Blog

Claude vs. Codex: Anthropic’s Compute Crisis Meets OpenAI’s Developer Offensive

Tuesday, June 16, 2026

Hello, this weekly digest processes the most important new videos from around 40 curated AI and coding YouTube channels — with substance, no superficial top-5. One complete summary per video, plus a weekly overview of dominant themes. Read at your leisure — or copy a summary into the LLM of your choice and dive deeper. Click the link under each summary to watch the original video.

Model Releases & Benchmarks

GPT 5.5 from OpenAI dominated the week’s benchmark discussion: AI with Arnie, Tech With Tim, and Melvynx attested to the model’s clear improvements over GPT 5.4, particularly in coding, spreadsheets, and browser automation. Through the Codex interface, GPT 5.5 can independently test the browser, execute clicks, and take over the entire desktop via computer use. Google meanwhile quietly tested an upgraded Gemini 3 Flash in the arena, which according to WorldofAI achieved quality close to Gemini 3.1 Pro — with Google I/O (May 19–20) as the likely announcement date for Gemini 3.5 Pro. China’s Baidu delivered another counterpoint with Ernie 5.1: better benchmark results than DeepSeek V4 at lower costs. The SubQ model from Sub Quadratic with claimed 12-million-token context window and 52x efficiency garnered attention but remained skeptically reviewed due to missing technical reports and unclear benchmarks (Tim Carambat).

Local & Open-Source AI

LM Studio experienced a significant functionality upgrade according to a tutorial by Bart Slodyczka: the app now integrates MCP tools via Node.js, supports PDF analysis, web search, and image processing, and can serve as a local model server for agent frameworks like Claude Co-work, Claude Code, OpenClaw, or Hermes. Leon van Zyl demonstrated LocalForge (introduced in the video as “Honeyfree”) as a free local coding agent that with models like Qwen 3.6 or JML4 via LM Studio/Ollama builds a Kanban board for autonomous feature implementation — recommendation: set context window to at least 64,000 tokens. DeepSeek V4 established itself in several workflows as a cost-effective entry point for basic coding tasks, while more expensive models are reserved for polishing (WorldofAI demo with Claude Code + DeepSeek V4). Google released Gemma 4 under Apache 2.0 license specifically for agentic workflows and on-device reasoning.

Claude Code & Anthropic Tooling

The week brought a flood of practical Claude Code content. Julian Ivanov compiled 20 tricks for Claude Code — from plan mode and slashinit via ultrathink keyword and worktrees to hooks for background notifications. In parallel, he presented the desktop app update, which now enables full Claude Code usage including auto mode, file preview, and multiple parallel sessions. Nate Herk introduced six production-relevant skills: the official Skill Creator (meta-skill for automatic skill generation), Superpowers (senior developer workflow with isolated environments), GSD (sub-agents with clean context against context drift), /review and /ultra review (multi-reviewer agent fleet in sandbox, from Opus 4.7), Context Mode (SQLite snapshot instead of garbage data), and ClaudeMem (vector search memory across sessions). Additionally, he showed how to build a voice sales agent with Claude Code and 11 Labs in 15 minutes, and how Higgsfield via MCP connector serves as a creative agency for image generation and video production. TheAIGRID provided tutorials on the official Claude integration in Microsoft Word and PowerPoint (each as an add-in, from Pro plan, with model choice between Sonnet 4.6 and Opus 4.6).

Coding Agents (non-Claude)

Codex from OpenAI was positioned in multiple videos as a full-fledged Claude Code competitor: Nate Herk built a complete YouTube analytics dashboard including GitHub deployment and weekly automation; Everlast AI listed 34 tips for the desktop app, including Chronicle Research Preview and sub-agents for context management. Melvynx compared Codex and Claude Code directly: Codex impressed with sub-agent overview, PR management, browser preview, and message queuing, but took 29 minutes for a task that Claude needed 40+ minutes for. As open-source alternatives in focus: Leon van Zyl showed OpenCode (not to be confused with OpenClaw) as a CLI tool with provider selection between OpenAI, Anthropic, Gemini, OpenRouter, and others. The Pi Agent Harness (Alejandro AO) positions itself as a minimalist, fully extensible harness with four core functions, skill system, package management, and session management — without pre-built features like MCP or sub-agents. Tech With Tim demonstrated Mistral Vibe with the open-source model Devstral 2 for specialized parallel sub-agents (test writer, code reviewer, deploy prep).

Software Engineering & Dev Culture

Nate B. Jones’ analysis of the Mozilla Mythos experiment was the week’s most substantive software engineering piece: Claude Mythos Preview found 271 security vulnerabilities in a Firefox release cycle (vs. 22 with Opus 4.6) because it doesn’t just search for known patterns but runs through a research loop of code reading, hypothesis formation, test case generation, and explanation synthesis. His thesis: readability becomes a security property, and there’s a 4–5-month window to refactor codebases for AI security systems. Fireship covered a critical Linux kernel vulnerability (CVE-2023-31431, present since 2017, discovered by an AI agent tool) that enables local root access via a 732-byte Python script — all Linux distributions after 2017 affected. Cole Medin demonstrated live the PIV Loop (Plan, Implement, Validate) as a reproducible system for AI-coding assistance with Claude Code and Jira. Matt Pocock introduced the triage tool for GitHub issue backlogs: a state machine system with two types (bug/enhancement) and five states that prepares issues directly for autonomous agent processing.

Personal AI OS & Agent Frameworks

The Hermes agent dominated this week’s comparison discourse with OpenClaw. Alex Finn ran a three-hour “Agent Olympics” livestream in which four combinations (OpenClaw and Hermes each with ChatGPT and Opus as backend) competed against each other — OpenClaw with Opus won, but Hermes showed considerable progress with seven new features like Kanban board, slashgo for long-term missions, multi-profile agents, model catalog, and automatic curator feature (cleanup of unused skills after seven days). Hermes suffers from compaction errors (loss of work state), however, which secures OpenClaw consistency advantages. Leon van Zyl successfully built Hermes as a coding agent for simple web apps with Vercel deployment via Telegram. Nate B. Jones analyzed OpenClaw as an agentic runtime: the decisive architectural move wasn’t model diversity (dropdown) but memory as an independent context layer outside all models — the Open Brain Recipe as an open-source repo defines provenance labels for every stored memory. Mark Kashef demonstrated a self-built AI operating system based on Claude Code with “Hive Mind” visualization, 3D graph view, and Meta CLI integration for ad analysis.

AI Automation & Workflows

Nate B. Jones delivered two structuring concept videos: First, a taxonomy of agent scaffolding layers (prompt → skill → plugin → MCP/app connector → hook/script), emphasizing that drawing plugin boundaries will be a highly paid skill in 2026. Second, the anticipation gap — the core problem with current consumer agents isn’t capability but lack of proactivity; he proposes a five-tier permission ladder model. Cole Medin presented Archon, an open-source harness builder for YAML-defined AI coding workflows, heading toward a community marketplace for shared workflows; in a second livestream he demonstrated a complete video generation workflow (Archon + 11 Labs + Remotion). n8n showed in an interview with product manager Sindhuja the development of an instance-level MCP that makes n8n controllable from any platform. Nate Herk introduced Printing Press — a CLI factory with over 50 pre-built CLIs (including ESPN and Hacker News) as a more token-efficient alternative to MCP servers in agent workflows.

AI Video & Content Creation

Cole Medin built a complete AI video workflow in a livestream: Archon orchestrates the pipeline, 11 Labs generates speech, Remotion creates visual elements — demonstrated with a marketing video for Archon itself, with iterative improvements via feedback loops. Nate Herk showed the Higgsfield integration via MCP in Claude Web and Claude Code: a single prompt generates product photos, Instagram ads, and UGC videos; Claude Code additionally creates a Google Sheet database with 45+ generations for status tracking. WorldofAI presented Open Design as an open-source alternative to Claude Design (Anthropic): 31 composable capabilities, 72 design systems, compatibility with 15 coding agent CLIs, deployable locally.

AI Business, Marketing & Freelancing

Dave Ebbelaar sketched a three-tier “Data Freelancer Blueprint”: demonstrate projects (Get Going), research hourly rates and leverage warm network introductions (Getting Paid), systematically improve leads and delivery (Get Good) — with N8N and Airtable as low-code entry points. Kyle Balmer analyzed the playbook of the big labs (consulting expansion) and derived a market gap from it: execute the same strategy in the mid-market, with workshops as the cheapest entry point ($2,000–5,000 per hour). TheAIGRID presented Meli, Google’s new AI marketing tool from Google Labs, which generates social media creatives, videos, and campaigns from website input, powered by Gemini.

PKM & Knowledge Management

Fireship delivered a compact, non-AI-specific operating system explainer video — from bootloader through virtual memory, kernel, scheduler, and threads to the shutdown process — as a conceptual foundation for anyone wanting to understand what happens below the agent layer.

Prompting & AI Literacy

Kyle Balmer picked apart Marc Andreessen’s viral system prompt: superlatives like “world-class expert” provably don’t improve model performance, “never hallucinate” doesn’t work structurally, and maximum length produces padding instead of quality. As an alternative, he presented the RISEN framework (Role, Instructions, Steps, End Goal, Narrowing). In a second video, he explained AI hallucinations mechanically (LLMs as stochastic token prediction, not database query) and named practical countermeasures: enable web search (–45% errors), stronger models, NotebookLM as a personal RAG system. A third video addressed AI writing patterns and their avoidance: OpenAI banned “Goblin” after an RLHF reinforcement loop; Wikipedia maintains a 15,000-word guide with AI detection markers; automatic AI detectors don’t work reliably and systematically disadvantage non-native speakers.

AI Industry & Strategy

The restructuring of the Microsoft-OpenAI partnership was the week’s biggest strategy topic. Theo (t3.gg) analyzed how OpenAI has terminated exclusivity with Microsoft since the O1 reasoning breakthrough and now offers models via AWS Bedrock and Google Cloud — a direct attack on Anthropic’s enterprise advantage there. Kyle Balmer highlighted that OpenAI (with “The Deployment Company”) and Anthropic (with Blackstone, Goldman Sachs) entered the consulting business nearly simultaneously and deploy AI engineers directly to companies — modeled on Palantir. TheAIGRID documented Anthropic’s growing market share despite community criticism. Nate B. Jones investigated Stripe’s agent commerce infrastructure: Links Wallet for agents with programmatic payment authority, Metronome for usage tracking, Radar against token fraud — the central thesis: purchase decisions will form in the buyer’s environment, not on the seller’s website. On the GitHub landscape: Theo discussed alternatives (Forgejo/Codeberg, Pierre, Graphite after Cursor acquisition, Entire from the ex-GitHub CEO) given growing reliability issues.

AI & Society / Future of Work

Melvynx refuted the job loss doomer thesis with historical data (agriculture, electrification, spreadsheets) and current figures: 90 percent of surveyed companies report no significant employment impact in three years, software developer demand spiked sharply since early 2025. Nate B. Jones added a practical audit framework: categorize work into Theater (T), Commodity (C), On-the-Line (L), and Durable (D) — his core thesis: performance systems reward visible outputs, not real value creation, which is why roles can look good while their economic foundation erodes. Everlast AI spoke with two robotics experts: a TU Munich professor on humanoid robots (VLA models, Nvidia Omniverse, liability questions) and Goodbites founder Dr. Susemihl, whose fully autonomous robot kitchen produces 50,000 meals daily and supplies the US Army — winner of the Robotics Award 2026 at Hannover Messe. TheAIGRID discussed the alignment problem and why AI CEOs fear their own product despite — or precisely because of — their AGI work.

In Brief

Zed (code editor) introduced the ACP protocol (Agent Client Protocol), making in-house agents from Codex, Cursor, or Claude usable directly in the editor, supplemented with multi-repository support via work trees (Melvynx). · NeuralNine showed a step-by-step VPS setup for OpenClaw with Ubuntu, Node.js, and Telegram integration, plus a Python quickstart tutorial for Slack bots with GPT-4o integration. · TheAIGRID documented AI-accelerated quantum computing progress: Google research suggests under 1,200 logical qubits might suffice for certain encryption attacks; Cloudflare moves its target for quantum-safe infrastructure forward to 2029. · WorldofAI reported from the Anthropic developer conference with announcements on dreaming functionality (agents review past sessions), multi-agent orchestration, and three model development goals: infinite context window, multi-agent coordination, persistent long-term reasoning. · Perplexity launched a finance agent with licensed data (WorldofAI). · Baidu Ernie 5.1 outperforms DeepSeek V4 in benchmarks at lower cost, according to reports (WorldofAI).

AI Explained

No new videos in this period.

AI Foundations (1 new video)

• FULL Claude Cowork Tutorial For Beginners in 2026! (Zero to PRO)
  4.5.2026, 17:28:28

  The video provides a comprehensive introduction to Claude Cowork, an advanced feature of AI Claude that goes beyond basic chat functionality. It starts with an explanation of the pricing structure, which requires at least Claude’s Pro version, and highlights the unique features of Cowork that are not accessible through the web interface but rather through a desktop application.

  The tutorial showcases the Claude Cowork user interface and explains various tabs such as Projects, Scheduled Tasks, Live Artifacts, Dispatch, and Settings. A central focus is on the differences in how you interact with Claude Cowork compared to Chat mode. Cowork is designed to accomplish tasks, while Chat mode is more geared toward brainstorming and strategy development.

  A practical example demonstrates how to create a folder on your desktop and have Claude Cowork interact with it to conduct research on the year 1800. In this process, Subagents are employed to research in parallel and create Markdown files. This feature enables you to handle multiple tasks simultaneously and save results directly to your computer.

  Another important aspect is the creation of Live Artifacts, which are described as mini-applications connected to real-time data. An example shows how an interactive dashboard for the year 1800 is created, featuring various information and learning tools.

  The video also demonstrates how to connect tools like Gmail with Claude Cowork to automate tasks such as categorizing invoices. This integration allows you to read and write data from various applications, significantly boosting efficiency when handling emails and other tasks.

  Additionally, Claude Skills are introduced, described as automations that perform specific tasks consistently in the same way. One example is a daily briefing automation that summarizes emails and calendar entries, giving the user an overview of their day.

  Projects in Claude Cowork are described as containers that hold specific instructions and files to accomplish certain tasks. An example shows how a project for creating SEO-optimized blog posts based on YouTube transcripts is built.

  The video concludes with an invitation to join the AI Foundations community, where additional courses and resources for automation with Claude are offered.

  Final comment: The video explicitly focuses on Claude Cowork and is geared more toward intermediate to advanced users.

AI with Arnie (1 new video)

• Is this AI breakthrough real?
  May 7, 2026, 15:15:00

  The video tests and compares the new models from OpenAI, particularly GPT 5.5 and the image model GPT Image 2.0, as well as competition from Deep’s version 4. The testing covers various applications including website creation, bee colony simulation, a 3D motorcycle racing game, an interactive factory and production simulation, traffic simulation, creation of ComfyUI and N8N workflows, and financial data analysis. The benchmarks show that GPT 5.5 demonstrates significant improvements over previous versions in many areas, particularly in Terminal Benchmark and Vending Benchmark. The video also discusses current issues at Anthropic, specifically rate limits, performance problems, and model unreliability, as well as the pros and cons of OpenAI and Anthropic plans. It’s recommended not to rely on a single provider and to use both models to balance different strengths and weaknesses.

  The video explicitly covers OpenAI (GPT 5.5, GPT Image 2.0, Codex), Anthropic (Cloud Code), and Deep’s version 4. It’s geared more toward intermediate and advanced users, as it includes detailed tests and technical analysis.

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

• You’re Wasting 40% Of Your AI Time On Something Fixable
  9.5.2026, 15:00:09

  # Summary: The Scaffolding Layers of AI Agents

  The video explains the various components that make an AI agent functional – not the language model itself, but the “armor” around it. The speaker organizes these components by complexity and reusability:

  **Prompts** are designed for one-off tasks. They don’t scale well for repeated work and require manual input each time. **Skills** are Markdown documents that describe reusable processes (e.g., how your brand handles customer service). Skills are universal and tool-agnostic. **Plugins** are larger workflow packages that combine Skills, data connections, scripts, hooks, and assets into a single installable unit – for example, pulling Salesforce data, processing it, and checking the results.

  **MCPs and App Connectors** are the “internet plugs” of the system: they connect the agent to live data from external tools like Slack, Figma, or GitHub.

  **Hooks and Scripts** are deterministic controls that shouldn’t rely on the model’s judgment – for example, validating JSON, running tests, or formatting code. They often belong inside the plugin.

  The central idea: these aren’t competing tools, but Lego blocks. Repeated work stays a Skill. When a workflow grows, with data sources or validations, it becomes a Plugin. The speaker emphasizes this isn’t purely technical – non-technical people can build plugins by using their domain knowledge to map out and structure workflows. According to the speaker, this is a highly paid skill in 2026 because few people understand where workflow boundaries should be.

  The video also criticizes how these concepts often remain vague and only engineers understand them – so automation stays a technical privilege. Instead, everyone should know: one-time = Prompt, repeated = Skill, ready-to-travel/with tools = Plugin.

  Practical examples: weekly business reports (Plugin with spreadsheets, Slack, Docs, dashboards), editorial reviews, design workflows with Figma integration, outbound emails with CRM data, customer service plugins for returns vs. activations (separate, not all in one).

  The speaker announces a Workbook on Substack with decision trees, starter plugins, testing checklists, and trust questions for safe installations.

  **Format:** Deep-dive / opinion, explicitly addressing Codex, Claude, Claude Design, and generically the concept of agents; the video also targets CTOs and executives who should understand the mental model.

• 271 Vulnerabilities: What Mozilla’s AI Found Changes Everything
  8.5.2026, 14:00:50

  # Summary: The Reversal of Trust in Code – From Myth and the Future of Software Engineering

  The central argument of this video is a fundamental shift in software security architecture: the statement “A good human engineer wrote this” loses its power as a trust guarantee, while AI systems gain significance as quality seals.

  **The Myth Experiment:** Mozilla gained early access to Anthropic’s Claude Myth Preview and deployed it on Firefox. The result: 271 identified vulnerabilities in a single release cycle – a browser that already had fuzzing, sandboxing, memory-safety work, internal security teams, and bug-bounty programs. For comparison: the previous collaboration with Claude Opus 4.6 found only 22 security-critical bugs (14 of them high-severity). This shows not just better code review, but a new industrial process for vulnerability detection.

  **Meaning vs. Implementation:** The author separates two things mixed in code: *meaning* (what the system *should* do) and *implementation* (what it *actually* does). Security vulnerabilities often arise in the gap between intended meaning and actual behavior. Myth doesn’t just search for known harmful patterns – it actively participates in a research loop: reading code, forming hypotheses, using tools, generating test cases, reproducing bugs, refining findings, and explaining them.

  **The Trust Shift:** Code was trusted because human judgment was the only capability for producing and understanding software at the right abstraction level. If machines surpass humans at exhaustively searching code consequences, human authorship no longer serves as a trust anchor – but becomes a source of unverified risk.

  **Parallel to Earlier Shifts:** Software has undergone such shifts before (compilers instead of assembly, garbage collection instead of manual memory management, automation instead of manual deployment). Humans weren’t eliminated from computing but *shifted upward* to higher abstraction levels. Security could be the next such shift.

  **The Practical Consequence:** The human engineer’s role shifts from personal code implementation to defining meaning and intent. The process becomes: Humans describe intent → Models propose implementations → Other models attack them → Tools produce evidence → Senior engineers review the Myth review cycle and decide on ship-worthiness.

  **Impact on Evals and Pipelines:** Currently, many people write evals at 80% functional code and only 20% non-functional requirements. This should be reversed: at least 50% should address code hygiene, architecture, clear function boundaries, and security standards. Myth can take over this human security-review work – but the code needs to be readable and well-structured.

  **The Golden Refactor Window:** There may be a 4-5 month window (until around end of 2026) to refactor codebases so they can be understood by AI security systems. Bad code isn’t just annoying – it’s a security risk. It can become structurally resistant to the tools that could make it safer. Readability becomes a security property.

  **Concrete Guidance for Different Roles:**
  – **Individual Contributors:** Write better specs, understand clarity of intent, structure code so machines can understand and defend it.
  – **Team Leaders:** Start architecting agentic pipelines now – design them modularly so Myth-equivalent systems can be plugged in later.
  – **CTOs:** Begin budgeting and planning; reshape your organization’s trust model.

  **The Bigger Shift:** From a world where the codebase itself is the gold standard, to a world of Intent Bundles (intent + implementation + verification through agentic pipelines) reviewed together. The valuable engineer won’t be the one writing the cleverest prompts, but the one who can define a system that can be safely implemented: through sharp standards, verifiable boundaries, APIs that minimize authority leakage.

  **Human Responsibility Doesn’t Decrease – It Concentrates:** Not on typing every line, but on defining where meaning enters the system. Senior engineering always depended on understanding meaning, seeing hidden couplings, knowing when a product choice creates a security problem – that now becomes the core role.

  **Warning Against Overinterpretation:** Not all AI code is secure today. Don’t replace every senior engineer with a model. Claude Myth is specific. But there are signs (ChatGPT 5.5 has similar security-sniffing attributes; open-source models will follow) that Myth-like capabilities will spread by year-end.

  **The Paradox:** In a world with Myth, human-written code might be perceived as *insecure* – not because humans are fundamentally bad, but because code that hasn’t been exhaustively searched adversarially carries new risks. Generated code will be trusted not because a model produced it, but because it went through a verified process.

  **Zero-Days:** A model finding a bug doesn’t magically heal the system. The security process doesn’t end at discovery. So: search existing code now with Myth-equivalent capabilities and patch aggressively. Mozilla releases Myth selectively because the organizations receiving it control some of the internet’s most powerful systems – they want them hardened, not attacked by adversaries in 3-4 months.

  **The Deeper Trust Model:** Code isn’t trusted because it’s human-readable (that’s a side effect), but because good architecture makes it attackable by friendly machines. Narrow modules are easier to constrain, explicit boundaries easier to test, small interfaces easier to verify.

  **The Ask:** Not natural language in, app out, but: natural language in → think through → traces, proofs, type systems, tests, adversarial review as part of the agentic pipeline → verification → humans inspect and sign off on the intent bundle. That’s the standard being built toward.

  Anthropic Claude (specifically Myth Preview), ChatGPT, and future open-source models are central; opinion/deep-dive on fundamental shifts in software security, trust models, and engineering culture.

• Your AI Agent Is Locked To One Model. OpenClaw Just Killed That.
  7.5.2026, 14:00:11

  # OpenClaw in April 2026: From Demo Project to Production-Ready Agentic Runtime

  OpenClaw evolved in April 2026 from a viral open-source agent framework into a serious agentic runtime – accordingly, the system expanded with massive release velocity across task management, memory features, provider support, channel updates, and automation. The surface description “a model with access to your computer” becomes incomplete: OpenClaw is becoming the action layer for agents, not just a chatbot wrapper.

  **The New Product Foundations for Serious Work:** Task Flow now orchestrates durable multi-step workflows with its own state and revision tracking; Memory evolved from a gimmick (“the bot remembers your name”) into an operational context layer – with concepts like Memory Wiki, Active Memory, and provenance-rich recall that track whether memories were observed, confirmed, or inferred by the model. Channels now handle the heterogeneous needs of different platforms (Slack, Discord, Teams, etc.) with correct threading and permission handling as a core runtime component rather than a distribution feature.

  **The Model Wars:** Anthropic restricted Claude subscriptions for always-on third-party agents in April – the reasoning is rational (agents consume more tokens, aren’t normal chat users, Anthropic wants API pricing instead of flat-rate), but unpopular with developers who used Claude as cheap background brains. OpenAI took the opposite position: Codex became accessible to all paid ChatGPT tiers; Sam Altman explicitly announced that OpenClaw now runs natively on OpenAI infrastructure. Google launched Gemma 4 under Apache 2.0 for agentic workflows and on-device reasoning.

  **The Strategic Rethink:** The crucial insight isn’t that OpenClaw can now use different models (that’s a dropdown), but that it *should*: cheap local models for classification and triage, GPT 5.5 for complex repo work, Claude API for high-judgment tasks, others for summarization. But this requires architecture where the model isn’t the product surface – it’s a swappable reasoning engine within a stable workflow loop. A durable workflow pattern shows this exemplarily: code review that triages GitHub issues, compares against historical fixes, knows risk files, and stores test lessons – the useful knowledge isn’t in the prompt, it’s in codebase history, reviews, deployments, and accumulated lessons. If that memory only lives in a chat transcript or a provider product, the workflow pattern breaks.

  **Memory as an Independent Context Layer:** Core to the solution: Memory must live outside any single model. So an “Open Brain Recipe for OpenClaw” was published as open-source – it defines how the agent retrieves context before critical work (project conventions, people, decisions, prior failures), how it writes back afterward (outputs, lessons, source channel, model use, task ID, confidence), and what provenance labels accompany each memory (observed, inferred, user-confirmed, imported). Concrete recipes: code-review memory for reusable PR lessons, task flow worklog for long-running agent attempts, memory-provenance recipe for clear source labels. This isolation makes the workflow resilient to model switching, pricing changes, and better local models.

  **The Architectural Conclusion:** Build one runtime, route different brains through it, build workflows vertically (Sales Ops, Research, Meeting Follow-Up, Compliance, Finance) – the constraint isn’t model access, it’s ownership of memory, tools, permissions, operating rhythm. OpenClaw users shouldn’t bet on vendor loyalty, but on runtime design where the model stays swappable and memory belongs to the user.

  —

  **Explicitly covered providers/models:** OpenAI (GPT 5.5, Codex), Anthropic (Claude, Claude API), Google (Gemma 4), Open Router, DeepSeek, Ollama, LM Studio; focused on OpenClaw as framework. **Format:** Deep-dive with strongly articulated architecture thesis; high difficulty level (targets builders, not beginners).

• The Work Primitive: What Every AI Product Leader Gets Wrong
  6.5.2026, 14:01:00

  # Summary: Work Primitives and the Strategic Layer Beneath Agent-Driven Workflows

  The author argues that the central strategic battlefront for AI agents isn’t the superficial ability to operate a computer, but control over **semantic work primitives** – meaningful units of work that agents need to understand and execute.

  **The Three Layers:** Access, Meaning, and Authority. Computer Use gives agents access; MCPs and APIs give them access to richer interfaces. But real power lies at the *semantic level*: does the system understand what an action means? Moving a calendar event isn’t just a click – it notifies five people, possibly breaks customer commitments, creates conflicts. A refund isn’t just a button, it’s an action with money flow, fraud risks, tax implications.

  **Why Coding Agents Worked First:** Not because code is text, but because the dev environment already has rich semantic meaning – tests, linters, dependencies, Git history. The agent gets semantic feedback directly. With calendar or sales workflows, this structure often doesn’t exist; meaning is hidden in policy, relationships, and unwritten history.

  **The Interface Hierarchy:** Agents should use the richest semantic interface (API > Connector > Browser > Desktop Control), not the reverse. This isn’t just engineering preference – it’s necessary for reliable high-stakes work.

  **The Platform Battle:** Two approaches compete: (1) work backward from semantic work meaning to agents (Perplexity strategy), (2) move forward from models and code to work (Claude, Codex). Hyperscalers can pursue both; non-hyperscalers must choose their lane. Salesforce deliberately exposes MCPs and APIs; SAP blocks agents – the author sees Salesforce as correctly positioned.

  **Central Tension for All Software Companies:** Too little semantics = agents click clumsily through UI; too much = product becomes backend infrastructure for foreign agentic interfaces. The question is: who defines work meaning?

  **The Real Goal:** Not software that makes every button clickable, but software that describes the *action behind the button* – authorized, verifiable, reversible, and composable from the start – legible to agents from day one, not just technically legible.

  **Explicitly Named Tools and Providers:** Codex (with Computer Use and auto-review feature), Claude (prefers MCPs), Perplexity (Personal Computer, Finance workflows, Comet), Salesforce (360, MCPs/APIs), SAP (agent-blocking), GitHub, Stripe, Shopify.

  **Format:** Deep-dive / opinion (strategic framework for product leaders).

• Consumer AI Has a Problem Nobody’s Naming.
  5.5.2026, 14:00:58

  # Summary

  The speaker criticizes that today’s AI-agent landscape has a fundamental problem: while agents are technically capable, they force users into a new management layer instead of offering true assistance. The central problem is the **Anticipation Gap** – agents are reactive (you have to summon them), not proactive. The ideal future agent would spot problems before they become work: it notices your flight is delayed and asks if you want to rebook. School sends an email with a signature form due Friday – the agent alerts you. A tense work thread needs a careful response – the agent drafts a calm counterproposal.

  The biggest obstacle isn’t technical capability, but agents understanding *when* they’re allowed to appear, *what* really matters, and *how* to act without being intrusive or mistakenly proactive. Current products like Poke (messaging interface), Clickie (cursor assistant), and Cluey (invisible help) try different approaches but all still miss true proactivity – their responses feel too generic or they require user intervention.

  The speaker proposes a **Permission Ladder principle:** Level 1 (Reading), Level 2 (Suggestions), Level 3 (Drafts), Level 4 (Actions with approval), Level 5 (Autonomous). For consumer agents, you should consciously choose which level to target – not “manage my whole life,” but a few specific domains with enough context and control.

  A promising path: Proactivity might first emerge at *work* (similar to how Slack did), since clearer metrics and more structured contexts exist there. Signs of near-term breakthroughs include strategic hires (like OpenAI hiring Peter Steinberger), increased load-relief in existing agents, and model release notes announcing long-running consumer intent with memory.

  The video argues that demand is enormous and technical capability exists, but the product approach must be fundamentally different: not “reactive until you call me,” but “I appear when it matters, ask permission for important decisions, otherwise vanish.”

  **Explicitly Mentioned Tools/Platforms:** OpenAI, Claude/Anthropic, Codex, GitHub, Cursor, Symphony Protocol, Stripe, Poke, Clickie, Cluey, Co-work, Chronicle (Codex), open-source/OpenClaw. **Format:** Opinion/reflection with deep-dive into product design and use cases; designed for advanced product thinkers and AI-interested audiences.

• AI’s ‘Thin Ice’ Moment: Is Your Job Already Gone?”
  4.5.2026, 14:01:31

  # Summary

  The author argues that the real AI danger for knowledge workers isn’t whole jobs disappearing, but individual tasks within a job becoming obsolete – which can undermine an entire role’s foundation, similar to how online booking changed travel agencies. He presents an audit framework that sorts the past two weeks of work activity into four categories: **T (Theater)** = organizational performance without real value, **C (Commodity)** = real but interchangeable routine, **L (On the Line)** = transitional work facing increasing automation pressure, **D (Durable)** = work whose value depends on non-fully-describable judgment.

  Most knowledge workers discover their weeks are disproportionately T and C, while D work – genuine judgment under uncertainty, holding questions rather than just answering – is much smaller than their professional identity suggests. The core problem: performance systems reward visible outputs, not real value creation, so roles can still “look good” today even though their economic foundation is eroding.

  The author recommends six concrete steps: (1) gradually eliminate theater tasks, (2) don’t redirect saved time into more commodity work – invest it in D development, (3) document durable judgment weekly, (4) systematically reduce commodity burden through project selection, (5) communicate judgment outcomes without explaining deeper mechanics (Partial Legibility), (6) switch roles if they have insufficient D potential. The key insight: the transition isn’t forced – there’s still time to reorient yourself before the organization does.

  **Format & Tools:** Opinion/reflection with practical audit tool; Claude and Codex mentioned as helpers for data processing.

• Stripe, Visa, Mastercard, Microsoft, Meta. All Building The Same Thing.
  3.5.2026, 17:00:44

  # Summary: The Shift of Power from Sellers to Buyers Through Agent-Driven Commerce

  The core of Stripe’s recent announcements isn’t that AI agents can now buy coffee, but a fundamental restructuring of the internet economy: for the first time in decades, power is shifting from seller to buyer.

  **Three Central Shifts:**

  1. **The Old Sales Funnel Was a Machine for Making Human Intent Visible.** Websites, apps, checkouts, and landing pages were controlled environments where companies could observe demand. This structure justified an entire industry (8,000+ martech companies in the 2010s) around human attention. Agents change that radically: the purchase decision forms not on the seller’s website, but in the buyer’s environment, before they ever visit the store.

  2. **Payment Authority Moves With the Task, Not at Checkout.** Stripe’s Links Wallet for agents lets an agent work with programmatic access to limited payment instructions – with a one-time virtual card or shared payment token. That’s not the same as traditional checkout. Payment method is bound to the task, can be capped by amount, currency, merchant, or approval status. In the old model, the seller extracted payment authority. In the new model, the buyer’s agent brings payment authority.

  3. **The Competitive Question Is No Longer “Are We Using AI?” but “Can Agents Call Us?”** When agents arrive (not if), the purchase journey will be simultaneously customer-driven and agent-driven. This means buyer comparison, seller legitimacy, and price clarity must exist long before the checkout moment.

  **Why All This Even Has to Work:**

  For an agent to buy something for a buyer, it needs: structured product information an agent can read (not just marketing copy), clear prices, return policies, delivery windows, payment options, and “intent hooks” – ways to translate vague human requests (“authentic coffee”) into precise commercial queries. That’s a much higher bar than “SEO for agents.” An agent can’t live with ambiguity like a human; it needs clarity to act.

  **What This Means Concretely for Businesses:**

  Websites may become less central, but commercial reality must be far more explicit. Product catalog, prices, policies, payment methods, usage limits, fulfillment constraints need to be exposed as agent-accessible surfaces – through protocols, APIs, feeds, or platforms like Stripe itself. Discovery becomes less about “winning Google rankings” and more about “being an available option in the agent’s decision process.”

  **On Instant Checkout and OpenAI:**

  The Walmart test showed that instant checkout (purchases right in chat) actually converted worse than sending back to the website. The problem is structural: people don’t want to buy individual items in chat when they already have carts, loyalty programs, delivery expectations, and existing merchant relationships. OpenAI’s corrected approach was to focus on discovery and treat checkout as adjunct, not primary experience.

  **Payment Structures in the Agent Age:**

  Stripe’s announcements around Streaming Payments (with Metronome for usage tracking and Tempo for stablecoin micropayments) reveal a much broader picture. Not every transaction is a one-time purchase. There are time-based intents (“buy later”), budget limits (“spend up to $100 finding the best supplier”), usage-based models (per query, per token), outcome-based (pay if ticket is resolved), and hybrids. These were awkward for humans to manage before – now they’re agent-native transactions. That requires different billing and settlement systems than classic checkout pages.

  **Fraud and Trust in the Agent Age:**

  Stripe’s Radar announcement is critical: in a world where agents burn tokens, fraudsters can cause massive damage. A free trial that was once harmless becomes token burning. Radar uses Stripe’s network (Link as wallet, Stripe Signals for risk data across payments, business, signup, and agent behavior) to detect fraud. That’s like a trust guarantee for a buyer-driven economy – when commerce leaves the seller funnel, trust must come from elsewhere.

  **On Brands:**

  An agent doesn’t feel nostalgia or aspiration like a human. It won’t be moved by landing-page emotions. But that doesn’t mean brands disappear – they change location. In the old web, brand worked through persuasion: you land, see design and social proof, the seller builds brand impression. In the agent web, brand becomes part of buyer memory: your preferences, prior purchases, trust history, loyalty memberships, stated dislikes become agent inputs. The agent can carry “I like 49th Parallel Coffee” as a constraint – or “I avoid this airline.” That’s hard for sellers because they can’t reset the conversation. Brand becomes not like a billboard – it becomes an entry in the buyer’s operating context, in the ledger. The brands that matter to agents are those that became reliable preferences: clear data, clear policies, consistent fulfillment, strong reputation.

  **The Larger Implication:**

  If Stripe lowers costs for transactions and trust, companies that don’t rest on real buyer relationships may disappear – companies that only exist because tired buyers land there. Agents will make those “frustrated landing” moments rarer. The internet economy will become more rational, efficient, less emotional. That doesn’t mean marketing vanishes – it just works differently. You can’t emotionally persuade an agent. You have to make your offering relevant.

  **Critical Questions for Every Business:**

  – Can an agent call your business programmatically (not scraped, but via APIs)?
  – Can the agent understand when you’re relevant and compare you against alternatives?
  – Can the agent act without human oversight?
  – Can the agent distinguish facts from marketing fluff?
  – Can the agent read your prices, terms, cancellation policies, error handling?

  The reason commerce is hard isn’t because button-pressing is hard. Commerce is hard because economic action has real consequences. Stripe knows this because Stripe has always lived where software meets money.

  **Outlook:**

  The old internet asked: “How do we get customers into our store?” The new internet asks: “How do we become usable by the buyer’s agent if the buyer never shows up?” The infrastructure of the sales funnel will collapse. Either you invest in high-quality experiences for real people (and in-person marketing), or you make your entire infrastructure agent-friendly and cleanly contractible. When the buyer wants “authentic coffee” and the agent visits your website, every hook needs to be there so the agent reads the real qualities and maps back to your Ethiopian honey-processed coffee.

  —

  **Technology & Format:** Deep-dive analysis of Stripe’s agent-commerce strategy with practical implications; no specific AI models featured, but rather broader economic shifts and Stripe’s role in them.

Alejandro AO (1 new video)

• Pi Agent – Crash Course | Minimal Coding Agent
  6.5.2026, 06:17:00

  This video introduces the Pi Agent Harness, a minimalist agent framework that can be customized to meet user needs. The focus is on enabling users to get started quickly without covering all features in detail. The key points are:

  1. **What is Pi?**
  – Pi is a minimalist agent harness that starts with just four core functions and can be customized individually through extensions and modifications.
  – Unlike other agent harnesses, Pi doesn’t offer pre-built features like MCP (Multi-Chat-Persistency), Sub-Agents support, planning mode, or integrated to-do lists. Instead, these features can be added as needed.

  2. **Installation and Basic Configuration**
  – Installation is done through a simple command in the command line.
  – After installation, users need to authenticate with an API key or subscription (e.g., from Hugging Face, Anthropic, ChatGPT, or GitHub Copilot) to use LLMs (Large Language Models).
  – Models can be selected using the `/model` command and switched between different providers.

  3. **Model Selection and Customization**
  – Users can choose between different models and save them in favorite lists.
  – The thinking level of the model (e.g., low, medium, high) can be adjusted.

  4. **Prompt Templates**
  – Custom prompts can be created and saved to simplify frequently used commands.
  – These prompts are stored in the `.pi` directory structure and can be manually edited.

  5. **Skills**
  – Skills are functions that extend the agent harness. They can be loaded from various directories, including `.agents` and `.clod`.
  – Skills can be invoked and used via the `/skill` command.

  6. **Themes and User Interface**
  – The Pi user interface can be modified by customizing themes.
  – Custom themes can be created and stored in the `.pi` directory structure.

  7. **Context Files**
  – Pi uses the standards `agents.md` and `clod.md` for context management.
  – These files can be stored in various directories (e.g., workspace or home directory) and loaded by Pi.

  8. **Extensions**
  – Extensions enable customization and enhancement of Pi’s functionality.
  – Example extensions include welcome messages and security prompts before dangerous commands (e.g., `rm -rf`).

  9. **Packages**
  – Packages are bundles of extensions, skills, and prompts that can be installed together.
  – Examples of packages include `PySubagents`, `ContextMode`, and `MCPAdapter`.

  10. **Sessions**
  – Sessions allow for managing, editing, and navigating threads.
  – Sessions can be named, exported, and saved in various formats.
  – Users can switch between different sessions, duplicate them, or merge them.

  The video explicitly focuses on the Pi Agent Harness and is geared more toward intermediate users who already have experience with agent harnesses and want to customize them individually.

Alex Finn (4 new videos)

• Hermes Agent is blowing me away…
  9.5.2026, 20:54:26

  The video compares the AI agents Hermes and OpenClaw and recommends Hermes due to its reliability, self-improvement, and user-friendliness. The author describes Hermes’ advantages, including regular, thematic updates, self-improving capabilities through usage, and a strong emphasis on experimentation and local models. Installing Hermes is described as straightforward, with options for different models and communication services, with Telegram and Opus recommended. The author showcases three use cases: one for beginners discovering new AI tools daily, one for advanced users performing daily proactive check-ins, and one for experts creating AI-generated videos. The author emphasizes the importance of brain-dumping and reverse-prompting to use the AI agent personally and effectively.

  The video explicitly covers the AI tools Hermes Agent and OpenClaw and is aimed more at intermediate and advanced users.

• LIVE: Anthropic and Elon just teamed up to take down OpenAI
  6.5.2026, 20:12:34

  The video covers the strategic alliance between Anthropic and Elon Musk’s XAI (X.AI), characterized by a major computing power deal. Anthropic gains access to SpaceX’s Colossus-1 cluster, which will significantly improve their ability to develop and train AI models. This partnership marks a turning point in competition with OpenAI, which has taken a dominant position in recent months with Codeex. Anthropic has faced reduced limits and less powerful models during this time, which the new alliance aims to address. Elon Musk, previously critical of Anthropic, is now providing massive computing resources, changing the dynamics of AI competition. The video also discusses Elon Musk’s long-term strategies, which may focus on larger goals like autonomous vehicles, space exploration, and robotics rather than remaining in AI chatbot competition. The alliance could usher in a new era of innovation and improvement in AI tools that consumers will benefit from. The video emphasizes the importance of using both leading AI tools, Claude Code and Codeex, to benefit from their respective strengths.

  **AI Tools/Models/Providers:** Anthropic, OpenAI, Elon Musk (X.AI), Claude, Codeex, Grock, Gemini, Open-Source
  **Target Audience:** Intermediate

• Hermes Agent might have just killed OpenClaw
  5.5.2026, 21:11:59

  The video presents Hermes Agent as a more reliable alternative to OpenClaw and discusses seven new features that improve productivity and user-friendliness. These include:

  1. **Kanban Board**: Enables multitasking through parallel processing of multiple task threads. A manager agent fills tasks with details and moves them through various statuses (Triage, To-Do, Ready, In Progress, Block, Done).

  2. **Slashgo**: A high-level mission function that assigns long-term tasks to the agent that can be worked on over an extended period. Prompt quality is crucial for good results.

  3. **Profiles (Multi-Agents)**: Allows creation of multiple agents with their own memories and abilities to optimize performance and avoid overload.

  4. **Model Catalog**: Simplifies switching and assigning models to specific tasks, improving cost control and efficiency.

  5. **Compression**: By adjusting the compression threshold to 0.5, less drastic compressions are performed, improving memory capability.

  6. **Curator Feature**: Automatically culls rarely used skills every seven days to reduce bloat and maintain performance.

  The video criticizes OpenClaw for frequent updates that cause instability and performance issues, and highlights Hermes’ targeted, reliable updates. It’s recommended to leverage Hermes’ new features to boost productivity.

  **Closing Comment**: The video explicitly covers Hermes Agent and OpenClaw and is aimed more at intermediate to advanced users.

• LIVE: Is Hermes better than OpenClaw? FINALE!!!
  4.5.2026, 21:53:53

  The YouTube video shows a live stream where the host tests various AI agents (OpenClaw and Hermes) in a competition called “Agent Olympics.” The stream is unusually long (3.5 hours) and is divided into various sections ranging from technical tests to personal discussions to spontaneous decisions.

  **Content Summary:**
  1. **Agent Olympics:**
  – The host tests four combinations of AI agents (OpenClaw and Hermes with different backend models like ChatGPT and Opus) on five different tasks.
  – Tasks include creating infographics, animated music videos, and other complex projects.
  – Results are rated live, with OpenClaw with Opus ultimately emerging as the winner.

  2. **Technical Discussions:**
  – There are extensive discussions about the stability and reliability of various AI agents, particularly Hermes, which is criticized for “compaction” errors (loss of working states).
  – OpenClaw is praised for its consistency and user-friendliness.

  3. **Personal Topics:**
  – The host discusses his sleep issues and experiments with various solutions like kiwis and magnesium.
  – There are discussions about work methods, including the use of treadmills and standing desks, with the host expressing his preferences and dislikes.

  4. **Community Interaction:**
  – Viewers are actively engaged in the chat, asking questions and providing feedback.
  – The host spontaneously decides to create a second YouTube channel called “Alex Finn Labs,” leading to an entertaining interaction with a viewer who had already reserved the desired channel name.

  5. **Announcements and Future Plans:**
  – The host announces plans to publish more videos about Hermes and multi-agent setups in the future.
  – There is discussion about whether live streams should be held at later times to reach a broader audience.

  **Closing Comment:**
  The video explicitly covers the AI tools OpenClaw, Hermes, ChatGPT, and Opus. It is aimed more at intermediate and advanced users, as it covers technical details and advanced applications of AI agents.

Bart Slodyczka (1 new video)

• LM Studio Is Getting Insane — Start Using It Now
  4.5.2026, 11:31:00

  The video demonstrates how to run AI models locally on your own computer using LM Studio. LM Studio is a free desktop app that lets you search, download, and use various AI models. The video explains how to check your computer’s compatibility with the models, download models, and load them into LM Studio. It demonstrates how to interact with the models in a chat-like interface, upload PDFs, analyze images, and perform web searches. Additionally, it shows how to integrate the models with other business tools like Claude Co-work, Claude Code, OpenClaw, or Hermes Agent. The difference between cloud-based AI and local AI is explained, with local AI being private and free but often slower and less powerful. The video also covers technical aspects such as using Node.js and configuring MCP-Tools (Model Context Protocol) for integrating external tools. Finally, it shows how to integrate a local AI model into the Claude Co-work app.

  The video explicitly covers LM Studio, Gemma 4, Claude, Node.js, and MCP-Tools, and is better suited for intermediate users.

Ben AI

No new videos in this period.

Brian Casel

No new videos in this period.

Coding with Lewis

No new videos in this period.

Cole Medin (3 new videos)

• AI YouTube Is Only Claude Hype Now
  7.5.2026, 00:01:02

  **Summary:**

  The video presents the live creation of a roadmap for Archon, an open-source harness builder for AI-coding. The host begins with an introduction to Archon, which enables packing AI-coding processes into workflows and running them in parallel across different codebases. Workflows are defined in YAML and can include both agent-based and script-based actions.

  A central theme is the concept of a marketplace for Archon workflows, which should allow users to share and install their own workflows. The host discusses various approaches for implementing this marketplace, including whether workflows should be hosted in separate repositories or integrated directly into the Archon repository. The pros and cons of both approaches are discussed, with security and user-friendliness at the forefront.

  During the video, the current Archon roadmap is also created and updated, with various features and improvements added, such as support for the PI-Coding Agent and optimization of the setup process. The host uses Claude Code and other tools to visualize and refine the roadmap.

  **Final Comment:**

  The video explicitly covers open-source tools and models, particularly Claude and Claude Code, and is better suited for intermediate to advanced users.

• 🔴LIVE – What’s Next for Archon – Live Roadmap Session
  5.5.2026, 03:58:13

  **YouTube Video Summary:**

  The video demonstrates the creation of AI-generated videos using Archon, an open-source harness builder. The process involves using Archon for the workflow, 11 Labs for voice generation, and Remotion for video generation. The creator demonstrates the creation of a video about the new TI-84 EVO calculator and then explains the steps for creating a marketing video for Archon itself.

  **Steps and Tools:**
  1. **Archon Workflow:**
  – Archon is used to create a workflow that divides video generation into multiple steps.
  – The workflow includes planning the video, generating audio and video, and validating and iterating.

  2. **11 Labs for Voice Generation:**
  – 11 Labs is used to generate voice output.
  – The creator experiments with various parameters such as speed, stability, similarity, and style exaggeration to achieve the best sound quality.

  3. **Remotion for Video Generation:**
  – Remotion is used to create the visual elements of the video.
  – The creator shows how the workflow plans and generates different scenes and animations.

  4. **Iteration and Improvement:**
  – The creator demonstrates how to iteratively improve the generated video by providing feedback and adjusting the workflow.
  – It shows how to enhance voice output and sound effects.

  5. **Archon Workflow Details:**
  – The workflow is defined in a YAML file and includes multiple steps orchestrated by Archon.
  – Each step can use different models and providers, increasing workflow flexibility and reliability.

  **Final Comment:**
  The video explicitly covers Archon, 11 Labs, Remotion, and open-source models. It is better suited for intermediate and advanced users who are familiar with AI tools and workflows.

• 🔴LIVE – Full AI Video Generation Workflow Using Claude Code + Remotion + Archon
  3.5.2026, 03:33:47

  The video is a revised version of a live workshop on AI transformation conducted jointly with Leor Weinstein. The focus is on creating a foundational system for reliable and repeatable results with AI-coding assistance. The process is divided into three phases: ideation with coding agents, building an iterative loop (PIV-Loop), and gradually advancing the coding agents over time.

  1. **Ideation with Coding Agents:**
  – Unstructured conversations with the coding agent to gather ideas and clarify requirements.
  – Use of tools like Claude Code and Jira to manage and organize work.
  – Creation of a Product Requirement Document (PRD) through specific commands and skills that structure the conversation.

  2. **PIV-Loop (Plan, Implement, Validate):**
  – **Plan:** Analyze the codebase and create a detailed plan for implementing a Jira ticket.
  – **Implement:** Delegate coding to the coding agent based on the created plan.
  – **Validate:** Automated validation by the coding agent, followed by manual code review and manual testing.

  3. **System Evolution:**
  – Retrospective analysis after each PIV-Loop to improve systems and processes.
  – Adaptation of rules, commands, and skills to prevent future errors and increase efficiency.

  The workshop emphasizes the importance of maintaining control over the process by handling planning and validation yourself, while delegating the actual coding to the AI tool. It demonstrates how to work efficiently with tools like Claude Code and Jira to boost productivity and automate repetitive tasks.

  The video is better suited for intermediate and advanced users, as it assumes viewers already have basic knowledge of software development and project management. It specifically covers tools like Claude Code and Jira.

“`html

Dave Ebbelaar (1 new video)

• If I Started AI Freelancing in 2026, I’d Do This
  4.5.2026, 15:15:23

  **Summary:**

  The video content describes a three-stage framework that the author calls the “Data Freelancer Blueprint” for succeeding as a freelancer in AI and data. The three stages are:

  1. **Get Going:**
  – Overcome psychological barriers and start with simple but useful projects, often considered “boring,” such as data automation or reporting.
  – Create three end-to-end projects that you can demonstrate, and learn how to integrate and deploy code into real systems.
  – Update your LinkedIn profile to clearly communicate what problems you can solve.

  2. **Getting Paid:**
  – Determine your hourly rate through research.
  – Use your network for warm introductions and conversations with decision-makers.
  – After the conversation, create a detailed project proposal with milestones, deliverables, and cost estimates.

  3. **Get Good:**
  – Focus on improving leads, sales, and delivery.
  – Build long-term contracts to secure stable income.
  – Use various channels like LinkedIn, YouTube, and freelance platforms to generate leads.

  The author emphasizes that freelancing in tech is a safe and lucrative way to start a business and encourages viewers to take the first step.

  **Final Comment:**
  The video features tools like N8N and Airtable (low-code/no-code solutions) as well as Python and TypeScript (custom code solutions) and is aimed at intermediate and advanced freelancers looking to break into or scale their AI and data business.

“`

David Shapiro

No new videos in this period.

Everlast AI (4 new videos)

• “You’re in for a surprise” Robotics professor on China, humanoids & the future
  7.5.2026, 15:15:00

  This video is an interview with Alexander from Everlast AI about the current state and future of humanoid robotics, particularly focusing on the robot assistant Gami from TU Munich. The main topics include the exponential development of robotics, the integration of Visual Language Action Models (VLAs), challenges in manipulation and robot sensitivity, as well as ethical and legal questions regarding AI liability. Alexander emphasizes the importance of design and interaction design for societal robot acceptance and discusses the role of open source and proprietary solutions in robotics. He sees major progress in simulation and collective robot learning but warns of challenges in safeguarding and liability. The video explicitly discusses models and providers like Nvidia (Omniverse, Kosmos) and emphasizes the importance of fundamental research and industrial application. It’s aimed at intermediate and advanced viewers, as it contains in-depth technical and ethical discussions.

  Final comment: The video explicitly addresses Nvidia (Omniverse, Kosmos) and is intended for intermediate and advanced viewers.

• Robotics expert: Forget humanoid robots! THIS is what will really happen (Dr. Hendrik Susemihl)
  6.5.2026, 15:15:00

  This video is an interview with Dr. Hendrik Susemihl, founder of Goodbites, a company developing fully autonomous robot kitchens. Goodbites produces over 50,000 meals daily and won the Robotics Award 2026 at the Hannover Messe. The company supplies autonomous robot kitchens to the US Army and is a global leader in this field.

  Hendrik Susemihl discusses the importance of Physical AI, AI that receives a physical body, and how it will fundamentally change our everyday lives. He explains why cooking is harder to automate than factory work and how Goodbites won the US Army as a customer. He also addresses Germany’s role in the global robotics race and emphasizes the need to think boldly and act quickly to keep up with competition from the US and China.

  Susemihl highlights that while Germany is technologically leading, it lacks the ability to build global companies. He stresses the importance of speed and willingness to experiment, as well as the necessity to think boldly to stay competitive. He sees Germany’s biggest mistakes as being too passive and not experimenting enough.

  In summary, the video discusses the importance of Physical AI and Germany’s role in the global robotics race. It emphasizes the need to think boldly and act quickly to stay competitive.

  The video addresses specific tools and providers like Goodbites and is aimed at intermediate and advanced viewers.

• Don’t use Codex until you’ve watched this video! (The ChatGPT SuperApp)
  5.5.2026, 15:15:00

  This video shows 34 tips and tricks for using Codex, a new SuperApp from OpenAI that’s supposed to replace ChatGPT. Here are the main points:

  1. **Basics**: Codex can be used like ChatGPT but offers additional features like direct access to local folders and files.
  2. **How it works**: Codex can directly access local folders and edit files, which is a major advantage over ChatGPT.
  3. **Excel spreadsheets**: Codex can create and edit Excel spreadsheets directly in local folders.
  4. **Design customization**: The Codex interface can be customized to individual needs, including font size, color scheme, and typeface.
  5. **Skills and plugins**: Codex offers various skills and plugins that extend functionality, such as creating Word documents and PDFs.
  6. **Subagents**: Codex can use multiple subagents simultaneously to handle complex tasks and manage context.
  7. **Browser and computer use**: Codex can control the browser and local applications, making task automation easier.
  8. **Automations**: Codex can set up automations to handle recurring tasks.
  9. **Chronicle Research Preview**: A feature that builds memories and context based on daily workflows.
  10. **Open Source**: Codex is open source, allowing users to add additional features.

  The video is aimed at intermediate users and explicitly addresses Codex, OpenAI, and specific tools like Remotion and Paper.

• AI News: Codex “Goal” works for DAYS, Claude takes over ADS, Opus 4.7 hacked & OpenAI “Goblins”
  3.5.2026, 08:15:00

  This video covers various current developments and trends in AI technology. It begins with the introduction of Project Deal from Anthropic, which enables trading between AI models and allegedly brought eBay to a standstill. Additionally, the new Codex Goal feature is presented, which allows AI agents to work autonomously for several days. New features from Gemini and Claude Code, which are directly integrated into CAD software, are also mentioned. It points out the token problem with Claude Code, which has been officially confirmed, while OpenAI is massively subsidizing Codex.

  Another focus is on AI system security. Johann Rehberger, a well-known AI security researcher, demonstrates how to manipulate Claude Opus 4.7’s memory using a ChatGPT image. He also discusses security vulnerabilities in Claude Codework and other attack vectors.

  The video also introduces the new Meta CLI, which allows AI agents to completely control and manage ad campaigns. It shows how to use this tool to analyze campaigns and create new creatives.

  Additionally, updates in Microsoft 365 and open-source alternatives to American AI tools like MicS for the legal sector are presented. The video ends with a discussion about AI job losses and AGI’s impact on the labor market, based on statements by Demis Hasabis and discussions with Professor Peromitsicit and Professor Dr. Andreas Moring.

  **Final comment:** The video addresses OpenAI, Claude Code, Gemini, Codex, and specific tools like Meta CLI and is aimed at intermediate and advanced users.

Fireship (2 new videos)

• Every operating system concept in one video…
  7.5.2026, 17:32:34

  The video explains in detail how an operating system works from the moment the power button is pressed until shutdown. It starts with the bootloader, which loads the operating system, then moves on to privilege rings, which separate the rights of the kernel and applications. Virtual memory is described as a system that allows multiple applications to run in parallel without interfering with each other. The kernel sets up the file system, loads device drivers, and enables interrupts, which allow the system to respond to inputs. The kernel then starts the first process (PID1), which is the ancestor of all other processes. System calls enable applications to communicate with the kernel, and the scheduler manages CPU time across many processes. Threads allow applications to execute multiple tasks simultaneously, and inter-process communication (IPC) enables different processes to communicate safely. Finally, the shutdown process is described, where all processes are terminated and the system is safely shut down.

  The video addresses operating systems and their components in general, without naming specific tools or vendors, and is more suitable for intermediate or advanced viewers.

• 732 bytes of Python just borked every Linux machine on earth…
  4.5.2026, 18:40:40

  The video covers a critical security vulnerability in the Linux kernel, referred to as “copy fail” (CVE-2023-31431), which has existed since 2017 and was discovered by an AI tool. The vulnerability allows a local user to gain root access by writing four bytes to the page cache of a read-only file. All Linux distributions updated after 2017 are affected. The vulnerability was exploited through a Python script that uses the ONC ESN protocol and the AF_AGL interface. Although the vulnerability is not remotely exploitable, it is strongly recommended to update systems. The video also mentions the role of AI in discovering security vulnerabilities and promotes Code Rabbit, an AI tool for improving code quality.

  The video explicitly addresses AI tools such as the AI agent tool used by Theori and Code Rabbit, and is intended for intermediate to advanced 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)

• These 20 Claude Code Tricks Save You Hours
  7.5.2026, 19:15:23

  The video presents 20 tricks and best practices for effective use of Claude Code that are designed to help in various work situations. The key tips include:

  1. **Use planning mode**: Create a plan before starting a project or making major changes and discuss it with Claude to avoid misunderstandings.
  2. **Ask questions**: Specifically ask Claude to clarify before it starts working to ensure all details are addressed.
  3. **Discuss instead of command**: Treat Claude as a sparring partner and present problems openly instead of prescribing finished solutions.
  4. **Autoaccept mode and permissions file**: Use autoaccept mode for more efficient work and create a permissions file to restrict access to sensitive files.
  5. **Use voice input**: Enter long prompts via voice input to save time and convey context better.
  6. **Slashinit for project onboarding**: Use the slashinit command to quickly onboard Claude to a project and create project memory.
  7. **Maintain Claude MDI**: Actively maintain the Claude MDI file and store important information there to provide Claude with relevant info in every session.
  8. **Emergency brake and rewind**: Use the Escape key to stop or reset Claude if something goes wrong.
  9. **File mentions as shortcuts**: Reference files directly in the chat to make it easier for Claude to access important information.
  10. **Use screenshots**: Use images or screenshots to better explain designs or layouts.
  11. **Self-control**: Instruct Claude to check its own work to avoid correction loops.
  12. **One task per session**: Work on only one task per chat to avoid overloading the context window.
  13. **Use command/context**: Check the context status to see what’s taking up the context window and free up space.
  14. **Slashcompact and Slclear**: Summarize the context window with slashcompact or clear it with slclear to free up space.
  15. **Custom skills**: Outsource regular tasks to skills to avoid having to explain them to Claude each time.
  16. **Use models based on task**: Deploy different models (HighQ, Sonnet, Opus) depending on task complexity.
  17. **Hooks for notifications**: Use hooks to be notified when Claude has completed a task.
  18. **Subagents and agent teams**: Use subagents for token-intensive tasks and agent teams for complex, parallel work.
  19. **Ultraathink keyword**: For difficult problems, use the keyword “Ultraathink” to give Claude more time to think.
  20. **Use worktrees**: Use worktrees for parallel work on different features of a project to avoid conflicts.

  The video explicitly addresses Claude Code and is geared toward intermediate users.

• Claude Code Has Never Been This Easy!
  3.5.2026, 15:26:48

  The video presents the latest updates to Anthropic’s Claude Code desktop app, which has significantly improved the use of Claude Code. Claude Code is a tool that runs on your own computer and has access to a folder to create, read, and execute files. Previously, Claude Code could only be used in limited ways in the desktop app, but the update now enables much more comprehensive usage. Users can now work directly with Claude Code in the desktop app, view files, edit them, and even preview the created apps. The video shows how to create a flashcard app by giving Claude Code access to a folder and then describing the app. Claude Code plans and creates the app automatically, with the user able to accept or adjust the plan. The desktop app now also offers an auto mode where Claude Code executes commands that are not dangerous and only asks for permission for riskier commands. Additionally, the user can view the created files, edit them, and even open the terminal to execute further commands. The video also shows how to use multiple sessions simultaneously and what limitations still exist, such as displaying images and PDF files. Overall, the desktop app now offers much better overview and functionality for using Claude Code.

  The video explicitly addresses Anthropic’s Claude Code and is geared toward intermediate users.

Kyle Balmer | AI with Kyle (4 new videos)

• How to Prompt AI Better Than Marc Andreessen (A Billionaire)
  8.5.2026, 05:15:03

  # Summary

  The author critically analyzes Marc Andreessen’s viral system prompt and explains why it’s from 2023 and doesn’t work.

  **Problematic aspects of the Andreessen prompt:**

  Opening with “You are a world-class expert in all domains” and statements about superior intellectual abilities don’t actually make AI better – this is scientifically disproven. Adjectives like “sharp” or “erudite” without concrete instructions are ineffective. “Never hallucinate” doesn’t work because the model doesn’t know what it’s hallucinating (it doesn’t have an indexed fact database; it fills knowledge gaps with probabilities). Asking for responses to be as long and detailed as possible leads to token waste and unusable padding.

  **What works better:**

  The author presents his “RISEN” framework (Role, Instructions, Steps/Sequence, End Goal, Narrowing):
  – **Role**: Focuses the response style, doesn’t make the AI smarter
  – **Instructions**: Concrete verbs instead of adjectives (“challenge weak assumptions” instead of “be intelligent”)
  – **Steps**: Needed when a process follows; modern models already think sequentially
  – **End Goal**: Define specifically what success means – not just “as long as possible”
  – **Narrowing**: Constraints and formats at the end

  A revised prompt using this framework would be shorter, more precise, and waste less context window with each run. The author distinguishes between system prompts (for projects/custom instructions, write once, use repeatedly) and everyday conversation prompts (informal, no prompt engineering needed).

  **Practical**: Find custom instructions in ChatGPT under Account → Personalization → Custom Instructions, or in Claude under Settings → General → Instructions. In ChatGPT web, a full system prompt is only possible at the project level.

  The author offers a quick-start guide in the newsletter.

  —

  *Claude and ChatGPT are mentioned throughout as examples; the clip analyzes a viral system prompt as opinion/reflection with practical improvement suggestions.*

• The $10Bn Move OpenAI & Anthropic Are Making That You Can Copy
  7.5.2026, 05:30:15

  # Summary: Major AI labs move into consulting and advisory services

  On May 4th, OpenAI and Anthropic announced new ventures almost simultaneously to implement AI in enterprises not just as tools but as integrated services. Anthropic partners with Blackstone, Hellman & Friedman, and Goldman Sachs; OpenAI founded “The Deployment Company” with 19 investors including BCG, McKinsey, Accenture, and Capgemini. The core model: AI engineers are deployed directly into companies to understand workflows and restructure them around AI. The reason is the “Jagged Frontier” concept – AI isn’t equally good everywhere, but has particularly improved at coding and now scales to white-collar work generally. A Harvard Business School study showed consultants with AI work 25% faster with 40% higher quality. The playbook follows Palantir’s strategy: through deep system integration, these companies become indispensable.

  **Opportunities for individuals:** While the labs address Fortune 500 companies, a gap emerges for freelancers and small teams to offer the same service in the mid-market and smaller firms. The strategy is combining industry knowledge with AI expertise – exactly what the big labs are learning. Concrete approaches include advisory/consulting (retainer basis), implementation (project basis or internal SaaS), and education/content for your own industry. The easiest entry point is workshops ($2,000–$5,000 per hour, scaled across multiple participants); these open doors to further projects. Long-term, AI-driven automation will threaten all white-collar jobs – but over the next few years, lucrative opportunities emerge for professionals with industry knowledge. A free guide is available at aiwithkyle.com/workshop.

  The video explicitly addresses OpenAI and Anthropic (Claude) as providers and is opinion/reflection with strong practical focus.

• AI Hallucinations Explained: Why They Happen & How to Stop Them
  6.5.2026, 05:00:35

  # AI Hallucinations: Causes, Examples, and Solutions

  The video comprehensively covers what AI hallucinations are, why they occur, and how to handle them.

  **What are hallucinations?**
  There are two types: First, completely inventing non-existent information (like the Chicago Sun Times adding books by fictional authors to their reading list). Second, the more insidious form where true information is mixed with false – misquotes, incorrectly adapted numbers – which is harder to detect because it’s closer to reality. High-ranking institutions like Deloitte published reports with fabricated footnotes and hallucinations that were sold to governments.

  **How do hallucinations work mechanically?**
  Large language models don’t function like search engines retrieving information from a database. They stochastically predict the next token based on training data (billions of words from the internet, books, transcripts). They learn patterns of what correct answers look like but don’t generate facts – they create probabilistic likelihoods. This isn’t a bug but a core feature of LLMs: they are “dream machines,” as Andrej Karpathy explains. Everything an LLM outputs is technically a hallucination; value judgments alone distinguish “good” (intelligent) from “bad” (factually false) hallucinations. LLMs therefore have no concept of true or false – they can’t lie but can output information that doesn’t match reality.

  **When are hallucinations likely?**
  – Specific facts (quotes, statistics, names, dates, citations) – the model likes to invent these to be helpful
  – Niche topics with less training data
  – Recent events beyond the knowledge cutoff date
  – False premises: AI agrees with leading questions and hallucinates to confirm the premise

  Statistics: 92% of users don’t verify AI responses; 45% of AI responses have significant issues; 34% of the time the model is more confident when it’s wrong.

  **Practical measures to reduce them:**
  1. **Enable web search** – reduces factual errors by 45%
  2. **Stronger models** for important tasks (Claude Opus 4.6/4.7, GPT-5 Pro with Thinking, Gemini 2.0/3.1 Pro) – these use extended thinking or deep research
  3. **Use Notebook LM** – offers personal retrieval-augmented generation (RAG), constrains the model to provided sources
  4. **Better prompting**: instead of “Is this a good idea?” have it compare two options; specify sources and explicitly ask statistics be verified; tell the model accuracy is critical
  5. **Paid plans**: free models are weaker and often lack web search

  **Can they be fixed?**
  No – not fundamentally. Hallucinations don’t decrease but increase as models become more complex (reasoning models). That’s because hallucination is the core function of LLMs. They improve on factual questions because models can use search tools. The solution isn’t elimination but better systems around them (tools, document access, ecosystem integration).

  **Conclusion**: Users must shift their thinking – away from the search paradigm (database query) toward a generation paradigm. Those who understand that LLMs don’t work like smarter search engines but generate creatively will be more successful with AI.

  —

  **Models/tools discussed**: Claude (Opus), ChatGPT (including GPT-5), Gemini, Notebook LM, Perplexity; **Format**: Opinion/reflection with how-to elements.

• How to Write With AI and Never Get Caught
  5.5.2026, 05:00:05

  # Summary: OpenAI bans goblins – AI detection signs and countermeasures

  The video discusses why OpenAI explicitly banned the word “goblins” (and similar creatures like gremlins, trolls, raccoons) from its models, using this to comprehensively explain how AI writing patterns emerge, are detected, and can be avoided.

  **The goblin case:** OpenAI hardcoded into system instructions that models shouldn’t discuss goblins – unless absolutely relevant. This happened because the “nerdy personality” in ChatGPT created a reinforcement loop: a “nerdy” tone was associated with mythical beings in training signals, these were rewarded, appeared increasingly often, and solidified themselves through reinforcement learning.

  **AI detection signs (“tells”) are patterns like:**
  – Vocabulary: “delve”, “nuanced”, “tapestry”, “navigate”, “foster”, “leverage”, “pivotal”, “meticulous”, “vibrant”, “robust”, “showcase”, “multifaceted”
  – Structure: “It’s not X, it’s Y” (negative parallelism), bullet points for short answers, Markdown headers, key takeaways
  – Tone: Introductions like “Great question”, artificially inflated, seemingly weighty style with hollow substance
  – Punctuation: Em-dashes are overused

  **The Wikipedia resource:** Wikipedia banned AI writing and maintains a 15,000-word guide for editors with a comprehensive list of these signs. It warns against undue emphasis on significance, vague attribution, superlatives.

  **Practical defense (belt and braces):**
  1. *Before writing:* Feed the Wikipedia list as system instructions/custom instructions into the AI model to avoid these patterns from the start. A precompressed set called “Humanizer” offers pre-made instructions.
  2. *After writing:* Check output against the same checklist (as a skill or prompt) to eliminate remnants.

  **On AI detectors:** The speaker emphasizes that automated AI detectors (TurnItIn, Copyleaks, etc.) don’t work and cause harm. OpenAI shut down its own text classifier in 2023 – generation tools evolve faster than detection tools. These detectors systematically disadvantage non-native English speakers and are increasingly sued by universities for falsely flagging human work as AI.

  **Moving targets:** These signs aren’t static. “Delve” was the notorious word of 2024, then was explicitly removed from models. Em-dashes were “fixed” by OpenAI. This means: those wanting to avoid AI writing must update the list regularly; those wanting to detect it must practice on Reddit and LinkedIn and follow Twitter.

  **The disclosure paradox:** 94% of news consumers say journalists should disclose AI use – but 42% trust the story less once they see the disclosure. In practice: people often notice no difference and consume AI content without problems if not explicitly disclosed.

  The speaker offers free cheat sheets and skill files for Claude and ChatGPT to implement both sides (AI detection and AI avoidance) practically.

  **Explicitly discussed providers/models:** OpenAI (ChatGPT, Codex), Claude; independent practical resource: Wikipedia Signs of AI Writing. **Format:** Deep dive/opinion with practical application tools, audience clearly defined (everyone who writes with AI or needs to detect AI writing).

Leon van Zyl (4 new videos)

• I Turned Hermes Agent Into a Coding Agent
  8.5.2026, 11:02:33

  The video demonstrates how to use the Hermes Agent as a coding agent to create a web app and deploy it online. The process includes setting up the Hermes Agent on a VPS, integrating it with Telegram for communication, installing the Vercel CLI tool for deployment, and configuring the necessary skills for the agent. The creator tests whether the agent can create a personal portfolio page by scraping information from the creator’s YouTube channel and creating an appealing frontend design. The agent successfully creates the app, deploys it on Vercel, and provides a public URL that opens the app in a browser. Additionally, it’s shown that the agent is capable of making changes to the app and deploying them automatically.

  The creator concludes that Hermes as a coding agent is suitable for simple tasks and quick dashboards but not for complex software projects. The video explicitly covers Hermes Agent, OpenAI Codex, GPT 5.5, Vercel, and Telegram. It is aimed more at intermediate and advanced users.

• Build Apps with LocalForge: A Free Local Coding Agent
  6.5.2026, 12:25:00

  The video introduces the open-source tool “Honeyfree,” which allows you to autonomously plan and implement software projects. The user describes to the tool what they want to build, and it plans the features, adds them to a Kanban board, and implements them automatically. The tool supports various models like Alum Studio and Ollama and can break down complex tasks into smaller features. The user demonstrates creating a simple to-do app and shows how new features can be added and implemented. The video emphasizes that this is now possible with free models, which wasn’t the case a few months ago. It also explains how to download models like Qwen 3.6 or JML4 and use them in Alum Studio or Llama Studio. The user recommends increasing the context window length of the models to at least 64,000 tokens for better performance. The video also shows how to install and set up Local Forge to create and manage projects. It’s emphasized that while free models are good at writing code, they require more detailed instructions for better results. The user recommends using a paid model like Claude for planning features, while using free models for the actual implementation. The video ends with an invitation to sign up for a masterclass course that teaches building applications with coding agents.

  The video covers open-source models like Qwen 3.6 and JML4 as well as tools like Alum Studio, Llama Studio, and Local Forge. It is more suitable for intermediate and advanced users who already have experience using AI models and software development.

• OpenCode Tutorial for Beginners: Setup, Agents, Skills & MCP
  5.5.2026, 12:33:17

  The video is a tutorial that shows how to create a Next.js application using OpenCode, an open-source AI tool. The process begins with installing and setting up OpenCode, including connecting with various AI models and providers, both free and paid. The tutorial demonstrates how to add agent capabilities like frontend design and Next.js skills to improve the quality of generated code. It also shows how to use memory files and design systems to increase the agent’s consistency and efficiency. The tutorial continues with creating an application that allows users to input a rough idea of their app and receive a detailed project plan. The agent uses subagents to perform tasks in parallel and protect the main context. At the end, the application is tested and improved, with the agent completely redesigning the UI and running automated tests. The video is suitable for intermediate and advanced users interested in AI-powered coding tools.

  AI tools/models/providers: OpenCode (open-source), OpenAI, Anthropic, Gemini, OpenRouter, BigPikko, HY3, Minimax, Nvidia, Vercel, Cintra AI.

• I Built a Full App Using Only Cursor AI
  4.5.2026, 11:01:35

  In this video, an AI-powered YouTube summarizer is developed using the Cursor tool. The process begins with creating a user interface that takes a YouTube URL and delivers a video summary. The requirements include a short summary (TLDR), five to eight key points, a “Watch these moments” section with timestamps and descriptions, as well as the original video link.

  The creator uses Cursor and chooses the Composer 2 model to scaffold the project. They install Next.js and the Shad cn library for the user interface. With the help of Cursor’s agents, a basic user interface is created that meets the requirements. Subsequently, functionality is added to retrieve the transcript of a YouTube video using the YouTube Transcript API.

  For the AI-powered summary, the AI SDK from Cursor is used to return structured data. The creator chooses the “anthropic/claude-2” model from OpenRouter and integrates the API key in an .env file. The agent then generates the summary, including the TLDR, key points, and recommended moments from the transcript.

  The video explicitly covers the tools Cursor, Composer 2, Next.js, Shad cn, YouTube Transcript API, AI SDK, and OpenRouter. It is aimed more at intermediate and advanced users.

Liam Ottley

No new videos in this period.

Mark Kashef (2 new videos)

• Build Your Agentic OS Better Than The 99%
  9.5.2026, 20:00:14

  This video covers the practical implementation and optimization of agentic operating systems (Agentic OS) like Open Claw, Hermes, or Claude Code to achieve real business value. The focus is on the often-overlooked “behind-the-scenes grunt work” necessary to effectively deploy such systems. The author emphasizes the importance of data readiness and organization before diving into dashboards or flashy features.

  Key points include leveraging Skills (both project-specific and global), Hooks (event-based automations), and Claude MDs (as central control documents). The author introduces a tool called “Skill/Silver Platter” that helps organize and summarize data to enhance agents’ analytical capabilities. Tactical tips are also provided on how to effectively use integrations and APIs to optimize your systems.

  The video showcases three case studies (Marco, Sally, and Dr. Sana Anwar) that illustrate how the described methods can be applied across different industries and business models. The author emphasizes the importance of strategically deploying agents and continuously improving them, much like training and developing employees in a company.

  At the end, the “Skill/Silver Platter” tool is offered for download, and a course is referenced that dives deeper into the subject matter.

  The video explicitly covers Claude Code and is better suited for intermediate to advanced users who already have some experience with agentic systems.

• This Claude Code Setup Runs My Entire Business
  3.5.2026, 19:00:21

  This video provides a detailed demonstration of a custom-built AI Operating System (AIOS) that the creator uses for his business. The system consists of a variety of agents, each covering specific tasks and areas of expertise. At the core of the system is a “Hive Mind,” which serves as a shared state of memory and visualizes agent activities and knowledge. The creator shows various views and features of the system, including 3D and 2D graphical views, a list of all tasks, and a “War Room” function for communicating with agents via text and voice.

  A practical example is the use of a meta-agent connected to the Meta Command Line Interface to analyze advertising performance and generate reports. The system also enables task automation and reminder creation. The creator emphasizes that the system is built on solid data organization and that implementation requires time and iteration. He offers resources and a carbon-copy version of his system accessible via a link in the video description.

  Final note: The video explicitly covers Claude, OpenAI (Cloud Code), Gemini, and specific tools such as Telegram, Discord, Slack, Loom, and SQLite. It is designed for intermediate to advanced users.

Matt Pocock (1 new video)

• Burn through the backlog from hell with /triage
  7.5.2026, 15:00:43

  The video introduces a tool called “Triage” designed specifically for managing GitHub Issues and other backlogs. It helps transform unstructured ideas and bug reports into clear, actionable tasks that can be handled by AFK agents (autonomous AI agents). The tool uses a state machine system with two categories (Bug and Enhancement) and five states (such as “needs triage”, “ready for agent”, “won’t fix”) to clearly classify each issue. Users can work through individual issues or the entire backlog to categorize them and process them directly if needed. The video showcases a live demo where the creator applies the tool to their own “Sank Castle” repository to triage issues and even fix them directly. It also demonstrates integration with other Skills, like the “Diagnose” Skill, which automatically reproduces and fixes bugs. The tool is designed to streamline collaboration between human developers and AI agents by providing a clear structure for task management.

  The video explicitly covers Claude usage and is aimed at Intermediate or Advanced users who already have experience with GitHub, backlog management, and working with AI agents.

Melvynx (4 new videos)

• Zed AI Agents : La nouvelle feature Zed qui change TOUT
  8.5.2026, 16:10:57

  # Summary: Z – New Features and Agent Integration

  The video presents major updates for Z, a code editor with new AI agent integration. The main innovation is the **ACP Protocol (Agent Client Protocol)**, which allows you to use your own AI agents directly in Z instead of relying on Z’s proprietary paid subscription. This enables users to leverage their existing subscriptions from providers like Codex, Cursor, and Claude directly within Z.

  In the chat interface, you can switch between different agents – for example, from Codex to Cursor or Claude – without leaving Z. Each agent loads its specific model (e.g., GPT-4.5 for Codex, Composer 2 for Cursor) and responds through the provider’s API. Additional agents like Cline, Cortex Code, and others can be added via the ACP Registry.

  The second major innovation is **multi-repository support**: you can open multiple projects simultaneously in different work trees and switch between them. Separate chat sessions can be opened for each repo. The chat features a “Follow” function that live-tracks file changes as the agent works.

  Z remains a full-featured IDE with terminal and file management. The presenter praises the native Rust implementation for speed and responsiveness, but notes occasional performance issues and high RAM consumption when multiple windows are open (for example, 19 GB for Codex simultaneously).

  **Demo with Cursor, Codex, and Claude; topic: Z with ACP Protocol for agent integration.**

• Les DEVS sont inutiles ? Voici des CHIFFRES (preuves) qui montrent l’inverse
  7.5.2026, 15:45:02

  # Summary: The AI Apocalypse is a Marketing Myth

  YouTuber Melvin extensively references an article by David George (Anazinovosti Foundation, May 6) that challenges widespread fears of massive job loss due to AI as unfounded.

  ## The Central Thesis

  The “Apocalypse of Work via AI” is a rehash of the economic sophism of **”fixed stock of labour”** – the assumption that there is a fixed number of jobs that simply vanish through automation. This is historically false.

  ## Historical Evidence Against the Doom Narrative

  **Agriculture**: In the early 20th century, over 60% of jobs were in agriculture. Tractors should have caused catastrophic unemployment – instead, agricultural production sextupled, the population quadrupled, and basic food prices fell by two-thirds. Workers shifted to new sectors (factories, offices, services).

  **Electrification**: Electricity completely reorganized factories and created entirely new consumer product categories. Over 100 years passed between the first electric motor and measurable productivity gains. By 1930, labor productivity had doubled, but not through fewer jobs – through more production, more sales, more business activity.

  **Automobiles**: Price declines led to explosions in both production and employment simultaneously.

  **Spreadsheets (Excel/Visicalc)**: They displaced accountants but created 1.5 million new financial analyst positions – a net gain of 500,000 jobs.

  **Travel Agents**: Their numbers halved due to technology, but remaining agents earn significantly more, and new service categories emerged (coaching, concierge, luxury services) funded by higher incomes.

  ## The Jevons Paradox

  When inputs (intelligence) become cheaper, costs fall, quality rises, speed increases – and demand *explodes*. The economy doesn’t remain static. New products emerge, new demand follows. This leads to new jobs, often in completely unexpected sectors.

  ## The Labor Market Shifts, It Doesn’t Disappear

  The article shows: sectors that were dominant in 1850 (agriculture, transport, energy) were replaced by even larger ones (information, technology, communication). No single sector became as dominant as its predecessor, but total economic volume grew massively.

  ## Current AI Data

  – **90% of surveyed companies report**: No significant impact on overall employment in the last 3 years.
  – **Goldman Sachs analysis**: Professions with high substitution risk (telephone operator, clerk) are marginal; but designers, developers, and creative professions are *not* on the high-risk list.
  – **Augmentation outweighs substitution**: In earnings calls, AI-as-augmentation is mentioned about 8x more often than substitution.
  – **Software developers**: Demand has surged since early 2025 – precisely when AI agents and code tools like Claude Code proliferated.
  – **Product managers**: Open positions are rising continuously and higher than in 2022.

  The data trend shows: redistribution of jobs, not generalized unemployment. Roles augmented by AI (analytical, technical, management) are growing; routine, administrative roles are shrinking.

  ## Why the Doomers Fail

  They focus on task replacement and ignore:
  1. **New boundaries**: Most jobs existing today didn’t exist in 1940.
  2. **Human ambitions don’t freeze**: When work becomes cheaper, people shift their goals to higher-value problems.
  3. **Prosperity spillover**: When there are winners, they spend money – new services emerge (coaching, pet care, nails, education).
  4. **Robotics boom**: AI opens entirely new industries (robotics, AI datasets) that were previously impossible.

  ## Practical Perspective: Developers

  The YouTuber also shares the thesis of another AI YouTuber (Defun Intelligence): current enthusiasm around AI code tools leads to a short-term “breakthrough” – many non-technical people try to launch projects. But they’ll discover that the final 20% (quality, maintainability, scaling, production-readiness) is extremely difficult without real developers. This will lead to *increased* demand for good developers, not less.

  —

  **Format & Context**: YouTuber reviews and comments on an academic/VC article (deep-dive/review) with focus on data validation against the widespread “AI destroys jobs” narrative. No specific AI tools or providers mentioned except Claude Code.

• Codex App ou Claude App : laquelle tu dois utiliser maintenant (comparatif 2026)
  5.5.2026, 16:00:51

  # Comparison: Codex vs. Claude Code – Two AI Coding Interfaces in Practical Testing

  The creator has moved away from his earlier conviction that terminals are the best solution and increasingly uses the graphical applications Codex and Claude Code. In this video, he compares both solutions using concrete workflows for adding a YouTube video feature to an email campaign application.

  **Core insight on usage:** While he previously used individual Claude-3 instances via terminal, he now starts 5–10 parallel tasks across different projects, some taking 25+ minutes. These interfaces shine here: multiple chat sessions side-by-side with clear status overview.

  **Main differences:**

  *Codex excels with:* Developer features (subagent overview with names, PR management with direct links, Git actions directly in UI, browser preview with port overview, terminal splitting, message queuing for feedback without interruption). Model selection is intuitive (intelligence vs. speed). Workflows are better followed. Performance is faster (29 min vs. 40+ min for the same task).

  *Claude Code:* Simpler but also more spare on features. No direct Git action buttons, preview is small and isolated in sandbox (GitHub login doesn’t work). Less UX clarity when working with work trees. However: solid core functionality, better model quality in reasoning (creative solutions), and the creator still uses Claude Code in parallel because of his existing subscription.

  **Special features:** Environments (work-up/work-down scripts for duplicating databases), skill management with slash commands, auto-review by subagents, PR monitoring with scheduled checks.

  **Creator’s conclusion:** Codex is “three lengths ahead” in developer experience. Claude Code is becoming increasingly “normie-friendly” rather than power-user-friendly. He uses both, but favors the Codex interface, even though he often finds Claude’s models better conceptually.

  **Comparison video on Codex, Claude Code (Anthropic) — Demo/Opinion, focusing on developer ergonomics and workflow integration.**

• Claude met fin à la fête : les développeurs s’énervent (sont-ils méchants ?)
  4.5.2026, 17:18:07

  # Summary: The Shift from Claude to OpenAI in AI Development

  The video creator observes a fundamental shift in sentiment within the AI developer community: while Anthropic grows increasingly unpopular, OpenAI gains favor. The reasons are concrete and traceable.

  **Anthropic’s problematic moves:**
  Anthropic has officially blocked the use of open-source models (like Open-Claude and Hermes) with the Cloud subscription. This hits developers hard because these tools consume enormous amounts of tokens – the video creator would have owed $1,429 without the flat rate, while paying $100 monthly instead. Additionally, users attempting to bypass token limits are automatically shifted into costly “extra usage.” CEO Andrew Damodei also spreads drastic statements claiming AI could destroy 50 percent of all white-collar jobs within 1-5 years – rhetoric that repels developers. Internally, Anthropic maintains strict secrecy culture (comparable to Apple), discussing techniques sparingly (e.g., a red-team blog with poor responsive design).

  **OpenAI’s opposing course:**
  Sam Altman of OpenAI articulates a contrasting vision: “We want to create tools that elevate and emancipate people, not replace them through entities.” Practically, this means: OpenAI allows developers to use their ChatGPT subscriptions with code tools, regularly resets token limits, and communicates less fearfully. GPT-4.5 is currently the most intelligent available model for coding by benchmark standards and costs no more than Claude’s Opus model.

  **Market dynamics:**
  The video creator analyzes that OpenAI initially targeted the mass market (ChatGPT for ordinary users) while Claude focused on and won over developers. Now OpenAI has recognized that developers are the key and is pivoting – with noticeably higher token generosity and better pricing than Claude. The creator himself now uses a ChatGPT subscription again and finds that Claude’s limits are reached significantly faster.

  **Personal usage:**
  The video creator now works with both systems in parallel (plus a third tool called Z) and manages interaction not via terminal, but through GUI, since managing the volume of parallel chats is more convenient that way.

  **Format:** Opinion/reflection with concrete data examples; explicitly covers: Claude/Anthropic, OpenAI, GPT-4.5, Open-Claude, code agents, token limits, Cloud subscriptions.

n8n (1 new video)

• How we use n8n: Sindhuja, product leader
  4.5.2026, 14:17:20

  The video is an interview with Sindhuja, a Product Manager on n8n’s AI team. She shares her professional journey, which took her from being an engineer to founding an EdTech company and eventually to n8n. At n8n, she initially improved user adoption and retention before moving to the AI team to drive the use of AI for simplifying workflows. She discusses the development of an instance-level MCP that enables users to use n8n from various platforms, and emphasizes the importance of community feedback. Additionally, she uses n8n herself for personal and professional workflows, including an AI assistant within the product. She sees n8n’s future in combining creativity and reliability to make automation accessible to everyone.

  The video explicitly focuses on the n8n tool and is geared more toward intermediate or advanced users.

Nate Herk | AI Automation (7 new videos)

• This is The Most Powerful Tool to Give to Claude Code
  5/9/2026, 01:55:16

  The video introduces the **Printing Press** tool, a CLI factory and library that enables the creation and use of efficient command-line interfaces (CLIs) for various services. The presenter demonstrates how CLIs offer advantages over APIs and MCPs (Model Context Protocol) in terms of token efficiency and integration into agent workflows. Printing Press provides a library with over 50 pre-built CLIs and a factory for creating custom CLIs. The presenter shows how he built a CLI for a school community without an API in just minutes and explains the benefits of CLIs versus APIs and MCPs, such as reduced token usage and better integration in agent workflows. He also demonstrates how to use pre-built CLIs from the library, such as ESPN for retrieving NBA games, and how to create custom CLIs, for example for Hacker News. The presenter emphasizes that CLIs are the best option for agents, followed by APIs and then MCPs. He also shows how to share created CLIs with your team by hosting them in a private GitHub repository. Finally, he stresses the importance of CLIs for efficient agent usage and the ability to convert almost anything into a CLI.

  The video explicitly covers the **Printing Press** tool and is designed for **Intermediate** users who already have experience with APIs and agent workflows.

• Overwhelmed By AI? Just Copy My Tech Stack
  5/8/2026, 01:38:26

  The YouTuber presents his personal selection of AI tools that he uses daily, weekly, or occasionally, and shares his thoughts on how to manage rapid development and the abundance of new tools without feeling overwhelmed. He begins with his “daily drivers” (S-Tier), including CloudCode (his main tool), VS Code (as an IDE for CloudCode), and Glydo (for speech-to-text transcription). In the A-Tier, he mentions tools he uses weekly, such as Codex (another agent tool), Claude (for chat-based tasks), Hermes Agent (for general knowledge work), Perplexity (for research), and Groq (for specific search tasks). He emphasizes that his core stack consists of these few tools, although there are many other tools he uses occasionally or for specific tasks. The YouTuber also discusses his “specialists” (B-Tier), which he uses for specific tasks like Appify, GBT Image 2, Nano Banana 2, Key.ai, and HeyGen. In the C-Tier, he mentions tools he uses experimentally, such as Gemini, Anti-Gravity, Ollama, and Manifold. He also has a “graduated” category for tools he no longer uses, such as ChatGPT, Open Claw, Cursor, Notebook LM, Poppy AI, Anytten, and WhisperFlow. The YouTuber shares his thoughts on efficiency and productivity, emphasizing the importance of focusing on essentials and not getting distracted by new tools. He advises viewing tools as “harnesses” that can be integrated into existing projects and emphasizes the importance of focusing on your own goals (North Star) and only using tools that truly add value.

  The video explicitly covers multiple AI tools and models, including CloudCode, VS Code, Glydo, Codex, Claude, Hermes Agent, Perplexity, Groq, Appify, GBT Image 2, Nano Banana 2, Key.ai, HeyGen, Eleven Labs, Cloud Design, Gemini, Anti-Gravity, Ollama, Manifold, ChatGPT, Open Claw, Cursor, Notebook LM, Poppy AI, Anytten, and WhisperFlow. The video is aimed more at Intermediate and Advanced users who already have experience with AI tools and want to optimize their workflows.

• Claude Just Solved Session Limits
  5/7/2026, 01:33:43

  # Summary

  Anthropic announced a partnership with SpaceX that significantly increases its compute capacity and delivers immediate improvements to usage limits: the 5-hour rate limits for Claude Code are doubled (across all subscription tiers), throttling during peak hours for Pro and Max accounts is eliminated, and API rate limits for Opus are substantially increased (input tokens per minute increased from 30k to approximately 35k, output tokens per minute from 8k to 80k). These changes address months of ongoing congestion issues that have frustrated users when hitting limits.

  The SpaceX partnership brings 300 megawatts of capacity and over 220,000 Nvidia GPUs. Anthropic has also made agreements with Amazon, Google, Broadcom, Microsoft, Nvidia, and Fluid Stack. A long-term plan is emerging: Anthropic and SpaceX are interested in developing multiple-gigawatt orbital AI compute capacity (GPUs in space), because terrestrial compute resources are ultimately limited by power consumption, water needs, and cooling infrastructure.

  For developers, this means: workflows that previously failed at rate limits are worth revisiting; Opus can now be used more frequently instead of falling back to Haiku/Sonnet; the 1-million-token context window becomes practical in production environments; Claude Code can be used for regularly executed workflows instead of just prototypes; multi-agent workflows with multiple parallel agents become feasible.

  Claude Code appears to be Anthropic’s flagship product, which is why it received priority in these announcements over API features. In the longer term, Anthropic is signaling that compute is central for the next 5+ years and is also investing in trust with local communities around their data centers.

  **Explicitly covered: Claude, Anthropic, SpaceX, Opus, Haiku, Sonnet, and vendor partnerships (Amazon, Google, Microsoft, Nvidia); Format: Opinion/Reflection with news update elements.**

• Master 97% of Codex in 1 Hour (full course)
  5/6/2026, 01:21:13

  # Summary: Codeex – Complete Tutorial from Beginner to Automation

  Codeex is a desktop app that combines ChatGPT models with local file access and agent functionality – similar in structure to Claude Code but using OpenAI models instead of Claude. The main difference from web ChatGPT: Codeex can not only chat but also read/write files locally, edit Excel sheets, execute automations, and build websites/apps.

  **Interface & Core Concepts:**
  The layout has projects and chats on the left side like Claude Code. You can select models (GPT-5.5, GPT-5.4) and adjust speed and intelligence levels (Low to Extra High). A “pet” indicator at the bottom shows the current status. An agents.md file (like claw.md in Claude Code) serves as an onboarding document for each chat.

  **Practical Project: YouTube Analytics Dashboard**
  The creator builds a complete system from scratch:

  1. **YouTube Connection:** Rather than using pre-built plugins, he uses Plan Mode to ask Codeex how to pull YouTube comments. Codeex suggests the Google Cloud API, explains the steps, then executes the configuration (env.local file, API keys).

  2. **Creating Skills:** After successfully analyzing data (200 comments in Excel with patterns, tools mentioned, content ideas), Codeex automatically converts it into a reusable skill – a markdown file with instructions that can be accessed anytime (via /slash command).

  3. **Building the Dashboard:** With GPT-Image-2 integration, Codeex generates UI concepts, then builds a React dashboard on localhost with charts, insights, links to comments – with automatic visual review loops.

  4. **Deployment:** GitHub repository + Vercel – Codeex pushes code to GitHub, Vercel deploys automatically, every change goes live.

  5. **Automation (Scheduling):** Weekly routine every Sunday at 5 PM – Codeex executes the skill, updates Excel, pushes to GitHub, Vercel deploys anew – all without user intervention.

  **Browser Use & QA:** Codeex can interactively test the UI – clicks buttons, hunts for bugs, documents issues (e.g., external YouTube links not opening properly).

  **Important Mindset Shifts:**
  – Use Plan Mode first to discuss steps before execution
  – Treat errors as “golden knowledge” – store in agents.md so they aren’t repeated
  – Skills and automations aren’t perfect on the first try; improve iteratively
  – Keep context window in mind (shown at the bottom), consider model selection (Extra High costs more tokens)
  – Permissions: Default vs. Full Access depending on trust level
  – Not all tools need native plugins – APIs and curl requests work too

  **Integration with Other Tools:**
  Local project structure enables mix and match: Claude Code can work in the same folder structure, Cloud Code as well – just rename agents.md vs. claw.md and adjust if needed. Codeex and Cloud Code have different strengths: Claude better for brainstorming/planning, Codeex more pragmatic when executing and troubleshooting.

  The entire project is ultimately just a local folder – portable, version-controlled, extendable.

  —

  Explicitly covered: Codeex, GPT-5.5/5.4, Claude/Cloud Code (comparison), Google Cloud API, GitHub, Vercel, GPT-Image-2, Browser Use – **Tutorial & Demo (for beginners to advanced users)**.

• Higgsfield Just Turned Claude Into a Creative Agency
  5/5/2026, 03:05:58

  # Summary: Using Claude as a Creative Agency

  The creator shows how to connect Claude with Higgsfield (an AI platform for video and image generation) to build an automated creative agency.

  **Basic Setup:**
  Higgsfield is connected to Claude Web via MCP (Model Context Protocol). In Settings under “Connectors,” a Custom Connector is added and authenticated via OAuth. This allows Claude to directly access Higgsfield functions.

  **Practical Application in Claude Web:**
  A simple prompt (“Build me a headphone brand from the ground up with branding, product catalog, and generated assets”) automatically generates product photos, Instagram ads, and UGC videos. Claude orchestrates the generation and summarizes results. Iterations work through commands like “Make it faster, more energetic, with jump cuts and slow motion” — Claude then reformulates more precise prompts for Higgsfield.

  **Scaling with Claude Code (Desktop App):**
  In the desktop client, CLI commands are executed to install Higgsfield and Google Workspace Services. This enables advanced automations: Claude pulls all generated assets from Higgsfield, creates a Google Sheet database with 45+ generations, analyzes them by product/style, and plans new variations.

  **Key Concepts:**
  – **Research Docs:** A markdown file “Advertising Masterclass” with 2026 best practices is integrated into the project; Claude uses this as expert knowledge during ideation.
  – **Skills:** Recipes for consistent outputs. The creator reverse-engineers a skill from his favorite generations — for example, a “Hypermotion Video” skill that produces future videos in the same style.
  – **Status Tracking:** The Google Sheet gets a status column; Claude marks completed generations, job IDs, and URLs.
  – **Product Consistency:** Reference images must be uploaded daily to the project folder so Claude uses them during generation (otherwise the product looks variable).

  **Automation via Routines:**
  Claude Routines could, for example, plan 50 new variations on Sundays and then generate 30 of them on Mondays — while the user sleeps. Scaling works through different weekly routines for planning and production; outputs can later be piped to Meta Ads Manager.

  **Practical Learnings:**
  – MCP costs more tokens than CLI, so CLI is more efficient for agents.
  – Some generations get rejected for “sensitive content”; Claude can debug and correct prompts.
  – Text accuracy in videos (e.g., on bottles) is currently a weak point; workarounds include minimizing text or alternative label designs.
  – The quality of AI output depends heavily on prompting precision; therefore external expertise (research docs, skills) is important.

  **Workflows Visible in Video:**
  1. Brand brainstorm + asset generation (simple)
  2. Create Google Sheet database with all assets (moderately complex)
  3. Use Higgsfield Marketing Studio for launch videos (demo shows hypermotion videos)
  4. Plan bulk variations based on ad research (moderately complex)
  5. Build skills and set up routines for autonomous production (advanced)

  Claude, Higgsfield, and Google Workspace CLI are presented as an ecosystem to accelerate content production a hundredfold — from ideation to automated overnight generation. — **Demo + Tutorial with focus on Claude + Higgsfield integration, Google Workspace also mentioned; Intermediate to Advanced (routine setup requires familiarity with Claude Code).**

• Building Realistic Voice Agents Has Never Been Easier
  5/4/2026, 12:46:03

  # Summary

  The creator shows how to build a voice agent in about 15 minutes with Claude Code and Eleven Labs – without manual dashboard configuration.

  **The Concept:** A voice agent is a loop of speech input → transcription → LLM processing → optional database query or tool call → speech output. Each voice agent consists of four components: Persona (system prompt), Voice (selectable from various voices, including custom voice clones), Knowledge (business information/database access), and Tools (API calls, MCP servers, external automations).

  **Live Demo of the Workflow:** The creator builds a sales agent for an AI consultancy that pulls prospects from the website. He uses Claude Code in Plan Mode, which strategically asks questions (Cal.com API access? Desired persona? Data fields to capture?). Cloud Code then creates an architecture plan, which the creator accepts. After that, only API keys (Cal.com + Eleven Labs) are entered into an .env file, and Cloud Code handles the rest: agent configuration in Eleven Labs, tool setup (Check Availability + Book Call), widget integration into the HTML.

  **Iterations & Debugging:** On first test, issues emerge: voice too enthusiastic, first message not transmitted, tool finding incorrect availability (UTC instead of Central Time). Rather than reading documentation, the creator simply explains the problem; Cloud Code debugs it by analyzing the transcript and finds the error in the tool parameter. After re-prompting, availability and booking work correctly.

  **Security & Costs:** Since the widget runs on the website and the website owner bears the Eleven Labs costs, domain whitelisting, conversation length limits, rate limits, and strong knowledge grounding (real docs, not hallucinations) should be configured.

  **Deployment Options:** The agent can be deployed on a website (widget), in the Eleven Labs dashboard (testing), or via Twilio integration (phone call) – same engine, different interfaces.

  **Demo Result:** A functioning sales agent that automatically validates names/emails, queries availability (with Cal.com limits), and books meetings, completely built through natural language.

  —

  Claude Code + Eleven Labs; demo with live build of a sales voice agent.

• I Tried 100+ Claude Code Skills. These 6 Are The Best
  5/3/2026, 13:42:51

  # The Six Skills That Actually Make Money in AI Automation

  The content focuses on six practical skills for Cloud Code that aren’t designed for YouTube effect but solve real business problems: save time, cut costs, prevent errors.

  **1. Skill Creator** (official from Anthropic): A meta-skill that automatically generates other skills. Instead of manually writing markdown files, you describe the requirement in English, Claude drafts, tests, and packages the skill. Solves the problem that beginners fail at manual writing and build faulty skills.

  **2. Superpowers**: Forces Claude into a senior developer workflow – plan first, then code in an isolated environment, tests before code, self-review in two stages (spec match and code quality). Combats the main problem: hastily written code that breaks in production.

  **3. GSD** (Get Stuff Done): Fixes context decay issues. After ~30 minutes, sessions become sloppy, Claude forgets requirements. GSD spawns fresh sub-agents per task with clean context, has quality gates, and optional autonomous mode. Costs token excess but saves hours of rework.

  **4. /review and /ultra review**: Already built into Cloud Code. `/review` does structural code reviews locally, `/ultra review` (since Opus 4.7) spins up a fleet of reviewer agents in parallel in sandbox, each finds bugs independently. Requires Cloud Code 2.1.86+, costs 5–20 dollars after free attempts, runs in the background.

  **5. Context Mode**: Filters garbage data from tool calls (56 KB Playwright snapshot → 299 bytes). Works with SQLite database that tracks every event; when compacting context, a snapshot is injected instead of forgetting everything. Sessions run 3 hours instead of 30 minutes.

  **6. ClaudeMem**: Carries knowledge across sessions. Hooks into Cloud Code lifecycle, captures edits/decisions/bug fixes, compresses into local SQLite with vector search. New sessions automatically get relevant context. 10x token savings on retrieval vs. dumping everything. Auto-generates folder-level Claude.md files.

  **Bonus – Skill #7**: Official Front-End Design Skill from Anthropic, for less AI-generated look in design/slides.

  **Sales Approach**: Don’t sell the workflows, sell the outcomes – save 10 hours/week, reduce admin errors, increase lead velocity, maximize profit. Beginners should master one skill, build a few demos, show them – business owner sees value, not resume.

  **Installation**: All `/plugin install` commands are in the video description.

  —

  **Explicitly mentioned tools/models**: Claude (Anthropic), Cloud Code, Opus 4.7, Playwright; **Format**: Deep-dive with practical examples from real projects (real estate, HVAC, marketing agencies).

NeuralNine (2 new videos)

• OpenClaw: Simple VPS Setup Guide
  8.5.2026, 16:00:03

  The video shows a step-by-step guide to setting up Open Claw on a virtual private server (VPS) with Ubuntu. The focus is on speed and simplicity, with installation and connection to Telegram at the center. The process includes updating the system, installing the Node Version Manager (NVM) and Node.js, installing Open Claw via npm, and completing the onboarding process. OpenAI is selected as the model provider and Telegram as the communication channel. The user creates a bot through the Bot Father in Telegram and connects it with Open Claw. After setup, the bot’s functionality is demonstrated by creating and editing a shopping list. The video warns about the security risks of Open Claw and recommends running it in an isolated environment.

  The video explicitly covers Open Claw, OpenAI, Telegram, and NVM, and is aimed more at intermediate users.

• Coding Slack Bots in Python: Quick Start Guide
  4.5.2026, 16:00:21

  The video shows how to create a simple Slack bot with Python that can respond to messages and execute commands. The focus is on quickly implementing a Minimum Viable Product (MVP). Here are the main steps:

  1. **Create Slack App**:
  – Log in to Slack and create a new app via the Slack API.
  – Enable Socket Mode and assign the necessary permissions.
  – Generate and save the Socket token in a `.env` file.

  2. **Create Bot Token**:
  – Add bot token scopes for chat functionality.
  – Enable Event Subscriptions to respond to messages.
  – Create a Slash Command (e.g., `/add` for simple calculations).
  – Install the app in your Slack Workspace and save the bot token in the `.env` file.

  3. **Set up Python Project**:
  – Install required packages (`python-dotenv` and `Slack-Bolt`).
  – Set up the `.env` file with the Slack tokens.

  4. **Implement Bot Functionality**:
  – Simple response to the message “hello” with the username.
  – Implementation of the `/add` Slash Command for simple addition.
  – Response to general messages with a standard reply.

  5. **Connect Bot to a Language Model**:
  – Install the `openai` package.
  – Integrate GPT-4O to respond to bot mentions and generate answers.
  – Update Event Subscriptions for the mention feature.

  The video ends with a demonstration of how the bot responds to mentions and answers questions, including generating Python code.

  **Final Comment**: The video explicitly covers the use of OpenAI (GPT-4O) and is aimed more at intermediate users.

Nic Conley

No new videos during this period.

Nick Saraev

No new videos in this period.

Niklas Steenfatt

No new videos in this period.

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

No new videos in this period.

Productive Dude

No new videos in this period.

Sebastien Dubois

No new videos in this period.

Tech With Tim (4 new videos)

• One AI Agent Isn’t Enough Anymore
  9.5.2026, 16:00:42

  # Orchestrating specialized AI agents for software development

  This video shows how to build multiple specialized AI agents instead of a single generic all-in-one agent and run them in parallel from the terminal – using Mistral Vibe as an example, though the concept works with all popular coding tools.

  **The problem with generic agents:** A single agent that does everything quickly develops context issues (Context Illusion). The more information it needs to juggle – architecture, tests, configs, refactoring – the worse its performance becomes. Eventually it approaches the context limit and forgets things or makes inconsistent decisions, often requiring a full context restart.

  **The solution:** Specialized sub-agents that focus on individual tasks – one for tests, one for code reviews, one for deployment preparation, etc., just like different team members. Each sub-agent inherits the complete project context (file structure, Git status, code), but not the previous conversation history and tool calls. This means: it starts with 10–20% context utilization instead of 80–90%.

  **Mistral Vibe setup:** Installation via a simple terminal command, then type `vibe`. The tool comes with Devstral 2 (seven times cheaper than Claude Sonnet, similar performance, open-source), and agent TOML files land in the `.vibe/agents` folder. You can create agents manually or ask the tool to do it.

  **Agent types:** Main Agents (for a specific role in the main session) or Sub Agents (independent background processes). Sub Agents can run in parallel, don’t store session logs in memory, and automatically receive project context without conversation history.

  **Permissions & safety:** When configuring, use permission scoping – for example, give the test writer only Bash, read, and write commands, not Git or network access. You can also enable `auto_approve` so agents don’t constantly need confirmations, set a maximum budget, and limit the number of turns.

  **Practical example:** The creator builds a simple FastAPI backend and HTML/JS frontend, then creates three sub agents:
  1. **Test Writer** – writes backend tests with PyTest
  2. **Code Reviewer** – reads code, checks security and performance, makes no changes
  3. **Deploy Prep** – runs complete test suite, linter and code review, confirms deployment readiness

  All three are then executed in parallel in the background, significantly speeding up the workflow.

  **Team advantage:** Agents can be configured at the project level in `.vibe/agents` and committed to the Git repo – the team then shares the optimized agent definitions.

  Mistral Vibe (Devstral 2 model), demo.

• Codex is INSANE – Everything New in 10 Minutes
  8.5.2026, 16:25:45

  **Summary: GPT 5.5 – Capabilities and Applications**

  GPT 5.5 is OpenAI’s latest frontier model and is described as the best available model currently – significantly better than its predecessor 5.4 and Claude’s Opus 4.7 model, particularly in practical scenarios like programming, data analysis, and spreadsheets. It costs about twice as much as the previous model, but through a Pro subscription (200 USD/month) is still cheaper than top models from Anthropic. The model can be accessed via ChatGPT, IDEs, browser extensions, the Codex desktop application, or the API.

  The key new capabilities are:

  **Coding and browser automation:** GPT 5.5 can write extensive code and automatically test the application in the browser in Codex – it independently takes control, clicks through the interface, and validates functionality without requiring manual input.

  **Computer use:** After enabling in Codex plugins, the model can take over your entire desktop. It can open applications, perform actions (like starting Spotify and playing music), while simultaneously allowing your own computer use since it has a separate cursor.

  **Data, tables, and presentations:** GPT 5.5 creates Excel spreadsheets with extensive data (including research, color coding, charts) and can automatically generate multi-page PowerPoint presentations from them – a significant improvement over the previous version.

  **Additional features in Codex:** The tool offers plugins for connecting with other applications, automated workflows, MCP server integration, and GitHub connections. Web application deployment is free via here.now (24 hours or permanent with a free account).

  The video presents **OpenAI GPT 5.5 in the Codex application** – demo format with practical examples.

• How to Build an App With Claude Code – Full Tutorial for Beginners
  6.5.2026, 13:17:28

  # Summary: Building and deploying a web application with Claude Code

  This video is a comprehensive tutorial on creating and launching a complete web application with Claude Code – from setup to going live on your own domain.

  **Preparation and installation:**
  First, you need to install Claude Code (terminal version recommended, requires Premium subscription), a code editor (Cursor is recommended, free), and later a hosting service. In Cursor, install the extensions “Claude Code” and “Hostinger Connector.”

  **Planning before development:**
  Before generating code, create a detailed spec document (Markdown file) with Claude by describing your requirements and asking Claude questions – clarify colors, layout, content, desired tech stack. The creator uses a portfolio website example and chooses Next.js + Tailwind as the stack.

  **Building the website:**
  Once the spec is set, instruct Claude to build the website and run it locally. The AI tool generates the complete code and starts a local server. The creator shows that with just a few prompts, a functional, attractive website is created – preferably in small steps (MVP = Minimum Viable Product).

  **Version control:**
  Git is set up to save changes and enable rollbacks. Claude automatically creates commits.

  **Deployment via Hostinger:**
  A Hostinger account is created (starting at about 4 EUR/month), a domain selected, and registered for free. The compressed project file is uploaded to the Hostinger dashboard and deployed. An API token is generated in Hostinger and inserted into the Hostinger extension in Cursor. After that, changes can be pushed directly from Claude Code via natural language to Hostinger – without manual re-uploading. The creator shows how to rename the website and redeploy with just a text command in Claude Code.

  **Result:**
  The website is live under a real domain and can be updated anytime from Claude Code.

  Claude Code, Next.js, Cursor, and Hostinger were explicitly covered; **tutorial** with practical beginner focus.

• Claude Code + Nano Banana 2 = This Changes Everything
  3.5.2026, 18:18:44

  # Summary

  The video shows how to integrate image generation with Nano Banana Pro directly into Claude Code instead of visiting websites manually. The process works in three steps: First, create an API key at Google AI Studio and activate billing (with recommended spending limit). Then install two Skills – one converts normal text prompts into structured JSON schemas for more detailed image generation, the other calls the Nano Banana API. With this combination, Claude Code can automatically generate multiple images in parallel and embed them directly in websites. The video demonstrates concrete examples: a fully generated perfume comparison portfolio, then an image of the creator with Sprite and Mac Mini, and finally a reference image editing example where a Claude logo is replaced with a ChatGPT logo. The JSON structure enables very precise control over details like framing, lighting, and materials – details no one would write manually.

  **Explicit mentions:** Claude Code, Nano Banana Pro (image model), Google Gemini API, Skyworks 3.0 (mentioned as sponsor), Whisper Flow (dictation), ChatGPT. — **Format:** Tutorial with live demo.

TheAIGRID (7 new videos)

• How To Use Claude For Microsoft Word (Microsoft Word Claude Tutorial)
  9.5.2026, 20:15:01

  The video demonstrates how to use Claude in Microsoft Word. First, installation of the add-in is explained, with a note that a Claude Pro or Max plan is required. Users are encouraged to use the Sonnet 4.6 or Opus 4.6 models, with Opus 4.7 reserved for mathematical tasks. Important settings like “Working with Files” are highlighted, which enable sharing context across different Office documents.

  The video shows how Claude can be used in Word to write, edit, and format text. It demonstrates how to highlight specific sections and then expand or rephrase them. Additionally, Claude’s ability to reformulate and structure text in various styles is emphasized, including adding headings and bullet points.

  Another focus is the function to mark and analyze specific sections of a document, for example to highlight security aspects or important points. The video also demonstrates the integration of web research directly in Word, noting that this feature is suitable only for simple search queries.

  The video shows how Claude can collaborate with other Office applications like Excel and PowerPoint to integrate data from these applications into Word documents. This is demonstrated with an example where sales data from Excel is incorporated into a shareholder letter.

  Finally, the use of templates in Word is shown, for example for resumes, with a note that Claude may sometimes have difficulty with complex formatting. It’s recommended to use the “Undo” function if problems occur or to repeat the task with a higher-level model.

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

• OpenAI Is Losing The AI War
  8.5.2026, 22:34:30

  The video analyzes the current state of competition among leading AI companies, particularly between OpenAI and Anthropic. It notes that Anthropic has significantly gained market share in recent months, both among private users and in the enterprise sector. Anthropic has recorded unprecedented revenue growth and has surpassed OpenAI in many areas, particularly in coding and general tasks. The frequent product releases and strong market position of Anthropic, especially with models like Claude Opus and Mythos, are highlighted. Additionally, Anthropic’s resistance to certain government requests is presented as a positive PR moment that has strengthened trust in the company. Investors are increasingly skeptical of OpenAI, while Anthropic sees high demand on the secondary market. The video concludes with the observation that user preferences have clearly shifted toward Anthropic, raising the question of whether OpenAI can still catch up.

  The video explicitly covers Anthropic, OpenAI, Gemini, and specific models like Claude Opus and Mythos. It is aimed more at intermediate and advanced users who want to familiarize themselves with current developments and market trends in the AI industry.

• Claude For Powerpoint Tutorial – How To Use Claude With Powerpoint
  7.5.2026, 21:00:59

  The video demonstrates the use of “Claude for PowerPoint,” an official add-on from Anthropic that is integrated directly into Microsoft PowerPoint. It enables the generation of fully editable PowerPoint elements based on existing templates or new content. The add-on is available in various pricing tiers but is still in beta/research phase.

  To use Claude for PowerPoint, you need a Claude Pro account (from $20/month) or a Max Team/Enterprise subscription, as well as a current version of Microsoft PowerPoint (desktop or web). After installation via the Add-in feature in PowerPoint, you can choose between two models: Opus 4.6 for complex tasks and Sonnet 4.6 for quick edits. The settings allow you to set standards such as fonts, colors, and notes for each slide. Important is the option to confirm changes in advance to avoid unwanted edits.

  Claude automatically reads the template layouts, fonts, and color schemes and generates slides that follow the visual structure of the template. An important tip is to load the template before prompting, as Claude uses it as a reference. The video shows examples of creating presentations from text prompts, editing individual slides, converting PDFs and Excel data into slides, and generating presentations based on websites. Claude can also translate presentations into other languages and add speaker notes.

  Some limitations are mentioned, such as the 30 MB file size limit, potential issues with complex layouts, and limited graphic analysis. The add-on is currently only available for desktop and web, not for iPad or Android.

  The video explicitly covers the AI tool “Claude for PowerPoint” from Anthropic and is more suitable for intermediate users who are already familiar with PowerPoint and want to leverage AI tool functionalities.

• Why AI Ceos Are Now Afraid Of AI
  6.5.2026, 21:15:05

  The video discusses the ambivalence and fears of the richest and most powerful men in the world who are working on artificial general intelligence (AGI). They view AGI both as humanity’s greatest opportunity and as a potential end to civilization. A central problem is the alignment problem: How can one ensure that a superintelligent AI understands and follows human values and unspoken assumptions? So far, there is no solution to this. Additionally, there is growing fear of an arms race in AI, since whoever develops AGI first gains a dominant position. This could destabilize existing economic and political structures. Another critical point is recursive self-improvement, in which an AI could exponentially increase its own intelligence, leading to a sudden, difficult-to-control intelligence leap. Finally, the danger is emphasized that AGI could be deployed as a weapon, leading to autonomous cyberattacks, biotechnological threats, and massive disinformation. Current developments are occurring without international regulation and controls, further increasing the risks.

  The video explicitly covers OpenAI, Anthropic, Safe Superintelligence Inc., and models like GPT-6 as well as Grok, and is aimed more at intermediate and advanced users.

• How To Use Pomelli – Google Pomelli Tutorial –
  5.5.2026, 21:15:01

  The video introduces “Meli,” a new AI-powered marketing tool from Google Labs that generates social media and advertising content for businesses based on a website. The focus is on defining the “Business DNA,” which consists of Brand Aesthetic, Brand Tone of Voice, Business Overview, Brand Values, and Tagline. The creator demonstrates how to create these values with the help of AI tools like Gemini or ChatGPT by using screenshots of templates and having AI fill them in. Subsequently, it is shown how to set logos, colors, and fonts and create campaigns that showcase the product in various contexts. The creation of videos and photo shoots is also explained, with AI generating various variations that can be customized and downloaded. The tool is particularly useful for ideation and creation of marketing materials.

  The video explicitly covers Google’s Meli and uses Gemini and ChatGPT as supporting AI tools. It is more suitable for intermediate users who already have basic knowledge in marketing and branding.

• AI Helped Spark a Quantum Breakthrough. “The World ‘Is Not Prepared’
  4.5.2026, 21:22:09

  The video discusses recent advances in quantum computing research that have been accelerated by AI and the potential threats to modern internet security. Three main factors are highlighted: improvements in quantum computers, more efficient algorithms, and AI-driven discovery of these algorithms. This combination could significantly reduce the size of quantum computers needed to break modern encryption.

  Google has published research results showing that a future quantum computer with fewer than 1,200 logical qubits could attack certain encryption methods. Additionally, researchers from Caltech and Qatomic have argued that Shor’s algorithm could run with only 10,000 reconfigurable atomic qubits on cryptographically relevant scales. AI played a crucial role in developing these algorithms by searching through thousands of possibilities and improving algorithm efficiency.

  Companies like Cloudflare are responding to these developments by advancing their goals for transitioning to quantum-safe security. Cloudflare aims to be fully quantum-safe by 2029, including authentication. However, transitioning to quantum-safe cryptography is complex and requires updating multiple systems and rotating secrets.

  Final comment: The video explicitly covers OpenAI and open-source tools like Open Evolve and is aimed more at intermediate and advanced users.

• Hermes Agent Setup With Use Cases – Hermes Agent Use Cases
  3.5.2026, 21:30:05

  The video shows how to set up and use Hermes Agent on an affordable cloud GPU platform (hpcai.com). The process includes setting up a CPU-based instance for 24 cents per hour, installing Hermes Agent via a single command, and connecting to an inference provider like News Portal (with a $20 monthly fee for easier setup). Hermes Agent offers various features such as web scraping, cron job creation, image generation, lead generation, and price monitoring. Exemplary use cases include scraping YouTube channels, creating weekly reports, generating images for social media, finding leads for a business, and monitoring supercar prices. Hermes Agent can also be used for creating content ideas and automating tasks like Instagram DMs. The video emphasizes the versatility and efficiency of Hermes Agent for various applications.

  The video explicitly covers Hermes Agent and the cloud GPU platform hpcai.com and is aimed more at intermediate users.

Theo – t3․gg (4 new videos)

• Anthropic just…wait what
  7.5.2026, 09:08:36

  The video discusses the current partnership between Anthropic and SpaceX (or Elon Musk’s XAI) and its associated strategic implications for the AI industry. Anthropic, a leading AI provider, faces a massive compute problem as demand for its models, particularly Claude, far exceeds available computing resources. This scarcity has led to restrictions on usage limits and price adjustments, which however are not primarily aimed at profit maximization but rather at making efficient use of limited compute resources. The partnership with SpaceX, which has substantial compute capacity, is intended to close this gap. At the same time, the strategic importance of Cursor, an AI-powered coding tool, is highlighted for possessing valuable data essential for training AI models. The analysis shows that both Anthropic and XAI are attempting to compensate for their respective weaknesses (compute and data respectively) through this cooperation, while OpenAI remains as the main competitor in the background. The video analysis is detailed and suitable for intermediate to advanced users, as it provides in-depth insights into the strategic decisions and technical infrastructure of the AI industry. Topics explicitly addressed include Anthropic, SpaceX/XAI, Cursor, OpenAI, as well as specific compute infrastructures such as Nvidia GPUs and various cloud providers.

• Get In, We’re Leaving GitHub
  6.5.2026, 19:43:52

  The video discusses current challenges and alternatives to GitHub, particularly in light of GitHub’s increasing reliability issues such as random merge reversions and extended downtime. The author emphasizes the need to seek alternatives and evaluates various options including GitLab, Bitbucket, GitTea, Forgejo (Codeberg), and new approaches such as Pierre, Graphite, and Entire.

  **Key Points:**

  1. **GitLab**:
  – Often mentioned as an alternative, but has significant UX problems and is less user-friendly than GitHub.
  – The codebase is large and complex, making maintenance and improvements difficult.
  – GitLab is more of an enterprise solution focused on CI/CD and integration, but not necessarily a direct improvement over GitHub.

  2. **Bitbucket**:
  – Primarily promoted as a more cost-effective alternative for enterprises already using Atlassian tools.
  – Integrations with Jira and other Atlassian products are strong, but user experience and functionality are not comparable to GitHub.

  3. **Forgejo (Codeberg)**:
  – An open-source alternative that emerged from a fork of GitTea.
  – Offers a simple, self-hosted solution with good performance and transparency.
  – The author is impressed by the user-friendliness and community behind Forgejo and even donates to the project.

  4. **Pierre**:
  – A new approach that lays the foundation for a next generation of Git hosting solutions.
  – Focus on high throughput capability and integration of agents that generate large amounts of code.
  – Pierre has already achieved impressive performance metrics and could form the basis for future GitHub alternatives.

  5. **Graphite**:
  – Offers improved code review workflows and was recently acquired by Cursor.
  – Potential to create an entirely new type of developer platform that goes beyond traditional GitHub functionality.

  6. **Entire**:
  – Founded by former GitHub CEO Thomas Dohmke.
  – Develops tools to track and improve agent context when generating code.
  – Invests in the future of development, particularly in connection with AI agents.

  **Final Note**:
  The video explicitly discusses open-source tools such as Forgejo (Codeberg), Pierre, and specific companies like Graphite and Entire. It is designed more for intermediate and advanced users who want to engage with the technical details and strategic considerations of Git hosting solutions.

• Prime is (mostly) right about AI
  5.5.2026, 08:53:45

  The video discusses the changing economics of the AI industry, particularly regarding the use of AI models like Claude and GitHub Copilot. The speaker responds to a video from Primagen that analyzes current shifts in AI economics.

  The speaker emphasizes that recent changes in pricing models from Anthropic and Microsoft are not aimed at earning more money from end users, but rather at managing limited compute capacity (GPUs). For example, Anthropic has attempted to limit Claude Code usage in cheaper tiers to free up computing resources for enterprise customers. Microsoft has also adjusted its pricing models for GitHub Copilot, which the speaker attributes to limited availability of computing power.

  Another important point is the efficiency improvement of AI models. Despite higher token costs per request, newer models like GPT-55 are more efficient and cheaper to use, particularly in mid-range and lower-tier settings. The speaker argues that costs for actual work tasks are decreasing when considering model efficiency.

  The speaker also criticizes the assumption that Google subsidizes less. He argues that Google has actually subsidized very aggressively, but due to poor quality of its models and rapid adjustments to terms of service to avoid overload.

  Overall, the speaker agrees with Primagen’s analysis that AI industry economics are changing, but emphasizes that the main cause is limited compute capacity rather than corporate greed.

  **AI Tools/Models/Providers:** Claude (Anthropic), GitHub Copilot (Microsoft), GPT-55 (OpenAI), Google (Gemini)
  **Target Audience:** Intermediate

• Microsoft and OpenAI break up (Amazon is pumped)
  4.5.2026, 09:12:17

  # Summary: The Breakup Between Microsoft and OpenAI

  The partnership between Microsoft and OpenAI, which began in 2019 with an initial billion-dollar investment, is undergoing a fundamental restructuring. Originally, the agreement was designed so that Microsoft remained OpenAI’s sole cloud provider and could license all OpenAI developments until reaching AGI. The turning point came in September 2024 when OpenAI introduced the O1 reasoning models – a massive breakthrough that dramatically increased model intelligence. Microsoft then aggressively demanded access to details of this breakthrough technology, which OpenAI refused. This led to tensions that were reflected in internal Microsoft meetings where executives were frustrated that OpenAI wasn’t sharing their research quickly enough.

  In 2026, OpenAI then announced an “amended agreement” that fundamentally loosened the partnership: Microsoft remains the primary cloud provider, but OpenAI can now deploy all models on any cloud provider (not just Azure). Microsoft’s exclusivity is effectively over. OpenAI is now also closing deals with AWS and Google to make their models available there – particularly via AWS Bedrock, where Anthropic previously held a de-facto monopoly. This is the real heart of the story: Anthropic had been positioned, thanks to its tight Bedrock integration, to win enterprise customers faster than OpenAI because many companies were already on AWS. Through the new AWS partnership and distribution across multiple clouds, OpenAI can now directly compete with Anthropic and eliminate the advantage that cloud exclusivity had given Anthropic.

  The video author also sharply criticizes Microsoft’s Azure infrastructure for OpenAI models, which had significant performance problems (sometimes 2–15x slower than OpenAI endpoints), and documents how he drove Microsoft to quickly fix these issues through public criticism and benchmarking. In the end, the video describes an industry in upheaval: the real competition will increasingly not be between AI models themselves, but between custom chips and cloud providers (Nvidia vs. AMD vs. Trainium vs. TPUs).

  **Explicitly discussed tools/providers:** Microsoft, OpenAI, Anthropic, AWS (Bedrock), Google Cloud, Azure, DeepSeek, Trainium chips — **Format:** Opinion/deep-dive with strong personal perspective from the creator.

Tim Carambat (1 new video)

• A New AI Model Just Dropped With A CRAZY Claim.
  5.5.2026, 19:02:32

  The video discusses the announcement of a new model called SubQ from Sub Quadratic, which boasts a 12-million-token context window and allegedly 52 times higher efficiency compared to existing models. Creator Timothy Karen expresses skepticism due to missing technical reports and unclear benchmarks. He explains the underlying technology based on “sparse attention” and compares it with conventional and Flash Attention mechanisms. The benchmarks show the model performs well in certain tests like SWEBench verified and MRCRV2, but it’s unclear whether these results apply to the 12-million-token model or a 1-million-token preview. Timothy Karen has applied for early access and will share the results in a future video.

  The video specifically covers the SubQ model from Sub Quadratic and is aimed more at intermediate or advanced viewers.

Unsupervised Learning

No new videos in this period.

WorldofAI (7 new videos)

• Codex Super App, OpenAI Chaos Drama, Gemini 3.2 Pro In Arena, GPT-Realtime-2, & NotebookLM Update!
  9.5.2026, 07:19:53

  The video provides an overview of the week’s major AI developments. OpenAI announced hints at a future Codex Super App featuring capabilities like remote control, deeper integrations, and new connectors. GPT Real-Time 2 was also unveiled, a voice model with near GPT-5 intelligence for real-time interactions. Google is experimenting with new Gemini checkpoints, though they’re perceived as less capable. Claude Code introduced a financial data interface enabling advanced financial analysis and trading strategies. In China, Baidu released Ernie 5.1, which achieves better benchmark results than DeepSeek V4 at lower costs. Xi (likely a reference to XAI) is expanding Grock into a Super App with enhanced tool-calling capabilities and productivity tools. Additionally, private text messages between Sam Altman and the OpenAI board were leaked, providing new insights into power struggles at OpenAI. The video concludes by discussing the growing cultural rejection of physically present AI systems, the “anti-cyborg” movement.

  The video explicitly covers OpenAI, Google Gemini, Claude Code, Baidu, and XAI (likely Grock) and is better suited for intermediate to advanced AI enthusiasts.

• NEW Open Claude Code Is A FULLY FREE AI Coding Agent! (Tutorial)
  8.5.2026, 05:57:40

  The video discusses current issues with Anthropic, particularly aggressive rate limits and reduced model performance affecting the Claude Code user experience. The creator points out that even with a Pro subscription, usage restrictions are frustrating and costs for additional usage quickly add up. As an alternative, Freebuff is introduced, a free AI coding agent based on GLM 5.1 that requires no subscriptions or complicated setups. Freebuff offers fast, autonomous, and user-friendly coding experience with integrated sub-agents and intelligent follow-up prompts. The creator demonstrates Freebuff’s installation and usage, including integration with ChatGPT for certain use cases. Freebuff is presented as a promising and free alternative to Claude Code that offers faster and more reliable coding experience.

  The video explicitly covers Anthropic, Claude Code, Freebuff, GLM 5.1, and ChatGPT and is intended for intermediate to advanced users.

• Claude’s New “Infinite” Context Window Model, Doubled Rate Limits, Multi-Agent Cordination, & More!
  7.5.2026, 06:44:06

  The video summarizes the key announcements from the Anthropic developer conference, focusing on AI coding agents, agent workflows, and the future of software engineering with Claude. A central theme was long-term agent intelligence, with Anthropic introducing a new “Dreaming” feature that allows agents to review past sessions and improve future decisions. Multi-agent orchestration was also introduced, where a lead AI agent can delegate tasks to specialized agents working in parallel on complex tasks. Another important point was the substantial increase in Cloud Code rate limits for all paid plans, enabled by a new compute partnership with SpaceX. Anthropic also announced three main focus areas for the next generation of models: an infinite context window, advanced multi-agent coordination, and persistent long-term reasoning systems. These developments suggest Claude could evolve from a simple chatbot into a fully autonomous software engineering system.

  The video explicitly covers Claude from Anthropic and is better suited for intermediate to advanced users.

• Gemini Omni, Gemini 3.2 Flash, a 12M Context Window Model, Claude Replaces Analysts, & More! AI NEWS
  6.5.2026, 06:30:16

  The video provides a comprehensive overview of the latest developments in the AI industry, with special focus on Google’s upcoming announcements at the Google IO conference. Key points include alleged tests of Gemini 3.2 Flash and potentially more powerful variants like Gemini 3.5 or 4.0. A breakthrough in large language model architecture from SubQ is also mentioned, introducing a model with a 12 million token context window. OpenAI released GPT 5.5 Instant, which is faster and more efficient, while Anthropic expands its AI solutions for the financial sector. Google unveiled numerous updates for its AI tools including Gemma 4, Google AI Studio, and Notebook LM. Perplexity also launched a finance agent working with licensed data from various providers.

  Final comment: The video explicitly covers Google (Gemini, Gemma), OpenAI (GPT 5.5 Instant), Anthropic (Claude), SubQ, and Perplexity and is better suited for intermediate to advanced audiences.

• Open Design – Open Source Claude Design! Fully Free AI Design System!
  5.5.2026, 06:52:13

  The video introduces Open Design, an open-source alternative to Claude Design by Anthropic, enabling creation of UI designs, wireframes, interactive prototypes, and presentations through voice commands. Open Design offers several advantages over Claude Design, including the ability to integrate different models and tools, a local-first web deployment option, and compatibility with up to 15 different coding agent CLIs. It features 31 composable skills and 72 complete design systems, enabling creation of production-level designs. The video also demonstrates how to locally install and set up Open Design, including configuring agents, media providers, and MCPs. It shows how to create a blog post design and utilize Open Design’s various features to create high-quality UI designs. The video emphasizes Open Design’s advantages over Claude Design, particularly its flexibility and ability to use open-source models.

  The video explicitly covers open-source tools and is better suited for intermediate to advanced users.

• DeepSeek V4 + Claude Code = BEST AI Coder!
  4.5.2026, 07:30:58

  The video demonstrates how to combine DeepSeek V4 with Claude Code to create a cost-effective and efficient AI coding workflow. DeepSeek V4 is a powerful open-source model that is particularly token-efficient and supports long context windows. It works well for basic coding tasks like quick scripts, unit tests, and simple automation, but isn’t suitable for complex tasks like web development or security audits. By combining with Claude Code, DeepSeek V4 can be used for simple tasks while reserving more expensive models like GPT 5.5 or Opus 4.7 for complex work. The workflow is simplified through Gravity, which performs setup autonomously. A demo shows how DeepSeek V4 is used for the basic structure of an AI dashboard, while Opus 4.7 handles UI refinement and complex tasks. This hybrid approach saves costs and bypasses rate limits.

  The video explicitly covers DeepSeek V4, Claude Code, Opus 4.7, GPT 5.5, Gravity and is intended for intermediate to advanced users.

• Gemini 3.5 Flash In Arena! POWERFUL, Cheap, & Fast NEW AI Model! (Fully Tested)
  3.5.2026, 06:44:45

  **Summary**

  Google is testing an upgraded Gemini 3 Flash model behind the scenes, hidden under the same model slug in an arena but delivering significantly better output quality — users report reasoning and response quality almost two tiers above current standards, approaching Gemini 3.1 Pro. In parallel, Google informed Vertex AI customers that Gemini 3.1 Flash Light will soon be publicly available. The theory is that Google will release version 3.1 Flash before Google IO (May 19–20), then announce Gemini 3.5 Pro at the conference itself, with Gemini 3.5 Flash following in June/July — this would close the performance gap between 3.0 Flash and stronger 3.5 Pro.

  The creator tests the new model with several tasks: For front-end development (macOS browser UI, 360-degree product viewer, React animations) it shows quality at 3.1 Pro level. For 3D graphics (ThreeJS) it produces impressively detailed scenes — a PS5 controller scores 9/10, a 1970s TV simulator with nine different channels and real-time rendering works very well, only terrain navigation falls short. SVG generation (butterfly, pelican on bicycle) is solidly animated but shows minor accuracy issues. The model is accessible via Arena in Battle mode and proves significantly stronger for complex creative and technical tasks than typical Flash variants.

  **Finally:** Google / Gemini 3 Flash, 3.1 Pro, Vertex AI covered — demo & opinion/reflection.

---

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