32

  The video compares two open-source AI agents, OpenClaw and Hermes, developed by Nous Research. The main difference is that the Hermes agent features a built-in learning loop that makes it smarter over time and adapts its capabilities to the user’s workflow. In contrast, OpenClaw often forgets previous tasks and starts fresh each session, which can be frustrating. Both agents can run on your own machine or a VPS and be accessed through chat apps like WhatsApp or Telegram. However, OpenClaw has a larger ecosystem with over 247,000 GitHub stars and more than 700 community skills, while the Hermes agent is ready to use out of the box and already includes 40 integrated tools. Another advantage of the Hermes agent is its higher security, since the integrated skills come from the developers themselves rather than third parties, making it less vulnerable to malware. However, the Hermes agent is a newer project that hasn’t been tested as extensively as OpenClaw. The video also demonstrates how to set up the Hermes agent on a VPS and connect it to Telegram to use it as a personal assistant. The Hermes agent can automate tasks that would normally be handled by an employee, such as researching events for a newsletter. The Hermes agent is suitable for intermediate and advanced users looking for a powerful, learning-capable AI solution, while OpenClaw is better suited for tech-savvy users who prefer a larger ecosystem and more customization options.

Nick Saraev

No new videos in this period.

Niklas Steenfatt (1 new video)

• A NEW ERA
  19.4.2026, 12:46:18

  Summary: A New Era

  The creator tests Hermes, a newly emerging open-source AI agent project from Nous Research, which is praised in the comments as a potential successor to OpenCore. He compares both systems practically to clarify whether a switch makes sense.

  Hermes setup and first impressions: Hermes is set up on a Hostinger VPS with pre-installed Docker support, connected to the Anthropic Claude model via API key. The web interface has a vintage design and feels more user-friendly than OpenCore.

  Core difference in concepts: OpenCore focuses on integration with other tools (Tool Calling), while Hermes relies on adaptive learning – the difference is thus “Tool Calling versus Learning Loop”.

  Learning Loop as unique selling point: After each completed task, Hermes automatically evaluates its performance and creates new skills independently. This becomes evident when Hermes develops a specialized YouTube skill with fallback mechanisms (e.g., NoGPIO as a workaround) after a YouTube transcription task.

  Visual showcasing: On request, Hermes creates impressive Excalidraw diagrams to visualize its functionality. OpenCore (Amadeus) can likewise generate such visualizations, though it becomes clear that quality here depends more on the underlying language model (Sonnet 4.6) than on the agent itself.

  User experience: Hermes feels subjectively “smoother” – through step-by-step visualization of agent actions, emojis, and cooler slash commands (e.g., /by-the-way for follow-up questions). Nevertheless, the creator keeps Amadeus (OpenCore) for critical tasks, as this agent already has more contextual knowledge about him through longer collaboration and saved markdown notes.

  Knowledge transfer: The creator copies the OpenCore data directory into the Hermes container, so Hermes gains access to all collected notes and contexts from Amadeus – an interesting experiment in memory transfer between agents.

  Conclusion: Both tools are practical and very similar. Hermes is not the absolute breakthrough, but at least equivalent to OpenCore. The creator plans to deploy both agents together via Paperclip and have them cooperate.

  Explicitly discussed AI tools/agents: Hermes, OpenCore, Anthropic Claude (Sonnet), Paperclip, Excalidraw (as integration) — demo/tutorial hybrid with practical comparison elements.

No Priors: AI, Machine Learning, Tech, & Startups (1 new video)

• SAP: Bringing the ‘Operating System’ of a Company into the AI Era with CTO Philipp Herzig
  23.4.2026, 10:00:18

  The video is an interview with Philip Herzig, SAP’s CTO, discussing SAP’s AI strategy, the challenges of scaling, the importance of AI for business models, and the future of enterprise software. Herzig emphasizes that SAP has endured through multiple technology cycles because the company has focused on delivering results for customers. He explains that SAP offers the broadest range of enterprise software applications and platforms, spanning financial and HR systems to supply chain and logistics management. Herzig also discusses the integration of AI into SAP’s products and the necessity of designing AI to be scalable and secure. He mentions that SAP has already introduced products like “Jewel for Consulting,” which can increase consultant efficiency by 30%. Herzig identifies three key areas where AI will massively transform enterprise software: user interfaces, business processes, and the data layer. He stresses that the greatest challenge lies in designing AI to achieve the right outcomes within enterprises. Herzig speaks about the importance of verifiable results and the need to design AI so it can integrate with existing systems and data. He sees significant opportunities in combining structured and unstructured data to make better decisions and automate business processes. Herzig believes AI will fundamentally change how companies operate by automating repetitive tasks and enabling employees to focus on more strategic work. He sees SAP well-positioned to benefit from this transformation because the company focuses on delivering customer results and has a flexible architecture that enables AI integration. Herzig emphasizes that the biggest challenges for enterprises adopting AI are data integration, solution scaling, and system security. He sees SAP well-positioned to address these challenges because the company has a broad range of applications and platforms already deployed across many enterprises.

  The video explicitly addresses SAP and OpenAI and is aimed more at Intermediate and Advanced levels.

Productive Dude (7 new videos)

• ChatGPT 5.5 “Spud” just dropped… did it destroy Opus 4.7?
  24.4.2026, 12:01:55

  The video provides a detailed analysis and first impressions of GPT 5.5, which was released on April 23, 2026. The author compares the model with other AI models, particularly Claude models from Anthropic, and highlights the strengths and weaknesses of GPT 5.5. Key points include the model’s improved ability to handle tasks autonomously without constant supervision, as well as its strengths in areas such as code writing and debugging, online research, data analysis, document creation, and software operation. The author tests GPT 5.5 live in Codeex and demonstrates how the model can handle multiple complex tasks simultaneously, including creating a dungeon game, a landing page, and an artificial life simulation. The benchmarks show that GPT 5.5 performs well in areas like long context, terminal work, and complex reasoning tasks, while Claude 4.7 is superior in software development and agent-based tasks. The author concludes that GPT 5.5 is particularly suitable for long-context tasks, terminal work, and challenging reasoning problems, while Claude 4.7 is better for everyday software development and agent-based tasks.

  Final comment: The video explicitly covers GPT 5.5 from OpenAI and Claude 4.7 from Anthropic and is aimed more at intermediate and advanced users.

• Claude Live Artifacts just dropped… ALL your tools in one place!
  23.4.2026, 13:00:18

  The video demonstrates how to create a CRM system using Claude’s new Live Artifacts feature, which aggregates data from various sources such as Notion, HubSpot, Google Sheets, and Gmail and displays it in a unified dashboard. The process begins by connecting desired apps through the connector function in Claude. You then create a live artifact through targeted prompts that retrieves and displays data in real time. The user demonstrates creating a CRM dashboard with filters for different lead categories and an integrated email function controlled through Gmail. After some adjustments and troubleshooting, the finished dashboard displays all leads from Notion and allows direct email composition to contacts. The video emphasizes the versatility and efficiency of the Live Artifacts feature for data aggregation and analysis.

  The video explicitly covers Claude and is aimed more at intermediate or advanced users.

• ChatGPT Images 2.0 just dropped… text, transparency and more! 🤯
  21.4.2026, 22:40:54

  The video introduces the latest version of OpenAI’s image model, Image V2, and evaluates its capabilities and limitations. The creator highlights that the model is capable of following detailed instructions, placing objects precisely, and rendering dense text, although with some limitations regarding text consistency. A special feature is the “thinking” function, which enables the model to analyze and optimize prompts, resulting in better outcomes. The model demonstrates strengths in creating UIs, magazine collages, handwritten essays, and photorealistic images. It supports various aspect ratios and is available in multiple languages. The creator tests the model with various tasks, such as creating website mockups, redesigning YouTube thumbnails, and generating images with transparent backgrounds. It becomes apparent that while the model can deliver impressive results, it also has some weaknesses, particularly in rendering text and precisely implementing instructions. At the end, the creator attempts to create a thumbnail for the current video using Image V2 to demonstrate the practical applicability of the model.

  The video explicitly covers OpenAI’s Image V2 and is more suitable for intermediate users who already have experience with AI tools.

• Claude + Higgsfield: Build Beautiful ANIMATED Websites! (Claude Tutorial)
  21.4.2026, 13:00:45

  The video shows how to quickly create and deploy an animated website using AI tools like Higgs Field and Claude Design. The process begins by creating images and videos in Higgs Field, where an image is converted into a video with an animated scene. This video is then converted to WebM format and uploaded to Claude Design, where a landing page is created with the video and additional elements such as text and cards. The finished website can then be deployed via Netlify, with the entire process completed in less than 5 minutes. The creator emphasizes the simplicity and speed of the procedure and suggests using such websites as an outreach method to attract potential customers.

  Final comment: The video explicitly covers Higgs Field, Claude Design, Claude Code, and Netlify and is aimed more at intermediate and advanced users.

• Claude Routines Tutorial: Automate Anything with Claude Ai!
  20.4.2026, 13:00:19

  The video explains how to use Claude Routines, a new Claude feature that allows you to create and execute automated tasks. Routines can be triggered through API calls, schedules, or webhooks and connected to various tools like Slack, Gmail, and Google Calendar. The user shows how to create a routine to check Google Calendar and send daily meeting notifications via Slack. Additionally, it demonstrates how to set up a routine with an API trigger to process leads, conduct research, log into a CRM, and create an email draft. The steps include setting up forms, connecting APIs, and testing the automation.

  The video explicitly covers Claude and is aimed more at intermediate users.

• Claude + Heygen: INSANE Automated Videos! (easy tutorial)
  19.4.2026, 12:00:22

  Summary: Claude + HeyGen – Creating Automated Videos

  Carter Svrcek demonstrates an end-to-end system for fully automated video production using Claude Design, Claude Code, and HeyGen.

  Motion Graphics with Claude Design: Through natural language prompts, Claude Design creates animated motion graphics assets (lower thirds, titles, transitions) with green backgrounds. The generated project is exported as a ZIP file – the greenscreen basis enables easy background removal in video editing software like Premiere using keying tools.

  Rendering with Hyperframes: The ZIP file is loaded into Claude Code. There, you use the HeyGen Hyperframes skill to open the project as HTML, test it in preview, and ultimately export it as video.

  Fully Automated Video: Through HeyGen for Agents, Claude Code can orchestrate a complete video project: it researches current news, writes a script, uses a created avatar with specific voice, integrates the motion graphics overlays, removes the green background, and renders the final video – all in one prompt. For this, the HeyGen API is configured with local environment variables (avatar ID, voice ID, API key).

  Result: The system delivers a professionally-looking video with news content, avatar movements, and animated graphics within a short time. The quality significantly surpasses previous AI avatar solutions, though the creator recommends using “source avatars” for even better results.

  Tools/Providers: Claude (Claude Design, Claude Code), HeyGen (Hyperframes, Avatar skill, API), Anthropic; Format: Tutorial with demo.

• 23 INSANE Claude Code Skills… to build your SaaS! (from Gary Tan)
  19.4.2026, 00:11:31

  Summary: 23 Claude Code Skills from the G Stack by Gary Tan

  The G Stack is a set of 23 Markdown files published by Gary Tan (CEO of Y Combinator) with specialized skills for different roles (CEOs, designers, engineers, managers, QA). With 76,000 GitHub stars, it presents itself as a potential replacement for an entire software engineering team.

  Installation and Setup: The G Stack is installed via Claude Code by cloning the repository and then opening two project folders – one with the G Stack itself, one with your own project (in the demo: Lerty, an early-stage SaaS with 10K ARR).

  Most Important Applied Skills:

  • Office Hours: Conducts a consultative process, asks critical questions about the project, and generates concrete results (design docs, GitHub issues, product improvement suggestions) – not only evaluates positively but questions fundamental assumptions.
  • Auto Plan: Combines four specialized perspectives (CEO, designer, engineer, developer experience) in sequential analysis with automated decisions based on six defined principles; produces structured plans with four different verdicts and action items.
  • Design Shotgun: Generates multiple parallel design variants and displays them side-by-side in the browser for iteration – works as a brainstorming partner for visual components.

  The G Stack also organizes skills into categories like strategy & planning, design, quality & debugging, shipping & releasing, and upgrade management. Tips on prompt engineering (e.g., formatting gating questions with Claude’s question-answer tool) improve user-friendliness for many inquiries.

  Claude Code, G Stack (by Gary Tan/Y Combinator), opinion/reflection with application demo.

Sebastien Dubois

No new videos in this period.

Simone Rizzo (2 new videos)

• GPT 5.5 batte Opus 4.7? La verità dopo i test
  24.4.2026, 11:48:40

  GPT 5.5 vs. Opus 4.7 – Test comparison

  OpenAI has released GPT 5.5, an incremental update to its predecessor GPT 5.4. According to OpenAI’s official article from April 23, 2026, the model is distinguished by better understanding of user intent – it works more independently, requires less user guidance, and excels at code debugging, research, data analysis, and office tasks. GPT 5.5 can also handle unstructured multitasking requests, planning, tool use, and ambiguity management.

  On benchmarks (Artificial Analysis Intelligence Index), GPT 5.5 ranks first, with Opus 4.7 Max in third place. A major advantage: GPT 5.5 operates more efficiently – for example, on Terminal Bench 2.0 it requires only 14,500 instead of 18,000 tokens while delivering better results simultaneously. Particular improvements are evident in Genetic Coding, Computer-Use, and Knowledge Work.

  Practical tests with Open Code (to create fairer conditions than a cloud-vs.-ChatGPT comparison):

  1. Neuroflow website: Both models produced similar results. GPT 5.5 used 21,881 tokens in 2 minutes (€0.66), Opus 32,753 tokens in 9 minutes (€1.00). Tie.
  1. Solar system simulation: Opus performed slightly better (better shadows, more stable performance). GPT 5.5: 16,000 tokens / 5 min.; Opus: 23,000 tokens / 1 min. Slight edge to Opus.
  1. FPS zombie game: Opus impressed with better graphics, weapon design, and game physics. GPT 5.5: 23,000 tokens / 9 min.; Opus: 43,460 tokens / 7 min. Clear win for Opus.
  1. Evolutionary ecosystem game: GPT 5.5 chose Python (better simulation management), Opus chose Browser/JavaScript. GPT 5.5: 16,000 tokens / 4 min.; Opus: 31,000 tokens / 2 min. Slight edge to GPT 5.5.

  Overall assessment: Opus is on average 2.25× faster, GPT 5.5 produces 67% fewer tokens. Opus costs approximately 39% more per test. In terms of quality, both are on a similar level – barely noticeable to average users.

  Pricing criticism: Despite lower token consumption, OpenAI doubled output costs (from $15 to $30 per 1 million tokens). GPT 5.5 Pro costs even more ($180 output). The savings from fewer tokens are thus partially offset by higher prices.

  Tester’s conclusion: GPT 5.5 is an incremental update without groundbreaking new architecture or features. Performance and quality rank on par with Opus 4.7 – for specialized developers and researchers, a difference might emerge in complex scenarios, but for average users the difference is marginal.

  Explicit models/providers: OpenAI GPT 5.5, Anthropic Opus 4.7, Gemini 3.1 Pro, additional Chinese models mentioned (Autarch, Kimi K2); test tool: Open Code — demo/comparison.

• RAG è già vecchio? Il nuovo modo di costruire Knowledge Base per AI con LLM Wiki
  21.4.2026, 17:50:31

  Summary: Is RAG outdated? LLM Wiki as a new knowledge base methodology

  The video describes the evolution of three generations for managing knowledge databases for Large Language Models, with LLM Wiki presented as the newest method, developed by Andrew Carpenti.

  The three generations at a glance

  Generation 1 – RAG (Retrieval Augmented Generation, 2022): Documents are broken into chunks, converted to vectors via an embedding model, and stored in a vector database. When a query comes in, it’s also vectorized and similar document blocks are retrieved via vector search, then augmented to the LLM. Problem: context loss through chunking, invisible cross-references, similarity ≠ relevance, no knowledge accumulation across queries.

  Generation 2 – Agentic File Search: The system is based on the file system rather than vector databases. The agent intelligently navigates folder structures and files (in Markdown format), can run multiple agents in parallel, and selectively reads only relevant documents in full – saving tokens. Cross-references can be followed at runtime, and the file hierarchy becomes part of the model’s reasoning.

  Generation 3 – LLM Wiki: Knowledge is compiled once into a persistent artifact (a wiki structure) and then continuously updated – like Wikipedia. No reconstruction needed for each query. The agent navigates through linked Markdown files that represent a knowledge graph.

  Practical setup with LLM Wiki

  Architecture in three layers:

  1. Raw layer: Raw data (PDFs, CSVs, images, web content)
  2. Wiki layer: Structured Markdown files with concepts, entities, syntheses, index, and log
  3. Schema layer: Cloud.md file with rules for the agent

  Workflow:

  • Data Ingest: Load documents into /raw (via web clipper extension for browser or manually), agent structures these in wiki format
  • Queries (Q&A): Agent reads index, jumps between linked documents, answers based on local data
  • Linting: Automatic health checks of the wiki – finds inconsistent data, missing links, orphaned pages

  IDE & visualization: Obsidian (free for Windows/Mac/Linux) visualizes the files and offers a graph view showing the network. Optionally, the agent can format answers as MARP markdown presentations displayed via Obsidian plugin.

  Practical example in the video

  The speaker creates a wiki on “diets” live: loading multiple web articles and PDFs with public funding options into the raw directory, having the agent ingest these, and then asking questions like “Best diet for brain performance?” or “Which funding fits my AI consulting firm?” The agent automatically navigates through linked pages, pulls information together, and can export the answer as a presentation.

  Keys to success

  The core principle: the tedious work isn’t thinking, it’s bookkeeping (cross-references, consistency, updates). LLMs don’t get bored – they can effortlessly rework 15–30 files at once while humans hand off wikis and forget them. Therefore LLM Wiki is ideal for personal notebooks, enterprise documentation, legal/scientific collections, internal team wikis, and competitive analysis.

  Featured systems: RAG, Agentic File Search, LLM Wiki (Andrew Carpenti), Obsidian, Claude/Open Code (as coding agent), OpenClow with Memory Wiki Plugin; format: tutorial with live demo.

Tech With Tim (5 New Videos)

• Local Models Got a HUGE Upgrade – Full Guide (Ollama/OpenClaw)
  24.4.2026, 14:14:05

  Local Models Got a HUGE Upgrade – Full Guide

  This video shows how to run local AI models on your own computer and connect them with OpenClaw to become independent from cloud services and save costs.

  Hardware Requirements: Choosing the right model and hardware is crucial. On Mac, overall system specifications are relevant (viewable in the About window), while on Windows/Linux, the VRAM of your Nvidia GPU matters. Local models can use approximately 100% of these resources, which means older hardware (6+ years) is practically not viable.

  Installation with Ollama: The free tool Ollama is installed via terminal command and lets you download and run models locally. The video recommends Gemma 4 as the current best option with tool support for OpenClaw. When selecting a model, you should match the size (e.g., 9GB, 20GB) with your available RAM – larger models deliver better performance but require more storage and run slower.

  OpenClaw Integration: After ollama pull gemma4:31b (or similar), the model is configured with openclaw configure. You select Ollama as the provider, set it to “local only”, confirm the models, and run openclaw gateway restart. After that, the model is available in OpenClaw and can be switched using /model [modelname].

  Hybrid Approach: The video recommends using local models as your default and adding cloud models (like OpenAI or Anthropic) only for more complex tasks – either for maximum speed/cost savings or as a combination for optimal results.

  Bonus: The provider Link Ventures runs a program called ClawComp, which offers a free Mac mini for three months to develop with OpenClaw and local models (application deadline May 8, concluding with “Claw Week” in Cambridge).

  Ollama and OpenClaw were covered – tutorial format with technical focus for beginner to advanced users.

• Hermes Agent Full Tutorial for Beginners | Setup Guide
  23.4.2026, 14:00:05

  Hermes Agent – Setup Guide for Beginners

  Hermes Agent is a self-learning AI agent platform by developer Nous Research, distinguished primarily by an integrated learning loop that sets it apart from competing systems like Open Claw. The system continuously improves itself, creates new skills, plans tasks, and updates its memories. Unlike Open Claw, Hermes gets better the more you use it, while other systems can degrade over time.

  Security and Deployment: The agent has access to your entire computer (browser, terminal, file system). For this reason, running it on a dedicated machine is recommended – ideally a virtual private server (VPS), as it’s cheaper, isolated, and runs reliably. This tutorial uses Hostinger with one-click installation.

  Installation and Configuration: After VPS deployment, configuration happens via an interactive web dashboard with login credentials. OpenAI Codex is recommended as the model provider. For messaging integration, Telegram is suggested: you create a new Telegram bot via BotFather, receive a bot token, and configure allowed user IDs (at least one to keep the bot private). The bot can then be tested via Terminal UI with hello world.

  Gateway and Messaging: To reach the bot from outside (e.g., Telegram), the gateway must run with the command hermes gateway. Without an active gateway, the bot won’t respond to external messages.

  Adding Capabilities:

  • Firecrawl for Web Access: API key is set via hermes config set firecrawl_api_key; enables web searches
  • Voice Function: Installing required dependencies via pip install hermes-agent-all and system libraries enables voice input/output in Telegram
  • Cron Jobs: The bot can schedule recurring tasks (e.g., daily at 9am to research and deliver a morning briefing with top 3 AI trends)

  Skills Concept: Skills are reusable workflows that the agent references. You can assign skills directly to the bot ("Please create this skill for me") or it can develop them on its own.

  Additional Features: hermes setup allows configuration changes; hermes doctor diagnoses issues. The bot stores session histories and memory artifacts and can retrieve past conversations.

  Core idea: After basic setup, work creatively with the bot – it improves through use, creates its own skills, refines memories, and adapts to your needs.

  Explicit Tools/Providers: Hermes Agent (Nous Research), OpenAI Codex, Telegram, Firecrawl, Hostinger VPS, Whisper Flow (voice dictation, mentioned briefly), Open Claw (comparison), Eleven Labs (optional for text-to-speech) — Tutorial, beginner-friendly.

• Build an AI Email Assistant with Code | Full AI Tutorial
  22.4.2026, 13:00:17

  This video shows how to build an AI email assistant that replies to emails and sends automated messages.

  Final Product Demo: The assistant receives emails, automatically sends a welcome message on first contact from each contact, and replies with AI-generated text. A dashboard tracks all contacts and email activity; users can send broadcast emails to all collected contacts.

  Tech Stack: Next.js, Postmark (email service), Anthropic API (Claude), Better SQLite 3 (database), Drizzle ORM.

  Main Implementation Steps:

  1. Project Setup: Create a new Next.js app with TypeScript, ESLint, and Tailwind CSS; install dependencies (Postmark SDK, Anthropic SDK, Drizzle, Better SQLite 3).
  1. Configuration: Create a Drizzle config file and .env file for API keys (Postmark token, Anthropic API key, sender email).
  1. Postmark Setup: Create a server in the Postmark dashboard, verify sender signature (via email confirmation), configure inbound webhook URL later.
  1. Database Schema: Define two tables – messages (thread ID, role, content, timestamp) and contacts (email, name, welcome status, timestamp).
  1. Database Functions: Write saveMessage, getRecentMessageLog, upsertContactFromInbound, markWelcomeSent, shouldSendWelcome, getAllContacts.
  1. Email Handling: Create POST endpoint /api/inbound that receives inbound payloads from Postmark, extracts email details, sends welcome message if needed, calls Anthropic API (with system prompt), saves response, and sends reply email.
  1. Webhook Management: Use Ngrok to make localhost server publicly accessible; enter Ngrok URL + /api/inbound in Postmark as webhook.
  1. Testing: Send an email to the Postmark inbound address, observe webhook triggering in the terminal.

  The tutorial is structured as a step-by-step guide; complete code is available on GitHub. The host notes that the tutorial assumes basic knowledge of JavaScript, TypeScript, and Next.js.

  Explicitly Covered: Anthropic/Claude (API), Postmark (service), Next.js, Drizzle (ORM), Better SQLite 3, Ngrok — Tutorial format for intermediate developers.

• I Built an AI Agent in 20 Minutes – Here’s How
  20.4.2026, 13:53:04

  The creator shows how to build a functional AI agent with MaxClaw in a short time – a hosted version of OpenClaw on MiniMax infrastructure. His girlfriend uses such an agent daily on her smartphone for research and planning without any technical knowledge.

  The MaxClaw System:

  MaxClaw eliminates complex manual configuration normally required (API keys, model deployment, MCP config, security). Instead, it offers out-of-the-box tools like web search, image generation, video generation, code execution, and web scraping, plus over 10,000 pre-built experts to choose from. The platform works on web, desktop, mobile, Telegram, Discord, and Slack.

  The Underlying Model M2.7:

  The MiniMax model M2.7 stands out for its training process – the model improved its own training infrastructure across 100 iterations without human intervention and achieved a 30% improvement. It reaches 97% compliance on tool calling with 20-40+ tools, while comparable models only achieve 74%. This makes it competitive with frontier models like Opus and GPT but runs more affordably.

  Pricing:

  MiniMax offers credit-based plans starting at $19/month (with an additional 200 daily free credits) or a token plan starting at $10-20/month offering multimodal access (text, image, video, speech) with generous limits. The creator uses a +20 plan and gets 12% discount through his link.

  Practical Demos:

  The creator deploys several MaxClaw agents with different pre-built experts: (1) an image generation expert that creates a cartoon portrait of “Tech with Tim” based on web research; (2) a trend intelligence expert that aggregates current AI and OpenClaw news and can set up automated reports on request; (3) a default expert with coding skills that generates a landing page with AI images and generated video within seconds.

  In all cases, tools are automatically integrated – no manual configuration needed.

  Assessment:

  The creator considers MaxClaw ideal when you need a mid-complexity AI agent quickly without diving into technical depth. For super advanced, highly customizable setups, a self-hosted OpenClaw is better. Cost-wise, MaxClaw is significantly cheaper than running your own OpenClaw instance since hosting and all modality usage are included in one plan (versus separate API subscriptions that quickly cost $30-200+/month).

  Explicit Tools/Providers: MiniMax, OpenClaw, M2.7 model, MaxClaw platform. Format: Demo/Tutorial.

• I Replaced OpenClaw With Perplexity Computer…
  19.4.2026, 12:36:24

  Summary: Perplexity Computer vs. OpenClaw

  Perplexity has released a new Computer feature that functions as an agent tool without configuration overhead – unlike OpenClaw, which requires extensive setup. The Computer feature is cloud-native, costs $200/month for the Max plan, and requires no API key management or manual model selection.

  Comparison of Approaches:

  • OpenClaw: Full local control, infinitely configurable, but time-consuming to set up, prone to errors and security issues. The author notes that even experienced developers spend hours on setup.
  • Perplexity Computer: Cloud-based, zero configuration, automatic orchestration between frontier models (Opus, Gemini, Grok, etc.), parallel sub-agents, isolated compute environment, self-healing, always available – even without running your machine.

  Practical Demos:

  The author shows two examples: (1) He asks Computer for the top 5 AI agent tools with pricing and strengths/weaknesses – it generates a complete interactive web application with comparison charts and dark mode within 8:25 minutes. (2) A financial dashboard for Nvidia, Microsoft, and Anthropic with stock prices, analyst sentiment, and news – completed in 12 minutes.

  A third example demonstrates Slack integration: The author connects Google Drive, Gmail, and Slack, instructs Computer to monitor his inbox each morning, summarize important emails, draft replies, and format everything into a Notion dashboard. Computer responds directly in Slack and can create a test briefing immediately.

  Author’s Usage Conclusion:

  If you already use and configure OpenClaw productively, you won’t switch entirely – but for quick one-off tasks (reports, dashboards, ad-hoc automation), the author sees Perplexity Computer as superior because no setup is required. For recurring business automations, OpenClaw remains the better choice since it can be configured specifically for your needs. OpenClaw also offers cost savings for light users (e.g., $10/month VPS + $5 tokens versus $200/month).

  Explicit Tools: OpenClaw, Perplexity Computer, Anthropic (Claude/Opus), Google Gemini, Grok — Demo/Comparison

TheAIGRID (5 new videos)

• How To Use GPT 5.5 For Beginners – GPT 5.5 Tutorial
  24.4.2026, 17:56:32

  Summary: How To Use GPT 5.5 For Beginners – GPT 5.5 Tutorial

  The video shows beginners how to use GPT 5.5 optimally. The first important step is to enable the model in settings (top right → Configure) and ensure “Extended” is selected instead of “Standard” – “Standard” is sufficient for simple queries but shouldn’t be permanently enabled to save costs.

  A common beginner mistake is using GPT 5.5 in the native ChatGPT interface, which limits the model’s capabilities. Instead, you should use Codex (via the sidebar or app download) to implement actual tasks like creating Excel files, PowerPoint presentations, and web apps. In Codex, you should first configure settings to “same power and less technical details” and set the model to GPT 5.5 with medium reasoning level.

  The video demonstrates concrete examples: An Excel file for business planning was created in about 3 minutes, a PowerPoint presentation in 4–5 minutes. The Plan Mode function (Plus button) is particularly valuable for complex tasks since it creates a work plan you can review before execution and before credits are consumed. Other options include web apps and agentic tasks like website testing with Browser Use (via plugins). Important: Unlike the standard ChatGPT interface, Codex continues running even if you open a new chat. For difficult tasks, the instructor recommends increasing reasoning to “High” or “Extra High”. Finally, it’s noted that GPT 5.5 is more expensive and you should monitor rate limits in settings.

  OpenAI GPT 5.5 and Codex; beginner tutorial.

• How To Use ChatGPT Agents – Workspace Agents Tutorial
  23.4.2026, 03:45:11

  Summary: How To Use ChatGPT Agents – Workspace Agents Tutorial

  Accessing Agents: ChatGPT Agents are only accessible via the Business version – you must upgrade your account and then go to the Business section instead of your personal area. You’ll find the Agents tab in the sidebar.

  What is an Agent? An agent is a system that follows instructions, uses skills, works across multiple tools, and executes actions. It transforms a request through a structured process into completed work (e.g., writing follow-up drafts, preparing CRM notes, readying for review).

  Creating an Agent: You click “Create Agent”, can use templates or describe in natural language what the agent should do. The Agent Builder thinks through the best structure in real time. You authenticate yourself in required apps directly in the chat and can set a schedule for the agent. Important: the more context in the prompt, the better the initial draft.

  Testing: Before going live, you should conduct evaluations (Evals) – realistic test cases (common requests), messy inputs (incomplete/conflicting), and edge cases (unusual but possible queries). If something fails, you should specifically tell the agent what went wrong so it can update its instructions.

  Practical Examples:

  • Product Feedback Agent: Collects feedback from various sources, summarizes recurring issues, posts daily updates to Slack, and creates tickets in the ticketing system.
  • Reporting Agent: Creates weekly reports from Google Drive, uses Skills for consistent metric definitions, runs automatically on schedule.
  • Sales Agent (Spark): Researches new leads, scores them, sends outreach emails, automatically schedules follow-ups.
  • Vendor Risk Manager (Trove): Conducts due diligence, uses existing Finance Skills, creates structured risk assessment reports in minutes.

  Key Concepts: Agents can work with Skills (reusable process definitions), run on schedule or be manually triggered, have connections to apps (permissions are crucial), and their run history shows all steps and outputs for transparency.

  OpenAI as provider, ChatGPT platform: tutorial with practical demo examples.

• ChatGPT Images 2 Tutorial For Beginners (With New Tips And Tricks)
  21.4.2026, 22:11:11

  ChatGPT Images 2 Tutorial for Beginners

  The tutorial shows the key features and beginner mistakes when using ChatGPT Images 2.0.

  Basics and Templates: After opening the Images section, beginners can start with templates. Important: templates don’t use the uploaded image as the driving image, only the style – the prompt generates the final image.

  Precise Image Editing: Instead of simply writing “remove XYZ”, you should use the Select function, hover over the element to be changed, and then give a precise instruction (e.g., “replace with tail”). This makes editing significantly more accurate.

  Aspect Ratio: The aspect ratio should be mentioned in the prompt before generation (e.g., square format for Instagram) since ChatGPT doesn’t have a standard checkbox for it. Changes afterward are possible but less efficient.

  Structured Prompts Instead of Slot-Machine Approach: The new model can research and conceptualize. Instead of “make me an ad”, you should mention concrete work steps: “Research the latest OpenAI merch, identify the most interesting items, estimate resale values, then create a polished mock-up.” This way you use the model like a designer who can search the internet.

  Exact Placement Instructions: The model follows precise spatial instructions better. Example structure: “Create a photorealistic object. Place object 1 exactly in the center, object 2 directly next to it. Text says exactly [what]. No additional words. Layout clean and readable.” This approach is crucial for thumbnails, product mock-ups, and comparison graphics.

  Multi-Image Outputs Instead of Single Images: Beginners often create only individual images – but you can generate consistent series: multiple infographic pages, slide sequences, or complete academic posters. PDFs, YouTube scripts, or Notion docs can be uploaded directly and converted into visual material.

  PNG Icons with Transparent Background: The model can generate transparent icons that can be imported directly into editing programs – saving the otherwise necessary detour through external background removal tools.

  The video emphasizes that precise, structured prompts and multi-asset thinking make the difference between beginner and professional use.

  OpenAI / ChatGPT Images 2.0 – Tutorial.

• GPT 5.5 Just Leaked… And It’s Scary Good
  20.4.2026, 17:24:21

  GPT 5.5 (or “Spud”) – Leaked Information About OpenAI’s Next Model

  The video summarizes multiple leaked pieces of information about OpenAI’s upcoming model GPT 5.5 Pro (also called “Spud”). Greg Brockman from OpenAI discusses expected improvements: the model should solve significantly more complex problems, work more nuancedly, and understand contexts better. Brockman emphasizes the concept of “Big Model Smell” – the subjective feeling of qualitative superiority when an AI model is genuinely significantly more capable.

  The model is based on two years of pretraining work and is therefore not an incremental update but a new foundation – which suggests a capability leap. People who have already tested Spud describe it as on par with Anthropic’s Opus 4.7.

  For expected benchmarks, a performance increase of approximately 10–15% is predicted, which could put OpenAI ahead of Anthropic in several areas. Possible native multimodality is being discussed. Particularly relevant for enterprise: Spud should develop into an autonomous “Digital Worker” with improved computer use – advanced over GPT 4.5, which still required more oversight and focused on coding.

  Leaked outputs show impressive results: the model creates demanding applications and HTML games sometimes on the first try with higher coherence than previous versions. A new image generation model (Images V2) is set to launch in parallel and shows in test shots better taste in styling and detail accuracy than competing models – for example with gaming console styles and complex prompts with color and material specifications.

  Format: Opinion/reflection on leaked information about OpenAI GPT 5.5 / Spud; mentioned competitors Anthropic (Opus 4.7) and various external test sources.

• How To Use Claude Design – Tips And Tricks Most People Missed
  20.4.2026, 01:14:07

  Claude Design – Tips and Tricks Overview

  Claude Design is Anthropic’s new visual tool based on the Opus 4.7 model that creates not just static mock-ups but code-based, interactive prototypes – clickable interfaces, slide decks, landing pages, and experiences with voice, video, 3D, and shaders. The tool doesn’t replace Figma but the blank canvas: you use it for the initial draft and then export to Figma or Canva.

  Access and Usage: Accessible at claude.ai/design, but requires a Pro or Max subscription (not available for free). Claude Design has a separate usage limit that you should monitor in settings – many users quickly hit the ceiling.

  Building a Design System: Before starting, you should create a design system to align the AI with your brand. Best to upload brand images to a chat prompt with Claude and generate a style/brand guideline, then upload fonts and additional details and review all points under “needs review”. This takes initial effort but saves massive tokens and usage later.

  Workflow – Wireframe to High Fidelity: For a new design, Claude asks many details to work better (name, value proposition, features, style). With a wireframe version you can first test the structure and think before moving to the complete high-fidelity version. The “Edit” tool enables granular changes without big token waste (changing colors, text, numbers, drawing).

  Tweaks Feature: The tweaks system enables design variations through buttons without re-rendering the whole thing – saves tokens and allows quick iteration through sliders for hue, glow, size, position, etc.

  Important Pitfalls: Animations are extremely compute-intensive and quickly consume all your usage; generally the Max subscription is recommended. Claude has predefined “Skills” (Animated Video, Interactive Prototype, Slide Decks, Wireframe, PowerPoint, PDF Export) – you should activate these before the prompt to get better outputs. Always select the correct design system.

  Export Options: After editing you can export to various formats – to Figma, Canva, and other programs.

  Claude Design (Anthropic), demo/tutorial.

Theo – t3․gg (5 new videos)

• Why people spend millions on Github stars
  25.4.2026, 14:52:41

  Summary: Why people spend millions on Github stars

  The video analyzes a massive problem: millions of fake GitHub stars are being purchased to influence startup funding. Researchers from Carnegie Mellon, North Carolina State University, and Socket identified approximately 6 million suspicious stars distributed across 18,600 repositories through around 301,000 accounts. The infrastructure is professionalized: dozens of websites sell GitHub stars between 3 cents (disposable accounts) and 90 cents (premium aged accounts) per star. Fiverr gigs offer promotion starting at $5, WeChat groups process 20+ repos and generate $3.4–4.4 million annually.

  The problem is directly linked to VC funding: a Redpoint Ventures partner documented that median 2,850 stars are required at seed rounds and 4,980 at Series A. The ROI calculation is brutal — for $85–285, startups can buy the seed median, at $1,000–4,500 the Series A range, while these rounds bring in $1–10 million (ROI: 3,500x to 117,000x). GitHub trends were manipulated as a result: 78 repos with fake campaigns appeared on GitHub Trending, despite many developers taking years to land there.

  The analysis shows characteristic patterns of fake stargazers: ghost accounts (zero repos, zero followers), low fork-to-star ratios (vs. ~0.2 for legitimate projects like Flask), and watch-to-star ratios near zero. Examples like Union Labs (52% suspected fake stars, still ranked #1 in Runa Capital’s ROSS Index) and Free Domain (81% of stargazers without followers) illustrate the scale.

  The moderator criticizes GitHub harshly for lacking moderation tools — compared to Twitch, fundamental features are missing like automod, spam limits, or word filters. He warns: Goodhart’s Law applies (when a metric becomes a goal, it ceases to be a good metric). VCs rely on startups manipulating benchmarks instead of growing organically.

  Regulatory action is underway: FTC regulation (starting October 2024) explicitly bans buying/selling fake social influence indicators — penalties up to $53,000 per case. A Headspin CEO was charged with wire fraud (maximum 20 years) and securities fraud for faking metrics to scam investors out of $80 million. So far no one has been specifically charged for fake GitHub stars, but the SEC has a path forward.

  Better metrics instead of star counts: researchers recommend weighted popularity based on network centrality, monthly contributor activity, fork-to-star ratios (healthy: 100–200 per 1,000 stars, suspicious: <50), or package downloads. The moderator also mentions the forking ratio — T3 App has ~1,600 forks to 9,000 stars (20%), Browser Use and Lovable received heavy funding because demand and product were strong, not because of stars. GitHub removed 90% of flagged repos but only 57% of accounts — the labor force of the fake-star ecosystem remains intact.

  Side note on YouTube fraud: the moderator exposes tech YouTubers with millions of fake views (engagement rates 0.025% instead of 3–4%), and confirms he’s rejected sponsorships (including Higsfield for a seven-figure offer) because the products were sketchy — while other creators take these deals and undermine brand trust.

  Format & Tools: Opinion/reflection + deep-dive, based on Dagster/Awesome Agents investigations into GitHub star manipulation; no specific AI tools/models discussed.

• I don’t really like GPT-5.5…
  24.4.2026, 03:35:49

  The creator is overall disappointed with GPT-5.5 despite acknowledging its technical capabilities. His main criticisms: the model comes across as “lazy” and doesn’t fully honor the intent of requests – it delivers the minimum to complete a task without really pushing forward. A major issue is context management: once erroneous information lands in the context, you can’t remove it through prompting – you have to start a new thread. This means long conversation threads barely work anymore; the creator had to end threads significantly more often than with other models. Additionally, he criticizes the price increase ($5 per million input tokens, $30 per million output tokens – double that of GPT-5.4) and missing API support at launch.

  On the positive side, he praises token efficiency (GPT-5.5 uses about half the tokens of its predecessors at comparable quality), code generation (the best he’s ever seen), and especially the Pro model, which is impressive at solving complex puzzles (solved unsolved DEFCON challenges from 5–10 years ago). Frontend demos, however, showed little innovation, many containing unnecessary UI elements (“cards”) in his view.

  Concrete recommendations for usage: write significantly more detailed prompts, do more research before asking, create new threads regularly, give strict instructions, and provide resources directly to avoid wrong paths.

  The creator considers GPT-5.5 a fundamentally different model that breaks old workflows – it deserves a new name (like GPT-6). He expects upcoming models to better solve current edge cases based on this foundation.

  Opinion/reflection; discusses OpenAI (GPT-5.5) and competing models (Claude Opus, Google Gemini).

• Open source is dead now?
  22.4.2026, 13:54:08

  Is Open Source Dead Now? – A Critical Analysis of Cal.com’s Strategy Shift

  The creator argues that Cal.com’s decision to shift from open-source to closed-source architecture is a concerning signal for the future of open-source software – not because the decision is incomprehensible, but because it reveals the fundamental shift in the security landscape caused by AI.

  The central thesis: AI models have democratized the exploit-finding process. Previously, you needed both deep security knowledge and domain knowledge of a specific codebase – a rare combination. Today, it’s enough to set a model loose on the source code with minimal security knowledge. The model provides the domain expertise, the human just needs a bit of security understanding. Cal.com was previously not an attractive attack target because security experts didn’t understand TypeScript fullstack projects. With Claude Opus/Mythos, that changes radically – even OpenBSD was analyzed by Anthropic’s Mythos and a 27-year-old vulnerability was found.

  The core problem: Security becomes a “proof-of-work” question: who spends more tokens? Cal thought going closed-source would increase the requirements for reverse engineering – true, but only temporarily. Models improve at decompilation. The real barrier doesn’t disappear, it just drops for a moment. Scraping API endpoints provides additional clues. And agents have no time constraints – they can search night after night.

  Counter-arguments and differentiations:

  • Tanner: if open-source is your security problem, you don’t have a real security strategy – you just lose good-faith developers who hardened code.
  • OpenAI launched GPT-5.4 Cyber (whitelisted only) to help developers preemptively harden codebases.
  • Drew and others argue: open-source becomes more valuable, not less – because multiple organizations add their token budgets for securing it, while an attacker only has their budget.
  • The ASI study shows: Mythos needed $12,500 per attempt (100M tokens), 3 out of 10 attempts fully penetrated a 32-step corporate network.

  Future scenario: The creator expects a 3-phase development cycle: development (fast, human-led), review (quality via models), hardening (token budget limited). This means: open-source projects will be bombarded with CVE floods – and maintainers (like FFmpeg) partially refuse to engage in “CVE slop,” even though that’s reckless. Hackers will ignore these. The new reality: security is expensive but necessary, and open-source projects need financial support to keep up with this game.

  Conclusion: Cal.com bought time, not genuine security. The bigger concern: fear-mongering (anxiety about AI and open source) could push many projects toward closed-source – which would be worse long-term because scattered proprietary implementations are more expensive to secure than communally-protected open-source.

  —

  Video format: Opinion/reflection with news integration (Claude, OpenAI GPT-5.4 Cyber, Anthropic Mythos explicitly discussed); deep technical analysis without beginner focus.

• Did Anthropic just kill Figma?
  21.4.2026, 10:19:39

  Anthropic launched Claude Design – a new product to help designers and developers create user interfaces. The video creator tests the tool for the first time to see if it’s truly a competitor to Figma.

  The tool in action:

  Claude Design enables creating wireframes, importing a codebase, and defining design systems. You can comment, leave feedback, and request multiple design variations. It also offers the ability to export designs directly to Claude Code – with automatically packaged specs in a dev-ready folder. The creator redesigns a marketing website for a project called T3 Code, giving the model context about features, dark mode requirements, and existing design language.

  Strengths and problems:

  The tool shows real strengths: it can process high-resolution screenshots, suggests multiple design variations (if asked), and the comment workflow for collecting feedback is practical. The creator praises the polish compared to other Anthropic launches. However, bugs do occur – designs sometimes disappear, word wrapping looks poor in places, and logo generation shows hallucinations (e.g., a fake Apple SVG). Usage limits are aggressive: at 20x subscription, a design session quickly consumes over 50% of the weekly budget.

  Creator’s conclusion:

  Despite bugs and costs, the creator is pleasantly surprised. It’s the best software product Anthropic has released in his view. With some polish, it could become a serious competitor to Figma – especially in rapid marketing prototyping and presentation design. He sees the potential and plans to keep using the tool.

  Claude Design was the central topic; demo of a new Anthropic product with critical but fundamentally positive assessment.

• Did Claude really get dumber again?
  20.4.2026, 14:50:02

  Summary: Did Claude really get dumber again?

  The creator systematically investigates reports of Claude regressions and identifies multiple causes beyond simple model degradation. He argues that perceived “dumbness” stems from several factors:

  Expectation shift: as models improve over time, user expectations rise. Tasks that once impressed become standard; failure there feels more drastic even though the baseline has increased.

  Cloud Code harness problems: the creator claims Cloud Code itself is poorly engineered. A benchmark shows Opus in Cloud Code at 58% performance versus 75–82% in other harnesses like Forge Code. Examples like unnecessary errors in file operations show how the harness pollutes context and causes redundant API requests.

  Tokenizer changes: with Opus 4.7, the tokenizer was fundamentally altered — the same input now generates 1.35–1.47x more tokens, especially with code and documentation. This means more “noise” in context, distracting models more.

  API & routing problems: aggressive filtering on the API side leads to task refusals (e.g., the creator was blocked on a simple cryptographic puzzle question). Multi-GPU routing (AWS Trainium, Google TPUs, Nvidia) across multiple tool calls could cause inconsistent quality.

  1-million-token context default: Anthropic itself confirmed in a September postmortem that the 1M-token variant of the model caused quality loss with misrouting. Since March, 1M context is standard in Cloud Code — the creator suspects this variant intentionally runs faster on AWS/Google chips (not Nvidia) and that Anthropic deliberately shifts traffic there.

  Redacted thinking & missing history: since March, 100% of “thinking” data is redacted to prevent distillation. The problem: the API no longer receives this thinking data in history messages — it would need to be fetched via database lookup. The creator suspects Anthropic implemented this incorrectly, so models have less context.

  AMD report as evidence: a detailed report from AMD’s AI director quantifies regression patterns across 17,000 thinking blocks: thinking depth dropped 73%, stop violations from never to 10x/day, read-to-edit ratio fell from 6.6:1 to 2:1, API requests grew 80x with simultaneously worse output quality.

  The creator concludes: most regressions don’t stem from worse models but from technical incompetence in Anthropic’s engineering (poor harness, API changes, faulty thinking-redaction implementation). By comparison, OpenAI models show no similar chronic regression patterns.

  Explicit tools/providers: Claude/Anthropic (Opus 4.7, 4.6, Sonnet), OpenAI (Codeex), Google (TPUs), AWS (Trainium), Anthropic Cloud Code, Cursor.

  Format: Deep-dive + investigative report with empirical data.

Tim Carambat (3 new videos)

• DeepSeek V4 Is Here And It Has 1 Million Token Context??
  24.4.2026, 11:00:51

  The video covers the release of DeepSeek V4, a new local language model featuring two variants: DeepSeek V4 Flash (284 billion parameters, 13 billion active parameters) and DeepSeek V4 Pro (1.6 trillion parameters, 49 billion active parameters). Both models support a context window size of one million tokens, enabled by a new technique that reduces KV-cache memory requirements by up to 90%. This technique, based on hybrid attention architecture, could be relevant for smaller models in the future. The models demonstrate improved benchmark results compared to earlier DeepSeek versions, particularly at maximum reasoning level. However, they only perform well in some benchmarks compared to leading cloud models like GPT-5.5 and Opus 4.6. The technical innovations, especially the more efficient use of the context window, are highlighted as significant. The video is aimed at intermediate and advanced users interested in local language models and technical innovations.

  AI Tools/Models/Providers: DeepSeek V4 (Flash and Pro), TurboQuant, Opus, GPT-5.5, open-source models.

• 27B Beats 397B?! The New Qwen 3.6 Is All About Efficiency
  23.4.2026, 18:15:01

  The video covers the latest version of the Qwen model, Qwen 3.627B, placing it in the context of previous models in the Qwen series. The author, Timothy Carbat, founder of Anything LLM, focuses on local models and their advantages regarding privacy and cost savings compared to cloud models.

  Qwen 3.627B is described as promising, particularly due to its multimodality and its capabilities in the Software Engineering Benchmark (SWE bench). The author compares the model’s performance with earlier versions of Qwen as well as with cloud models like Claude 4.5 Opus and Gemini. It is noted that Qwen 3.627B performs well in some benchmarks, but some nuances and potential biases in benchmark results are also discussed.

  The author emphasizes the importance of transparency when reporting benchmark results and warns against uncritically adopting results. He speculates that Qwen 3.6 Plus, an API-based version, could potentially be an earlier checkpoint of Qwen 3.627B due to the similarity in benchmark results.

  In summary, Qwen 3.627B is presented as a model worth considering due to its density of intelligence and multimodality, particularly for users who prefer local models. The author plans to test the model himself and share his experiences with the community.

  The video explicitly focuses on the Qwen model (open-source) and is intended for intermediate to advanced users.

• I Just Tried The Brand New Ternary Model And It’s Great!
  21.4.2026, 19:38:40

  The video covers the development and significance of one-bit and ternary models in the field of local AI. The speaker, Timothy Carbat, founder of Anything LLM, explains that one-bit models reduce the complexity of conventional LLMs (Large Language Models) by constraining weights to a 1 or 0. This enables a significant reduction in memory requirements, allowing large models like 8B models to run on devices such as smartphones or low-end laptops. However, there is a slight loss of accuracy compared to the original 16-bit models.

  Prism ML, a startup with researchers, has developed a practical one-bit model called BitNet that is actually usable and offers a good balance between accuracy and efficiency. Recently, Prism ML introduced a ternary model that adds an additional “0” to the one-bit system, further improving accuracy while keeping memory requirements significantly lower than conventional models. These ternary models offer nearly the same accuracy as 16-bit models but with a fraction of the resources.

  The speaker demonstrates how to install and run a ternary model on a MacBook M4 Max with 48 GB RAM by explaining the necessary steps and commands. He also shows how to integrate the model into Anything LLM to perform various AI-driven tasks. Finally, he discusses the future of one-bit and ternary models and speculates whether larger models than 8B can be developed with the same efficiency.

  The video explicitly focuses on one-bit and ternary models, particularly those developed by Prism ML. It is intended for intermediate and advanced users interested in the technical implementation and benefits of local AI models.

Unsupervised Learning

No new videos in this period.

WorldofAI (7 new videos)

• Deepseek v4: Best Opensource Model Ever? (Fully Tested)
  24.4.2026, 07:31:43

  Deepseek v4: Best Opensource Model Ever? (Fully Tested)

  DeepSeek has released two new models as V4 Preview: the Pro model with 1.6 trillion parameters (49 billion active) and the Flash model with 284 billion parameters (13 billion active), both featuring 1 million context length and MIT license. DeepSeek markets these as the best open-source models, claiming they beat leading closed models like Claude Opus 4.5 and 4.6 – however, the tester criticizes this as “benchmark-maxing”.

  In practical tests, DeepSeek V4 Pro performs significantly worse than expected: In UI generation tasks (macOS clone, Slack clone, SaaS landing pages), the output appears uncreative and fundamentally lacking. With 3D models (PS5 controller) and interactive projects (Instagram feed clone, Minecraft clone), the picture is similar – code is messy, lacking polish, and low complexity. The tester notes that competing models like Qwen, MiniMax M2.7, and especially Claude Opus deliver significantly better results on the same tasks. In the Code Arena ranking, DeepSeek V4 ranks 3rd, behind GLM and Qwen – consistent with practical observations.

  Price is the main selling point: Pro costs 1.4 cents per 1 million input tokens and 3.48 cents per output token; Flash is even cheaper at 0.03 cents input and 0.28 cents output. The models are available on HuggingFace and free to use via the DeepSeek website. The tester’s conclusion: Despite low costs and efficiency on paper, V4 Preview significantly lacks real use-case quality – cheap doesn’t mean better.

  Deep-dive into DeepSeek V4 Pro and V4 Flash; demo with practical comparisons to competing open-source and proprietary models (Claude, Qwen, MiniMax, GLM).

• OpenAI GPT-5.5: BEST AI Model Ever! Beats Opus 4.7 & Gemini 3.1! Powerful & Fast! (Fully Tested)
  23.4.2026, 22:40:49

  Summary:

  OpenAI has released GPT-5.5, a model focused on multi-step, labor-intensive tasks rather than just question-answering. It uses approximately one-quarter the tokens of GPT-5.4 and one-third of Opus 4.7, achieving state-of-the-art results on coding and reasoning benchmarks like TerminalBench (82.7%) and SWE-Bench Verified (58.6%), and despite sometimes lower raw scores, operates faster and more cost-efficiently in real workflows. The model excels at code refactoring, debugging, browser control, knowledge work, and agentic workflows; pricing is $5 per million input tokens and $30 per million output tokens – roughly 20% more expensive than Opus 4.7, but more efficient overall.

  Practical tests show particular strengths in frontend development and game clones (Minecraft, CS:GO, Pokémon): the model generates complex UI components, SVG graphics, and 3D scenes in great detail, especially with detailed prompts and high reasoning levels. Weaknesses lie in highly specialized 3D tasks like 360-degree product viewers. The model is accessible via ChatGPT (with “Thinking 5.5” mode), the API, or open agents like Kilo CLI and Codeex; GPT-5.5 works especially well combined with these harnesses.

  Conclusion: Demo and benchmark comparison (OpenAI GPT-5.5, Opus 4.7, Gemini 3.1 mentioned; Kilo CLI, Codeex as harnesses).

• Kimi K2.6: BEST Opensource AI Model That Beats Opus 4.6 and Gemini 3.1 Pro (Fully Tested)
  23.4.2026, 06:18:00

  Summary: Kimi K2.6 – Open-Source Coding Model

  The Moonshot AI team has released Kimi K2.6, a powerful open-source coding model that, according to benchmarks, competes with or even surpasses proprietary models like Opus 4.6, Gemini 3.1 Pro, and GPT 5.4 High. The model stands out for extremely cost-efficient pricing (94–95% cheaper than Opus 4.6 at comparable performance) and impressive long-duration capabilities: it can work autonomously for 12+ hours, manage over 4,000 tool calls in a single session, and coordinate up to 300 parallel agents.

  Core strengths emerge in practical tests: The model generates aesthetically sophisticated, interactive websites with genuine design understanding (typography, dynamic motion, video integration), creates functional 3D viewers and SVG graphics, simulates complex systems like a working Linux system with terminal and text editor, and can even generate a playable Minecraft clone. In agent swarm mode, it delegates complex tasks like creating a 12,000-word AI market analysis report to specialized agents working in parallel – tasks that would otherwise take hours are completed in minutes.

  Access is through kimmy.com (chatbot interface), the API, Kimi Code (various usage plans), OpenRouter, or the open-source alternative Kilo Code. The model offers four specialized modes: Instant (quick responses), Thinking (deep research), Agent (specialized tools like websites/slides), and Agent Swarms (long-term tasks). With 256k context window and 95 cents per million input tokens / 4 dollars per million output tokens (16 cents on cache hits), it ranks among the most cost-efficient high-performance models.

  Deep-dive/demo of the open-source Kimi K2.6 model from Moonshot AI; no specific format but continuous product demonstration with benchmarks and live testing.

• Bart Mode + Claude Code: NEW Spec Toolkit Ends Vibe Coding! 100x Better Than Vibe Coding (Tutorial)
  22.4.2026, 05:59:41

  Summary: Bart Mode + Claude Code – NEW Spec Toolkit Ends Vibe Coding

  The video presents the new “Bart Mode” feature from Tracer, which takes spec-driven development to the next level – from manually babysitting AI agents to fully automated, intelligent workflow orchestration. Instead of executing individual specifications sequentially, Bart Mode breaks down a large project (Epic) into multiple tasks, executes them in parallel, and verifies results against specifications after each batch phase. If adjustments are needed, the process adapts autonomously – user involvement is only needed when truly necessary.

  The tutorial shows concretely: After installing Tracer in your IDE (Cursor, VS Code, or WinServe), you start in Epic Mode and define your project (e.g., a dashboard with authentication and API integration). Tracer then interactively works out technical details (tech stack, data model, API endpoints, UI flows) and automatically generates a detailed implementation plan with individual tickets. This can be refined further, then the /execute command launches Bart Mode orchestration – it works autonomously with the selected coding agent (in demo: Claude Code), executes the full workflow, and continuously verifies against specs. At the end, Tracer also includes a review phase to identify and fix issues.

  The result in the demo: A fully functional agent manager dashboard with login system, agent creation, and local deployment option – all without manual intervention between work steps. The feature is available for free, even with free agents like Gemini.

  Tracer with Claude Code as coding agent; tutorial with practical demo.

• Kimi K2.6, GPT 5.5, Deepseek V4, Codex Superapp, Gemini 3.5, Grok 5 = AGI, & More! Huge AI NEWS!
  21.4.2026, 07:17:15

  Summary: Massive AI Developments This Week

  The past days brought a flood of new AI models and features. Kimmy K 2.6 from Moonshot AI is an open-source coding model competing with proprietary models like Opus 4.6 – at 94% lower input and 95% lower output costs. It supports 12+ hour sessions, over 4,000 tool calls, and up to 300 parallel agents; locally it runs on M3 Ultra devices via MLX.

  GPT 5.5 (codename “Spud”) is expected to release this week or by Thursday from OpenAI. Early leaks show superior performance versus Opus 4.7 and Gemini 3.1 Pro, especially in coding, SVG generation, and 3D workflows; the model is token-efficient and more creative in output design.

  DeepSeek V4 is also expected this week – a 1.6 trillion-parameter model with sparse MQA kernels. According to unverified leaks, it achieves MMLU scores of ~99.4% and Swaybench of 83.7%.

  Codex from OpenAI is evolving into a super-app: It can now control applications on Mac (Computer Use), integrate external tools, generate SVG/images with GPT Image 1.5, and continue tasks across sessions. The new *Chronicle* feature (preview on Mac for Pro users) creates memory from daily work patterns, making it more context-aware.

  Grok 4.3 Beta from xAI (~500 billion parameters, training data through December 2025) shows strong capabilities in frontend and coding with multi-modal support and better document generation. Elon Musk announced: Grok 4.4 (1 billion parameters, May) → Grok 4.5 (1.5 trillion, end of May) → Grok 5 as AGI.

  Qwen 3.6 Max from Alibaba (preview) brings improved agentic coding capabilities and better instruction following.

  At Google, new Gemini checkpoints are in development – possibly Gemini 3.2 or 3.5 Pro before Google IO (in ~28 days). Internally, a Copilot-like “Agent” for Gemini Enterprise is being tested to automate Gmail, Spreadsheets, and Google Workspace apps. Google AI Subscription (Pro/Ultra) now offers enhanced coding limits directly in AI Studio.

  Bonus: A robot completed a marathon, partially faster than humans – engineered like an F1 pit stop with AI-controlled motion and dry ice cooling.

  Featured AI providers: OpenAI (GPT 5.5, Codex), Moonshot AI (Kimmy K 2.6), DeepSeek (V4), xAI (Grok), Alibaba (Qwen), Google (Gemini); format: news update.

• OpenAI GPT-5.5 Leaked: Super Powerful AI Model! Beats Opus 4.7, Gemini 3.1! Cheap & Fast! (Tested)
  20.4.2026, 06:23:32

  Summary: OpenAI GPT-5.5 Leaked

  OpenAI is preparing the release of GPT-5.5 model under the internal codename “Spud”. The model is currently being A/B-tested in ChatGPT; some users have access to the “Crest Pro Alpha” checkpoint, which could represent the GPT-5.5 Pro model with Extended Thinking. The author describes the new model as “insanely good” – faster, better outputs, stronger reasoning, and better problem-solving than its predecessor.

  The model is expected to come as both base and pro variants and is characterized as a “half-cooked version of GPT6”. The main features are higher token efficiency, more tokens per second (faster responses), and lower costs – a combination that represents a major leap according to the author. The author emphasizes that in some ways it appears better than Anthropic Opus 4.7.

  In the demos, strength is particularly evident in frontend generation and code-based visualization: a Windows 11 browser OS with SVG icons, a functional Minecraft clone with blocks, textures, and terrain generation, photorealistic 3D scenes in Three.js (Monica’s apartment from Friends, complex solar system with moons and asteroids), low-poly 3D generations, flight simulators with various environments. SVG generations (Xbox controller, cat, ASCII art) are also described as accurate and compositionally thoughtful. The author emphasizes that the model “thinks” rather than just generates – scenes are designed and ideas expanded independently rather than the user needing to micromanage everything.

  Release is expected on a typical OpenAI release day (Tuesday or Thursday).

  OpenAI (GPT-5.5/Spud, Opus 4.7 mentioned for comparison), format: demo with leak news element, average complexity.

• Full AI Agent Tutorial for Beginners 2026 – How to Build AI Agents in Minutes
  19.4.2026, 04:45:06

  Summary: Full AI Agent Tutorial for Beginners 2026

  The video introduces Toolhouse, a no-code platform for building AI agents without technical prerequisites. The core advantage: agents can be configured in minutes using natural language or even voice commands.

  Key Features and Demonstrations:

  The video shows three practical ways to build agents: (1) Dashboard with Voice Control – through voice, users can describe an agent (e.g., “Create me a deep research agent for large language models, daily at 9 AM”). Toolhouse then configures autonomous pipelines with all integrated tools. (2) Integration of Tools and Knowledge – agents gain capabilities like email sending via Gmail integrations or RAG (Retrieval-Augmented Generation) through PDF document uploads. The agent can then automatically leverage these integration functions. (3) CLI for Developers – via command line, agents can be created with th new, private files added as knowledge, and deployed with th deploy.

  Practical Use Cases:

  • An agent researches LLMs daily and automatically sends a summary via email
  • Agents can be integrated via API or embedded chat interfaces (e.g., via Lovable) into external apps
  • A RAG agent processes private PDF documents and answers questions about them

  The video emphasizes that backend infrastructure complexity is completely abstracted, with focus on automation and practical use.

  Featured Platforms: Toolhouse (central), Verbra (voice control), Lovable (chat interface integration), MCP server integration for external coding agents.

  Format: Tutorial with continuous live demo, designed for beginners.

Zubair Trabzada | AI Workshop (1 new video)

• Why You Must Build a Personal Brand
  22.4.2026, 14:00:29

  Summary: Why You Must Build a Personal Brand

  The creator argues that building a personal brand now – not eventually – is essential because artificial intelligence is massively disrupting the job market. He points out that AI influencers like Lil Miquela have over 2.4 million Instagram followers and earn over 100,000 dollars monthly, with deals from brands like Prada and Calvin Klein – and this person doesn’t even exist. The virtual influencer industry is projected to be worth over 8 billion dollars by year-end. Corporations prefer AI avatars because they’re cheaper, controllable, and available 24/7, without scandals or salary negotiations.

  The problem extends far beyond influencers: 37 percent of executives plan to replace employees with AI; Goldman Sachs forecasts the loss of 25 million full-time jobs in 2026 and 270 million by 2030. Jobs are most at risk when people are interchangeable – meaning if no one knows who you are, you’re replaceable.

  A personal brand doesn’t mean becoming a YouTuber or TikTok star. It’s simply about people knowing who you are, what you do, and why it’s worth listening to you. It can be as simple as: sharing helpful insights on LinkedIn weekly, consistently answering questions in online communities, or regularly writing about your own learning journey.

  The core argument: While AI handles writing, coding, design, and analysis, there’s one thing it cannot replace – human connection. Your story, your perspective, your authenticity, the way you explain things, the mistakes and experiences you share – that’s your unique value. As AI-generated content floods the internet, genuine human voices become increasingly valuable.

  Three key points: (1) You don’t need permission or perfect conditions – start now, be consistently yourself somewhere. (2) The window is closing; in two years, authentic, trustworthy human voices in their field will dominate, while those who wait become invisible. (3) Your personal brand is the only thing that truly belongs to you – platforms, algorithms, and tools change, but your name, your reputation, and people’s trust in you follow you everywhere.

  Opinion/reflection, no specific AI tool provider mentioned.

---

Automatically generated from the latest YouTube videos of the curated channel selection. For feedback, suggestions, or to unsubscribe: simply reply to this email.