Blog

Hermes Agent Meets Claude Code Agent View: Two Paradigms for Autonomous AI Assistants

Tuesday, June 16, 2026

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

Model Releases & Benchmarks

Google internally tested Gemini 3.2 variants (internally “Fanta,” “Sprite,” “Cola,” and the more promising “Cappuccino” checkpoint series for Gemini 3.5 Flash), whose code and UI generation according to WorldofAI tests still lag expectations — simultaneously, an unreleased multimodal video model called Gemini Omni was leaked, which early users could activate in the app and which qualitatively compares to Sora 2. OpenAI internally works on GPT-5.6 (codenames Ember Alpha, Beacon Alpha). The newly announced Subq model with a 12-million-token context window and allegedly subquadratic sparse attention architecture drew attention on AI with Arnie; according to the video, benchmarks are limited and contested, a technical paper is still missing. Anthropic’s Mythos Preview exceeded GPT-5.5 on independent evaluations (XBO and UK AI Security Institute) for cybersecurity tasks like attack chains and vulnerability discovery.

Local & Open-Source AI

Hermes Agent v2.0 received, alongside the desktop app, a new Qwen-3.6-Plus integration (free for a limited time), Light-Panda as integrated browser backend, and Computer-Use support. Alex Finn demonstrated Hermes on Nvidia DGX Spark with locally running Qwen 3.6 27B for use cases like AI stock reports, YouTube content repurposing, and vibe coding. WorldofAI showed the combination of OpenAI’s Codex CLI with Ollama, allowing models like Deepseek, Gemma 4, or Qwen 3 to be used locally and cost-free in Codex. ONNX format as a model framework-agnostic exchange format was presented by NeuralNine as a production-relevant alternative to PyTorch or TensorFlow deployment dependencies.

Claude Code & Anthropic Tooling

Beyond the Agent View dashboard and `/goal` feature — which multiple channels demonstrated in detail — Anthropic released Remote Control functionality, making Claude Code controllable via smartphone (Julian Ivanov). Nate Herk catalogued all five deployment methods for Claude automations from Loops to Cloud Routines to Trigger.dev. Matt Pocock brought new skills to the GitHub repository (70,000 stars): `/handoff` for context-aware agent handoffs, `/prototype` for UI variant generation, and first previews of `/review` and Writing Skills; he also replaced `/grill-me` with `Grill with Docs`, writing domain-driven design ubiquitous language into a `context.md`. Ben AI showed how to use Claude Design with prepared design systems, templates, and skills for consistent outputs, and presented five architecture skills for an AI operating system. Mark Kashef argued that most skill collections actively harm through token overload, and presented eight audit tips.

Coding Agents (non-Claude)

Melvynx documented his switch from Claude Code to OpenAI’s Codex: lower token costs, multitasking via work trees, better UI — and showed how skills are migrated via symlinks. Leon van Zyl built with Codex CLI a complete AI image generator (Next.js, Docker, Postgres, GPT-4 Image API) and delivered one of the week’s most practical Codex tutorials. Grok Build from XAI tested by Alex Finn live with 3D benchmarks (first-person shooter, Elon Musk animation, city flythrough) — the verdict: competent, but usability and cost remain weak points. MiniMax Agent based on MiniMax M1 (not M2.7 as stated in the video — positioned cheaply against GPT-5.5) demonstrated website deployment, CSV analysis with LaTeX PDF output, and video story generation.

Personal AI OS & Agent Frameworks

Cole Medin’s Archon — a specialized open-source harness builder for AI coding — went live this week with a workflow marketplace where users can share their own workflows; an automatic PR review workflow validates contributions. N8N unveiled its official MCP, allowing agents not just to execute workflows but also create and edit them, plus N8N Desk, an open-source desktop app similar to Claude Desktop but delegating execution to N8N cloud instances instead of locally. Nate Herk from NeuralNine built a fully automatic coding agent triggered via Slack or GitHub Issues, running Claude Code in headless mode and creating a pull request — orchestrated via two N8N workflows. Brian Casel demonstrated OpenClaw on a Mac Mini as the foundation for his daily marketing radar scan agent “Veil,” whose results land as markdown in Dropbox and are retrievable via Telegram. Sebastien Dubois categorized context management into eight maturity levels — from context-free queries to fully maintained, agent-readable knowledge systems.

AI Automation & Workflows

Nate B. Jones analyzed this week in multiple videos the infrastructure layer behind agent workflows: Pinecone overhauls its retrieval interface with NoQL for more agent-friendly data contracts; Page Index relies on document trees rather than chunks; SAP acquired Dreamio and Prior Labs for structured enterprise data; Microsoft pushes Graph RAG. The core thesis: the retrieval unit must match the job — chunks for FAQs, tables for financial analysis, graphs for dependencies. In parallel, Jones described the “LLM as Judge” pattern as a control layer for agent actions: a specialized judge model sits at the action boundary and decides between allow, block, revise, and human review — with four action categories from read-only to high-risk. Tech With Tim showed Browserbase as a platform for AI-driven browser automation (with Stagehand SDK) and Compose as a tool reducing context bloat in tool integration through on-demand discovery.

AI Video & Content Creation

Cole Medin demonstrated a complete workflow for AI-generated videos with audio: HyperFrames for rendering, Claude Code as control layer, 11Labs or Kokoro for voice output, and Archon as workflow manager — the open-source repository should be deployable in under ten minutes according to the video. Julian Ivanov used the same three-part harmony of Claude Code, HyperFrames, and Claude Design as a video editing replacement: Videouse cuts per transcript, HyperFrames generates motion graphics, Claude Design integrates branding — end result in 4K. Tech With Tim showed Higgs Field MCP Server as a unified interface for image (GPT Image 2) and video generation (Seedance 2.0) directly in Claude, including automatic prompt refinement by the agent.

AI Business, Marketing & Freelancing

Kyle Balmer laid bare his content production system: 20+ pieces daily with two people, Hermes/Claude for topic scouting and script generation, ChatGPT Images 2.0 for slides, Descript for transcripts, ManyChat for email giveaways (100–300 new addresses daily). The critical thesis: liveness and on-camera presence become more valuable through AI saturation, not less. Nate B. Jones detailed how Salesforce (Agentic Work Units, $800M ARR with AgentForce), Microsoft (Copilot Credits, Microsoft Agent 365 at $15/user/month), and ServiceNow (Action Fabric) are shifting their billing models from seat licenses to action units — and warned that negotiating position against vendors is only strong before go-live. TheAIGRID presented ten AI tools for small businesses, including Claude Co-work, Gamma.app, Canva AI, Zapier, and Chatbase.

AI Industry & Strategy

According to Ramp data, Anthropic for the first time has more verified business customers than OpenAI — with annualized revenues for both firms around $30 billion. Anthropic’s 80x growth versus the planned 10x creates a real compute problem, reflected in the new usage restrictions. Nate B. Jones described the emerging implementation layer as the actual value carrier of the agent era: workflow design, data access, authority limits, evaluations, and audit trails — a layer that neither Anthropic nor OpenAI can assume with a product launch because it’s customer-specific and deeply embedded in business objects (leads, cases, policies). McKinsey, BCG, Accenture, and PE firms with thousands of portfolio companies are positioning aggressively here. The security incident at McKinsey’s AI platform Lily — SQL injection enabled complete read-and-write access to millions of chat messages for $20 — was analyzed by Jones as a warning for all enterprise AI procurements: 22 of 200 API endpoints were deployed without authentication because security teams were brought in too late in the procurement decision.

AI & Society / Future of Work

Theo this week openly led the debate on cognitive atrophy from AI tools: developers lose skills when they no longer think through implementation details themselves — his counterproposal is to use AI for planning and research but deliberately retain implementation. The Coding-with-Lewis video on Meta Llama’s history traced how open-source community trust was squandered through manipulated Llama 4 benchmarks — a pattern relevant beyond Meta. Everlast AI discussed with experts China’s role as a technology power through 2030: systemic overcapacity, demographic pressure, semiconductor independence aspirations — and how Europe responds. The No-Priors interview with Jacob Helberg illuminated “Pax Silica,” the U.S. initiative to secure AI supply chains through a 14-nation ecosystem with forward-deployed industrial zones in the Philippines.

In Brief

Anthropic and Musk/SpaceX: Anthropic’s Claude models run on SpaceX Colossus-1 according to Everlast AI, which doubles weekly Claude Code limits. — AWS now allows agents to control desktop applications in managed workspaces (ERP, legacy software, mainframes). — Agentic Commerce: OpenAI/Stripe (ACP), Shopify/Google (UCP), and Google AP2 as competing authorization layers for autonomous purchases were detailed by Nate B. Jones. — Melvynx switched from Prisma to Convex as backend solution and unveiled his new NStack (TanStack Start + Convex + Better Auth). — Dave Ebbelaar showed hybrid search (BM25 + embeddings + reranker with Cohere) and agentic RAG in pure Python as production-ready patterns. — Nick Saraev published a complete course on building mobile apps with Claude Code. — Long Lake Management announced the acquisition of American Express Global Business Travel for $6.3 billion — positioned as the first AI-driven acquisition of a public company via AI platform Nexus.

AI Explained

No new videos in this period.

AI Foundations (1 new video)

• Full Hermes Agent Set-Up For Beginners in 2026! (24/7 AI Agent)
  12.5.2026, 14:56:27

  This video is a tutorial on setting up the Hermes Agent, a 24/7 self-improving AI agent that runs on a server and continuously learns from feedback. The agent consists of three layers: memory, skills, and cron jobs. Memory allows the agent to integrate learned knowledge into future sessions. Skills are automations or mini-workflows that the agent creates and improves. Cron jobs enable task scheduling at specific times. The agent goes through a loop where it sets goals, uses tools, delivers results, and improves through feedback.

  The author compares setting up the agent on a local computer with setting it up on a VPS (Virtual Private Server). A VPS is always online and scalable, while a local computer is only active when powered on. The VPS setup is described in detail, including using Hostinger for VPS deployment and configuring the agent via terminal. The agent can be used with various models like Claude, OpenAI, or Open Router and can be controlled via different messaging platforms like Telegram.

  At the end, the author demonstrates a practical example of how he uses the Hermes Agent to automatically create and publish blog posts based on his YouTube videos. The agent checks analytics, improves content, and can be controlled via Telegram.

  The video explicitly covers Hermes Agent, Open Router, Claude, OpenAI, and specific tools like Telegram. It is intended more for intermediate and advanced users.

AI mit Arnie (1 new video)

• Claude Code reicht nicht mehr
  15.5.2026, 13:06:39

  The video introduces the announcement of Subq (Subquadrantic), a new AI model with a context window of 12 million tokens, representing a significant advancement over existing models like ChatGPT (1 million tokens). Subq uses a subquadratic sparse attention architecture that enables linear instead of quadratic scaling, resulting in significantly higher efficiency and lower costs. The model is claimed to be 52 times faster and cost only 5% of competing models like Opus. The announcement was published on X (formerly Twitter) by Alexander and includes APIs for Subq Code and Subq Search.

  The advantages of Subq lie in the ability to maintain large codebases or lengthy documentation fully in context, which could render current workarounds like RAG (Retrieval-Augmented Generation) or agent workflows obsolete. The sparse attention architecture allows focusing on only relevant tokens, significantly boosting computational power and speed. However, detailed technical papers are not yet available, and the available benchmarks are limited and somewhat controversial.

  Subq’s potential is enormous, as it could revolutionize how we work with AI in areas like coding, law, research, and audits. However, the question remains whether the model can deliver the promised quality and reliability with large context windows. Available benchmarks show promising results, but further testing and validation are needed.

  The video explicitly addresses the new Subq model and is aimed more at intermediate or advanced users who engage with the technical details and implications of AI models.

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

• Anthropic’s Mythos Just Beat OpenAI’s GPT-5.5 At Real Hacking
  May 16, 2026, 15:01:03

  # Summary: Five Stories About AI Agents This Week

  The video covers five recent developments around AI agents and their practical application:

  **1. Notion Developer Platform:** Notion launched a Developer Platform with CLI, hosted functions (Workers), webhooks, and an external Agents API. This enables agents like Claude to be embedded directly in Notion workspaces as participants — not just buttons, but integrated workflow partners. Example: A sales deal in Salesforce triggers a Notion Worker that syncs data, an agent designs an onboarding plan that the CSM reviews before sending.

  **2. Claude usage limits are tightening:** Anthropic must ration compute capacity because agent usage is overwhelming subscription models. In April, Anthropic blocked third-party tools like Open Claw on personal subscriptions, driving many developers to OpenAI. Now Anthropic allows limited agent usage again under a monthly rate limit that converts to API token payments — a less transparent and less popular solution than OpenAI’s transparent model.

  **3. Anthropic overtaking OpenAI with business customers:** According to Ramp data, Anthropic now has more verified business customers than OpenAI for the first time. Both companies are at annualized revenues around $30 billion. Anthropic had planned 10x growth but is experiencing 80x — a real compute problem.

  **4. Mythos and GPT-5.5 demonstrate cyber capacity:** Independent evaluations (XBO and UK AI Security Institute) show that Mythos Preview outperforms GPT-5.5 on cybersecurity tasks (attack chains from reconnaissance to network takeover) within the same token budget. Mythos excels at code audits and vulnerability discovery but needs validation infrastructure. For software teams, this is a signal: build AI-powered security workflows, scan critical codebases with Mythos/5.5, prepare disclosure and patch processes.

  **5. AWS Workspaces for AI Agents:** AWS now allows agents to control desktop applications in managed Workspaces environments — ERP systems, legacy software, mainframe interfaces. This solves desktop automation for back-office work (claims, trade settlement, procurement), but requires governance: screenshot logging, audit trails, human approval gates, not immediate full automation.

  **Overarching point:** The AI agents era is now concrete — not in big model launches, but in workflow integration, billing upheaval, and practical enterprise automation. Scaling laws have made agents good enough to operate existing software; that changes what becomes quickly automatable.

  —

  Claude and OpenAI (GPT-5.5, GPT-5.5 Cyber), Open Claw, and AWS explicitly mentioned as providers/tools; news roundup format.

• Your SaaS Bill Just Got a Second Meter. You’re About to Pay It.
  May 15, 2026, 14:00:04

  # Summary: Agent Pricing and the Future of SaaS Models

  The focus is on the fundamental shift in SaaS pricing models driven by agent workflows. Salesforce reports $800 million ARR for its Agent Force product with 169% YoY growth, measuring not tokens but Agentic Work Units — the concrete actions agents perform within the platform. Simultaneously, Microsoft launches Agent 365 at $15 per user per month plus a separate governance license controlling how agents act in the Microsoft environment.

  The central issue: Traditional SaaS was based on seat licenses — you paid for the number of people using the software. This model breaks when an agent uses the software without any human user. Agents can read CRM entries, update support tickets, and execute workflows without a human being logged in. This removes “the human as unit of value.”

  **Three pricing models at a glance:**
  – **Salesforce**: Flex Credits and Agentic Work Units per action (updates, summaries, workflows)
  – **Microsoft**: Hybrid model with seat prices for Copilot plus additional Copilot Studio Credits for various agent features (Answers, generative responses, workflow actions)
  – **ServiceNow**: Action Fabric measures operational work units instead of API calls — triggering access management or starting onboarding incurs costs

  Additionally, access policies are tightening: SAP, for example, sets clear boundaries around external or non-SAP-approved agents accessing SAP data. This is decided contractually, not technically — potentially a toll booth for every agent from another vendor.

  **Fair vs. rent-seeking in agent licensing:**
  A fair model: transparent, predictable metering; unsuccessful actions don’t cost like successful ones; third-party agents have a defined path; clean distinction between read, draft, write, approval, and execution; settable caps; exportable usage data; rates remain stable over the contract term.

  A rent-seeking model: vague “AI access” with no transparency, only vendor-native agents practical, external agents blocked, paying for user data even if the customer generated it themselves, failed work counts as billable, the meter stays hidden until renewal, credit bundles that expire while overages bill immediately.

  **For developers this is critical:** Most teams only account for token costs, not realizing that contracts have already shifted to work units, workflow actions, or other meters. A developer the author spoke with used 8 billion tokens in one month — that’s the new normal. Agents must be cost-optimized: Which tool calls are expensive? Which are reversible? When do you read, write, or execute? An agent blindly executing every action with the same resource budget becomes a budget disaster in production.

  **Negotiation strategy before renewal:**
  You have leverage only while agents aren’t yet productive. After that, negotiation becomes asymmetrical — if daily workflows depend on them, the vendor can push through price increases. Concrete questions: What’s included in existing seats? Are agents acting on behalf of users covered? Does an independent agent need its own license? Can third-party agents use the same path as vendor agents? Which actions consume credits? Do failed actions count? Is the rate card fixed for the contract term? Can caps be set per department, workflow, or agent?

  The most critical question: Does the business model change if the agent reduces the number of human seats? If a support agent resolves hundreds of tickets — can you reduce support staff? That’s the point where vendor incentive and customer interest collide.

  **Core thesis:** Pricing follows platform control. The vendor that defines the new unit of work claims the right to monetize it. Salesforce measures agent actions because Salesforce defines many CRM actions. ServiceNow measures operational work because ServiceNow shapes the enterprise action layer. Microsoft wants Copilot Credits because Microsoft sits atop the productivity graph. Whoever doesn’t understand these incentives and only builds with the vendor agent, without securing third-party agent paths, risks having the agent economy ultimately owned by the vendor.

  **Opinion/reflection format with deep-dive elements; no specific AI models mentioned, but vendor platforms (Salesforce, Microsoft, ServiceNow, SAP, HubSpot, Workday, Atlassian, Zendesk) are central.**

• The Trillion Dollar Agentic Workflow Opportunity Is Here
  May 14, 2026, 14:01:09

  # Summary: Implementation Layer and the Future of Enterprise AI Agents

  This is a strategic analysis of the convergence of three forces: private equity needing to rescue its struggling SaaS portfolios, hyperscalers (OpenAI, Anthropic) realizing they need forward-deployed engineers on-site to implement enterprise AI, and companies desperately wanting to deploy agent workflows but not knowing how.

  The core thesis: The real value isn’t in the model, but in the **implementation layer** — the ecosystem of workflow design, data access, authority/limits, evaluations, and audit trails that transforms a generic AI agent into a functioning enterprise system.

  **Four pressure axes shape the market:**
  1. **Labs moving downmarket**: Anthropic and OpenAI are launching their own deployment companies, hiring embedded engineers, and building products like Claude Design or Finance Templates — a clear signal of where they see value.
  2. **Consultancies moving upmarket**: McKinsey, BCG, Accenture, and PWC are building deliberate agentic practices and deployment teams for production delivery — backed by decades of customer relationships.
  3. **Systems of record opening strategically**: Salesforce, ServiceNow, Workday, and SAP are building APIs and agent frameworks directly into their platforms. They don’t want a startup intermediary between data and agent.
  4. **Private equity as distribution**: PE firms control thousands of mid-market companies and can standardly roll out a deployment partner across entire portfolios.

  **The implementation layer in detail:**
  – **Workflow Design**: Which decisions can the model make? Which steps stay manual? Where are the transitions?
  – **Data Access**: Which data sources are authoritative? Which permissions apply at row/field level?
  – **Authority**: What can the agent do? What limits on commits or spending?
  – **Evals**: How is correct/complete/safe measured — not as a benchmark, but as scoring against business rules?
  – **Audit Trails & Ownership**: Logging, error reconstruction, reversal, continuous tuning.

  **The strategic core**: SaaS was “generic” — one solution worked everywhere. Agentic workflows are the opposite: disproportionate value lies in customization and object-oriented design. A sales agent must understand sales objects (leads, accounts, deals, funnel stages). A support agent must model cases, policies, customers, entitlements.

  **For builders**: Sit closer to the business object than abstract reasoning. Understand the implementation layer deeply enough to recognize what actually creates value versus pure model marketing. PE firms are already testing whether in-house teams can build with Claude in a weekend — that fails with real enterprise workflows because the implementation layer is too nuanced.

  **For buyers**: Ask hard questions about data objects, workflow details, and ownership post-launch. Generic answers like “the model will improve” or “your data helps us” are red flags.

  **For PE**: This is a trillion-dollar market because agents can run entire workflows — something reliably and scalably possible since around 2026. PE profits twice: push pressure (SaaS in danger) and pull (standardize AI across portfolio companies and resell).

  The message: The implementation layer architecture isn’t something Anthropic or OpenAI can bulldoze with a product launch tomorrow — it’s specific, custom, and biased toward in-house builds. Whoever controls this layer controls the value.

  **Demo/opinion hybrid with deep enterprise strategic dive; no explicit AI tool names except Claude, OpenAI, and Anthropic; PE and hyperscalers as central actors.**

• Pinecone Just Demoted Vector Search. Here’s the Knowledge Layer.
  May 13, 2026, 14:01:15

  # Summary

  The video analyzes a fundamental problem in agent architecture: production agents waste massive amounts of context window and tokens through redundancy rediscovery. Pinecone, SAP, Google, and Microsoft are simultaneously building systems to solve this memory problem — but not through better vector databases, but through structurally different retrieval approaches.

  The key insight: Classical RAG (vector search over text chunks) was optimized for chatbots — a user asks, three semantically similar paragraphs come back. Agents work differently: they execute longer-term workflows, need cross-referencing across hundreds of pages, and must deliver correct, contextual answers, not just relevant text passages. Pinecone is hardening their retrieval interface with NoQL to carry intent, filters, access policy, and provenance. Page Index argues that many documents shouldn’t be chunked at all since their hierarchical structure (like contracts with definitions 40 pages away) is the meaning — they use document trees instead. SAP acquired Dreamio (lakehouse with semantic layer) and Prior Labs (tabular foundation models) because most enterprise knowledge lives in structured tables, not prose. Microsoft is pushing Graph RAG for relationally structured queries.

  The decisive principle: The retrieval unit must match the job — chunks for FAQs, sections for filings, tables for financial analysis, graph neighborhoods for dependencies. Larger context windows don’t automatically help; the problem isn’t whether the answer is in there somewhere, but whether it’s presented in a form the model can reliably use (context rot).

  Concrete steps when building an agent: (1) Define the data contract first, not the database — what must the agent receive in what form to work reliably? (2) Write down the specific bundle the agent needs (not just “relevant context”) — e.g., for a refund agent: customer record, plan, region, product version, purchase history, valid refund policy, threshold, previous exceptions, current ticket, approved response language, permissions. (3) Then choose the primitives to deliver that bundle — mix of vector search, document trees, semantic layer, tabular reasoning, or graphs, whatever fits.

  Failure modes: Precompiled bundles can become stale; graphs can encode bad relationships; parsing errors in tables; semantic layers become politicized; agents can store their own inferences as confirmed facts and thus degrade future runs. At the same time: don’t overengineer — a simple help center assistant doesn’t need graph RAG plus document tree plus everything. Learn from your own agent runs: How many retrieval calls before real work? How often the same sources? How many tokens just for raw context?

  The video is conceptual: it’s not about a specific tool but the architecture decision underneath. The central thesis: teams that win think through the agent job first, not the trendy database.

  —

  **Explicitly mentioned providers/tools:** Pinecone (Nexus, NoQL), Page Index, SAP (Dreamio, Prior Labs), Google Cloud Next, Cloudflare, Microsoft (Graph RAG), Chroma, Weev8, Neo4j — **format:** opinion/reflection with deep-dive elements in infrastructure landscape.

• ChatGPT Has 900M Weekly Users. Almost None Can Buy In It.
  May 12, 2026, 14:01:07

  # Summary: The Six Battlegrounds of Agentic Commerce

  The video analyzes the fundamental reshaping of e-commerce through autonomous agents and identifies six competing camps fighting for control of emerging agentic commerce infrastructure.

  **The central problem:** While traditional online purchase has clear structure — human sees product, price, shipping, clicks checkout — autonomous purchase breaks this down. An agent could search products, make purchases, and authorize payments without human oversight at each step. This creates questions the existing infrastructure doesn’t answer: Who decides where the agent shops? Who proves the agent has authority to act? Who bears responsibility for mistakes?

  **The six layers of conflict:**

  1. **Agent Checkout (OpenAI/Stripe vs. Shopify/Google):** ACP (Agent Commerce Protocol) from OpenAI and Stripe enables purchases directly in the chat interface — fast and simple, but merchants lose control of customer journey and branding. UCP (Universal Commerce Protocol) from Shopify and Google preserves merchant rules, loyalty, discovery, and complex business logic while agents interact.

  2. **Authorization/Approval (Google AP2, Stripe, Visa, Mastercard, PayPal):** Payment ≠ Authorization. An agent needs prior-granted permission with clear scope (budget, constraints). Google AP2 creates a digital “approval” with proof that the user authorized the agent. Here Google, Stripe, Visa, Mastercard, and PayPal compete for the trust layer.

  3. **Stablecoins & Machine-to-Machine Payments (Coinbase X42, Stripe MMP):** For agent-to-agent transactions (API calls, tool access, data queries), credit cards are inefficient. Stablecoins like USDC enable fast, cheap machine payments. X42 makes payment part of the web request itself (HTTP 402).

  4. **Enterprise Governance (AWS with Coinbase/Stripe):** AWS controls the runtime environment where agents operate. AWS can manage budgets, approval rules, vendor access, and logs — without owning any payment rail itself. This is the most powerful position in the system.

  5. **Platform Control:** Large platforms (Amazon) won’t simply let external agents run free. Agentic commerce will be partly protocol-driven, partly platform-controlled, partly only open under strict conditions.

  6. **Accountability:** No single answer to “who bears risk.” Accountability must deliberately distribute across layers: ACP answers checkout questions, UCP answers merchant system questions, AP2 answers authorization questions, payment networks answer trust and dispute questions, wallet providers answer settlement questions.

  **Core thesis:** Agentic commerce will redistribute trillions of dollars in commercial authority over the coming years. Companies must consciously choose which layer to control. Consumers must understand how these systems work to avoid being passively pulled into systems they don’t comprehend. This is a bigger shift in internet economics than the original e-commerce revolution of the 1990s.

  —

  **Format & Coverage:** Opinion/reflection with deep-dive structure; explicitly covers: OpenAI, Stripe, Shopify, Google, AWS, Coinbase, Visa, Mastercard, PayPal (ACP, UCP, AP2, X42, Machine Payments Protocol, Bedrock Agent Core Payments).

• LLM Agents: The Security Breach Pattern Nobody’s Talking About
  May 11, 2026, 14:00:52

  # Summary: LLM as Judge – The Security Pattern for Agents

  The video covers a critical architecture pattern for preventing agent errors that happen in production: unintended emails, deleted data, unauthorized operations. The speaker works through several key points.

  **The core problem:** Agents are trained to act, but until now lacked a control layer deciding *when* and *how* they act. Simple prompts and manual approval don’t work: prompts don’t hold over long context windows, and when users are constantly asked for approval, they click “OK” out of habit without looking — the EU cookie directive problem.

  **The solution – LLM as Judge:** The Lindy example shows the pattern: a specialized “judge model” reads the executive agent’s planned action, checks the rationale against available context, and decides. Not two identical agents, but one focused on the task, one on intent violation. This leverages the new capacity of long context windows and is significantly more reliable on modern frontier models (like GPT 5.5 or Claude 3.7 Opus in May 2026) than on older or open-source models.

  **Classification of agent actions into four categories:**
  – Read-only: easy, minimal risk
  – Reversible writes (drafts, labels, local files): needs validation, but maybe no audit trail
  – External actions (emails, meeting bookings, pull requests, customer notifications): must *always* go through strong judge layer before execution
  – High-risk (spending money, deleting data, changing permissions, merging code): judge plus human approval, or very explicit automated policies

  **Proper implementation:** The judge must sit at the action boundary — exactly where the agent triggers a tool call. Not just yes/no, but four possible outcomes: allow, block, ask agent to rework, or escalate to human review. Balancing the escalation rate is critical.

  **Correlated Judgment:** If actor and judge are the same model, they share blind spots — but that’s significantly less problematic with modern frontier models (May 2026) than 6-8 months ago. On older or open-source models, a real risk.

  **Broader context:** Agents are evolving from isolated workflows to “managed workers” needing task assignment, supervision, and correction. The first agent wave focused on setup, the second (now) on the management system — the judge is the agent’s manager. The entire product is no longer just the agent but the system around it.

  The speaker references a detailed Substack article with implementation details per action type, judge design patterns, and metrics.

  —

  **Format & Models:** Opinion/deep-dive with strong architecture focus; explicitly mentioned: Claude (Opus 3.7), GPT 5.5, Qwen, Gemini, older models; Lindy and Codeex as product examples; not a tutorial.

• Anthropic And OpenAI Just Admitted The Model Isn’t Enough.
  May 10, 2026, 18:00:09

  # Summary: The Lily Security Breach and Its Lessons for AI Procurement

  In February 2026, the startup Codewall discovered a critical security vulnerability in Lily, McKenzie’s AI platform, accessed daily by 70% of 40,000 consultants. An autonomous agent gained complete read and write access within two hours on a $20 budget to tens of millions of chat messages, thousands of user accounts, and all system prompts — and could even rewrite them. The security hole was classic SQL injection, a vulnerability taught in every web security course since 1998.

  The real mistake, however, lay not in technical hygiene gaps (McKenzie has strong engineering teams) but in organizational structure: 22 of 200 API endpoints were deployed without authentication, including a productive, writable connection. This suggests technical security concerns weren’t prioritized in planning — a consequence of the traditional SaaS procurement process (strategy → procurement → security review → IT integration → development) that doesn’t work for agents.

  The key difference: while humans visually limit access through screens, agents must query every system in code and receive explicit responses about whether they’re authorized. This requires carefully coded permission models, audit trails, and workflow orchestration across all integrated systems — before the platform goes live. This complexity isn’t “implementation details to solve later,” but already a strategic decision.

  Vendor responses show the industry recognized this: Anthropic and OpenAI built enterprise services teams; SAP acquired Dreo and Prior Labs for data layer integration; Pinecone, Salesforce, and ServiceNow all released solutions in the same timeframe for governed agent workflows, audit trails, and controlled data access.

  Concrete questions for your AI investment: (1) Does your platform technically distinguish between human users and agents and separately bound their permissions? (2) Can your team disable agents from a console within five minutes if something goes wrong — or do tickets need opening? (3) Can your compliance team trace what agents did on behalf of users (for regulators)? (4) What’s the organizational *default behavior* under time pressure — do teams tend to push unauthenticated endpoints to production when it needs to move fast?

  The recommendation: Bring technical teams much earlier into the procurement sequence, not for cleanup after buying. A six-question checklist for developers covers additional failure modes (agent delegation, token costs at scale, reversibility on agent errors). The critical lesson isn’t “Lily missed security,” but “our procurement and build processes aren’t calibrated for agent-native workloads.”

  **Video format: Opinion/deep-dive with case study analysis. Explicitly covers: Anthropic, OpenAI, SAP, Salesforce, ServiceNow, Pinecone; no specific AI models mentioned.**

Alejandro AO (1 new video)

• MiniMax Agent: Alternative to ChatGPT + Codex?
  12.5.2026, 12:46:44

  The video showcases the features of MiniMax Agent, an AI tool positioned as a competitor to Codex and Cloud Cowork. MiniMax Agent is powered by the MiniMax M2.7 model, which, while not as powerful as GPT-5.5 or Opus 4.7, still achieves high benchmark scores and is offered at a more affordable price.

  The video demonstrates how to use MiniMax Agent to create a website and deploy it on MiniMax Cloud. It introduces both the web and desktop app of the tool. The desktop app is used for data analysis of a CSV dataset, generating a PDF report with LaTeX. Additionally, the “Skills” of MiniMax Agent are shown, including a video story generator that creates videos from text descriptions or images.

  The video also mentions MiniMax’s pricing models, which include a free version with daily credits as well as premium plans. Particularly highlighted is the tool’s ability to handle complex tasks such as website creation, data analysis, and video generation.

  Final comment: The video explicitly focuses on MiniMax Agent and is aimed more at intermediate and advanced users.

Alex Finn (4 new videos)

• LIVE: GROK BUILD CLI IS ACTUALLY INCREDIBLE???
  15.5.2026, 20:24:28

  ### Video Summary

  The video is a live-stream session where the host discusses various aspects of selling Software-as-a-Service (SaaS) while testing the capabilities of Grok Build, a tool from XAI. The stream opens with a casual introduction where the host encourages viewers to participate and chat. He talks about the importance of Twitter (X) for software sales and the need to build a community.

  He then goes into the three key factors that determine SaaS product success: distribution, network effects, and partnerships with government data. He emphasizes that a strong community and good sales are critical for success.

  A large portion of the video is devoted to demonstrating and testing Grok Build, a new XAI tool. The host runs several benchmark tests, including creating a 3D first-person shooter, an Elon Musk animation, and a 3D city flythrough. He evaluates the results and compares them to other AI tools like Claude, OpenAI, and Gemini. While Grok Build performs well in some areas, there’s also criticism about the tool’s user-friendliness and costs.

  The host also shares personal experiences with various AI tools and his thoughts on the current landscape of AI development. He emphasizes the importance of coding skills for AI tool success and praises Anthropic for developing Claude Code.

  Toward the end of the video, the host announces that he’ll be doing fewer live streams in the future and will focus more on creating pre-recorded videos. He invites viewers to join his Vibe Coding Academy Bootcamp, where he’ll discuss more about SaaS, AI investments, and his workflows.

  ### Final Comment

  The video explicitly covers AI tools and models from Claude (Anthropic), OpenAI (ChatGPT, Codex), Gemini (Google), and specific tools like Grok Build (XAI). It’s aimed more at intermediate and advanced users interested in developing and selling SaaS products and the latest developments in the AI industry.

• LIVE: The greatest Claude Code workflow ever
  13.5.2026, 20:12:33

  The video showcases a detailed, advanced Claude Code workflow that integrates various tools like Slack, Linear, GitHub, and Claude Code. The host explains how these tools work together to boost productivity, track changes, and organize development. The workflow includes creating tasks and projects in Linear, linking them to GitHub for branch management, and using Claude Code to automate and manage these processes. The host emphasizes the benefits of this workflow, such as increased speed, better error prevention, and improved traceability.

  Additionally, the host discusses personal experiences and challenges, like dealing with difficult times and the importance of persistence. He also shares his thoughts on using AI tools like Claude Code and Codex, as well as their differences and use cases.

  The video is aimed at advanced users who already have experience with Claude Code and similar tools and want to optimize their workflows. Specific tools like Claude, OpenAI, and Linear are discussed.

• Hermes Agent powered by local models on the DGX Spark is basically magic
  13.5.2026, 13:30:07

  The video shows how to set up a Hermes Agent on a local model on an Nvidia DGX Spark to create a 24/7 available AI employee. The process includes setting up the DGX Spark in headless mode, installing a local model (Qwen 3.6 27B), and integrating the model into the Hermes Agent. The creator demonstrates three use cases: a daily report on AI stocks for beginners, repurposing YouTube video content for advanced users, and vibe-coding a to-do list app for advanced users. The focus is on the benefits of local models, such as cost-effectiveness (aside from electricity), privacy, customizability, and educational value.

  Final Comment: The video explicitly covers Nvidia DGX Spark, Hermes Agent, Qwen 3.6 27B, and Tail Scale and is aimed at intermediate to advanced users.

• LIVE: Talking AI news (no Hermes use cases ignore the thumbnail)
  11.5.2026, 20:11:47

  The video is a live-stream session that primarily focuses on discussing Hermes Agent and its use cases. The host, Alex Finn, begins with an introduction to Hermes Agent and emphasizes the importance of use cases for utilizing the technology. He mentions that Hermes has released a new website with hundreds of use cases that he and the viewers will go through and test.

  However, a large portion of the video is taken up by various tangents and discussions, including:

  1. **Investment Corner**: Alex talks about current investment opportunities in the AI industry, particularly in companies like Nvidia, Micron, TSMC, and Tesla. He emphasizes the importance of current AI developments and the need to invest in the right companies.

  2. **Personal Stories and Anecdotes**: Alex shares personal stories and anecdotes that are often humorous and entertaining but don’t directly relate to the video’s main topic.

  3. **Chat Interaction**: A large part of the video consists of interacting with the live chat, where Alex answers questions, makes comments, and engages with viewers.

  4. **Hermes Agent Use Cases**: Toward the end of the video, Alex begins discussing some of the use cases from the new Hermes website. He mentions use cases such as creating research reports, managing tasks, and automating processes.

  5. **Announcements and Updates**: Alex provides updates on his own projects and announcements, such as launching a second YouTube channel and attending a Baby Keem concert.

  **Final Comment**: The video explicitly covers open-source AI models and tools like Hermes Agent and OpenClaw. It’s aimed more at intermediate to advanced users who already have some understanding of AI and its applications.

Bart Slodyczka

No new videos during this period.

Ben AI (2 new videos)

• 5 Skills to Build an AI Operating System Like The 1% (Full Guide)
  16.5.2026, 08:48:55

  The video shows how to set up a “Second Brain” or an AI Operating System to increase efficiency and productivity with AI tools. The author, Ben, emphasizes the importance of a well-structured and maintained system to optimize token costs and provide relevant context. He introduces five Claude Skills that help set up the system quickly and according to best practices:

  1. **OS Setup Skill**: Helps with the initial setup of the Second Brain, including populating context, structuring folders, and creating Claude.md files that serve as an instruction layer for AI agents.

  2. **OS Operator Skill**: Sets up a scheduled task that pulls real-time context from various sources (e.g., meetings, Slack chats) and updates the Second Brain accordingly. This includes creating daily summaries, task lists, and maintaining existing files.

  3. **OS Optimizer Skill**: Performs regular audits and optimizations to improve the efficiency and token usage of the Second Brain. This includes removing duplicates, correcting formatting, and improving folder structure.

  4. **Team OS Skill**: Enables sharing and synchronization of the Second Brain within a team, including setting up read and write permissions for different team members.

  5. **OS MCP Skill**: Creates an MCP (Model Context Protocol) from the Second Brain, so scheduled tasks and optimizations can run autonomously, even with the laptop closed.

  The author recommends starting with the Second Brain setup and expanding it gradually, as the benefits increase over time and with growing context volume. He offers additional resources and support for those who want to dive deeper into the topic.

  The video explicitly features Claude and is better suited for intermediate and advanced users.

• How to Actually Use Claude Design Like a Pro (Real Use Cases)
  12.5.2026, 07:40:42

  The video shows how to use Claude Design efficiently for various design tasks, including presentations, social media content, and websites. The main focus is on a four-step process: First, setting up a comprehensive design system containing colors, fonts, and styles. Second, using templates to predetermine formats and layouts. Third, using Skills to predefine texts and content. Fourth, integrating these elements into Claude Design to quickly create consistent designs. The creator emphasizes that this preparation avoids endless iterations and high token costs. Concrete examples and steps for setting up design systems and templates are shown, as well as using Skills to automate the process. Additionally, free resources and tools are offered to facilitate getting started.

  The video explicitly features Claude Design and is better suited for intermediate users who already have some experience with AI tools.

Brian Casel (1 new video)

• How I build agents that work the night shift
  12.5.2026, 12:01:07

  The video demonstrates a process the user calls “Radar Scan,” in which multiple AI agents run on a Mac Mini with OpenClaw. One of these agents, called Veil, handles marketing and performs a Radar Scan daily at 4:00 AM. The result is saved as a Markdown file in Dropbox and can be retrieved via a link in Telegram. The file is opened in a custom-built app called Brainown, which facilitates reading and writing Markdown files. The scan primarily summarizes tweets from the user’s industry that they follow, including posts from companies, influencers, and thought leaders. The user organizes their Telegram contacts into human and agent friends and receives a report overview in Telegram, with the option to open the full file in Brainown.

  The user covers OpenClaw and the custom-built app Brainown, making this video particularly interesting for Intermediate or Advanced users.

Coding with Lewis (1 new video)

• How Meta Went From Open Source Hero to AI’s Biggest Villain
  15.5.2026, 15:04:55

  The video tells the story of Meta and its open-source AI model Llama, which was released in 2023 and downloaded over a billion times within two years. Meta was praised for its open and accessible approach to AI, which enabled developers to use the model on their own hardware and build upon it. Llama 2 and Llama 3 followed, demonstrating impressive performance, with Llama 3.1 even competing with the best proprietary models from OpenAI and Google. However, Meta’s success was overshadowed by the release of DeepSeek v3 by a Chinese AI lab, which achieved similar performance at a fraction of the cost. This put pressure on Meta to release Llama 4, which ultimately led to a disaster. Llama 4 was released on a Saturday without warning and unveiled three models, one of which, Behemoth, was still in development. Llama 4’s claimed performance was quickly called into question when developers and researchers discovered that the models didn’t live up to the promises and that Meta had used manipulated benchmarks. The community reacted with disappointment and distrust, and Meta lost the trust it had built over years. The story demonstrates how important honesty and transparency are in the open-source AI community and how quickly trust can be lost.

  The video explicitly covers Meta’s Llama models and is geared more toward Intermediate or Advanced learners.

Cole Medin (4 new videos)

• 🔴 The AI Coding Marketplace is Finally LIVE!
  15.5.2026, 03:17:26

  The video demonstrates how to create fully animated videos with audio using AI. The process leverages multiple technologies, including HyperFrames for rendering, Claude Code for orchestration, 11Labs or Kokoro for voice generation, and Archon as a workflow manager. The creator provides an open-source repository that enables you to create your own AI-generated video in under 10 minutes. The workflow encompasses scripting, audio creation, visual rendering, and synchronizing all elements. The creator emphasizes that while the technology isn’t perfect yet, it’s advancing rapidly and already finding practical applications, particularly for explainer videos. The process is explained in detail, including options to customize templates and create your own. The video concludes with examples of generated videos.

  The video explicitly covers HyperFrames, Claude Code, 11Labs, Kokoro, and Archon and is aimed at intermediate users.

• Make the PERFECT Videos with Claude Code (Full Workflow)
  14.5.2026, 00:00:24

  **Summary:**

  In this live stream, development of a workflow marketplace for Archon, an open-source harness builder for AI-coding, continued. The focus was on creating a marketplace where users can share their own workflows and leverage workflows created by others. The process involved building an Archon workflow that automatically reviews and approves pull requests for new workflows. The stream began by merging an existing pull request that introduced the marketplace UI, then proceeded with creating a new workflow to handle pull request reviews.

  The process involved multiple steps, including creating a plan, implementing the plan, and reviewing the generated code. Various questions and adjustments were discussed to ensure the workflow functions correctly. The stream concluded with creating a pull request for a test workflow and demonstrating how the automatic review process works.

  **Final Comment:**
  The video explicitly covers the use of Claude (Claude Code) and is aimed at intermediate to advanced users.

• Building the App Store for Agentic Engineering
  12.5.2026, 04:02:35

  **YouTube Video Summary:**

  The video features a live demo of Archon, an AI tool developed as an open-source harness builder for AI-coding. The streamer presents his current AI-coding workflow and how Archon accelerates this process tenfold. He demonstrates using Archon for various tasks, including handling GitHub issues (brownfield development) and building new features (greenfield development).

  1. **Brownfield Development:**
  – The streamer shows how he uses Archon to handle multiple GitHub issues in parallel. Workflows incorporate planning, implementation, and validation steps.
  – The workflows are designed to generate comprehensive pull requests that can be manually reviewed.
  – A demonstration shows how Archon integrates with a “second brain” (a system for organizing knowledge and tasks) to optimize the workflow.

  2. **Greenfield Development:**
  – The streamer plans and implements a new feature for Archon: a workflow marketplace that enables users to create and share workflows.
  – The PIV-Loop (Plan, Implement, Validate) is used—a structured approach to AI-assisted development encompassing planning, implementation, and validation.
  – The demonstration shows how Archon workflows can tackle complex tasks like building a new marketplace feature.

  3. **Technical Details:**
  – Archon enables creating workflows that integrate various AI models and tools, including Claude, Codex, and others.
  – The workflows are designed to be deterministic and reproducible, enhancing reliability and efficiency.
  – The streamer emphasizes the importance of human-in-the-loop processes to ensure result quality.

  4. **Integration and Extensions:**
  – The demonstration shows how Archon can integrate with other tools like Beads (a memory system).
  – The streamer discusses Archon’s advantages compared to other tools like N8N, emphasizing Archon’s specialization in AI-coding.

  5. **Community and Further Development:**
  – The streamer mentions the Dynamis community, where he regularly offers workshops and courses to help users effectively use Archon and other AI tools.
  – He announces that he will continue live streaming to demonstrate Archon’s development and usage.

  **Final Comment:**
  The video explicitly covers the Archon tool and is aimed at intermediate to advanced users interested in AI-coding and workflow automation.

• 🔴LIVE – My AI Coding Workflow has 10x’d Again with Archon – See it in Action
  10.5.2026, 05:56:02

  The YouTuber expresses frustration with the current YouTube landscape, which heavily focuses on reporting Claude’s latest features. He aims to stand out by offering deeper, more technical content that emphasizes actually building projects and AI-coding principles. To achieve this, he plans to stream three times per week (Monday, Thursday, Saturday), showcasing projects like Archon and the Dark Factory experiment while working interactively with the community. He emphasizes that he will continue covering relevant AI news but with a focus on practical application and long-term value. The content is intended for intermediate and advanced users interested in AI-coding and systems. The YouTuber explicitly covers Claude and Claude Code.

Dave Ebbelaar (2 new videos)

• The Complete Guide to Hybrid Search in RAG (BM25 + Embeddings + Reranker)
  14.5.2026, 17:52:50

  This video is a detailed tutorial that shows how to build a hybrid retrieval system from scratch, combining BM25, dense embeddings, reciprocal rank fusion (RRF), and a re-ranker. The focus is on creating a production-ready system that will be relevant in 2026. The tutorial begins by explaining the dataset used, the Financial QA dataset, which is part of the BEIR benchmarks. This dataset contains financial questions and associated answers, which are used to evaluate the retrieval system.

  The tutorial walks through the process of creating BM25 and dense embedding indices, performing queries, and combining results with RRF. A re-ranker is then added to further improve the results. The tutorial emphasizes the importance of evaluation and shows how to measure system performance using normalized discounted cumulative gain (NDCG).

  At the end, it explains how to apply the system to your own projects, including creating your own evaluation dataset. The tutorial is intended for developers and engineers who already have foundational knowledge of retrieval systems and want to use them professionally.

  Final note: The video covers OpenAI (for embedding models) and Cohere (for the re-ranker) and is geared toward intermediate to advanced users.

• Building Agentic RAG From Scratch in Pure Python
  10.5.2026, 09:57:56

  This video shows how to build an agentic RAG (Retrieval-Augmented Generation) system from scratch in pure Python. The focus is on making enterprise data or private information accessible to large language models for use in AI automation. The author, Dave Abalar, explains the differences between classical semantic RAG and agentic RAG, with the latter surpassing classical approaches through a feedback loop and repeated use of the language model’s intelligence.

  The system is built in several steps: First, simple tools are defined that can list, search, and read files. These tools work with Markdown files in the file system. The author shows how to implement these tools in Python, including the use of regular expressions to search for patterns in files. A simple agent is then created with these tools that can answer questions about the contents of the Markdown files. The agent uses the tools in a loop to find the right information and correct itself.

  The author also covers production best practices, such as using Rust-based tools like ripgrep for faster and more secure file searching, as well as implementing error messages that the language model can interpret to improve the agent. He shows how to deploy the system in various environments such as VPS, container apps, or serverless functions.

  At the end of the video, a complete example of an agentic RAG system in Python is presented that follows production best practices and can be deployed in real-world projects.

  The video covers OpenAI and is suitable for intermediate or advanced users.

David Shapiro

No new videos in this period.

Everlast AI (3 new videos)

• China-Insider packt aus: Es gibt KEIN zurück mehr! DAS rollt bis 2030 auf uns zu (Jörg Wuttke)
  14.5.2026, 15:15:38

  The video offers a deep insight into China’s current economic and political situation, particularly in the context of innovation, technology, and demographic challenges. Here are the key points:

  1. **Innovation and overcapacity**: China has a deep pool of engineers and leads in many areas such as AI, humanoid robotics, energy, and electric vehicles. However, China systematically produces overcapacity, leading to deflation and economic problems. This is because state-owned enterprises don’t disappear from the market even when they’re unprofitable.

  2. **Demographic challenges**: China is aging rapidly and has a low birth rate. The population is expected to decline to about 500 million by 2060. This creates a labor shortage that China seeks to offset through AI and robotics.

  3. **Technological independence**: China aims to become independent from Western technologies, particularly in semiconductors. This is a response to the US’s restrictive policy limiting high-tech exports to China.

  4. **Export pressure**: Due to overcapacity and weak domestic consumption, China relies on exports. This creates intense competitive pressure on global markets, especially in Africa, South America, and Southeast Asia.

  5. **Political and economic strategies**: China relies on long-term plans, such as the 15-year plan, to achieve its goals. However, there are also internal challenges, such as population aging and migration from rural areas to cities.

  6. **Corporate strategies**: Chinese companies are more risk-taking and customer-centric than their Western competitors. This enables faster development and adaptation to market needs.

  7. **Geopolitical implications**: Demographic and economic changes in China will significantly influence the global geopolitical landscape. Africa will gain importance as a supplier, while China’s influence declines.

  The video is more suitable for intermediate and advanced viewers, as it deals with complex economic and political relationships. No specific AI tools or models are discussed, but rather general trends and strategies.

• KI-Experten reagieren: DAS übersehen alle! Musks Claude-Deal, Chinas Roboter-Druck & EU kippt AI-Act
  12.5.2026, 15:15:26

  The video is a discussion about current developments and trends in artificial intelligence (AI), technology, and the future. The participants are Professor Dr. Andreas Moring, Professor Dr. Pero Mitic, and Kim Isenberg, who discuss various aspects of AI usage, robotics, automation, and their impacts on society and the economy.

  **Main topics and points:**

  1. **AI models and infrastructure:**
  – **Anthropic and Elon Musk:** Anthropic now uses Elon Musk’s data centers (Colossus 1 and 2) for its AI models. This is seen as a strategic move since Musk needs to rent out his computing power to avoid leaving it unused.
  – **Gemini and Apple:** Apple is using Google Gemini for Siri after negotiations with Anthropic failed. This shows that Apple doesn’t necessarily need the best AI model, but rather one that’s good enough for the mass market.
  – **Humanoid robots:** Boston Dynamics is showing progress in robotics, particularly with the Atlas robot, which can perform complex movements. This underscores the importance of robotics in manufacturing, logistics, and services.

  2. **Economic and political aspects:**
  – **China vs. Europe/USA:** China is far ahead in many areas of technology and infrastructure, particularly in robotics and AI. Europe, on the other hand, is struggling to keep up, both due to political decisions and economic challenges.
  – **Job losses due to AI:** Companies like Coinbase, PayPal, and Cloudflare are laying off employees, citing AI adoption. There is debate about whether these are truly AI-driven layoffs or if economic difficulties are the real reason.
  – **Regulation and data protection:** The EU is discussing stricter regulations, such as banning VPNs for child protection. This is viewed critically as it’s seen as a restriction of freedom.

  3. **Healthcare and AI:**
  – **AI in healthcare:** Google presents studies showing that AI diagnoses are often better than those by doctors. This could revolutionize the healthcare sector, but also raises questions about self-perception and technology dependence.

  4. **Education and future competitiveness:**
  – **Education systems:** It’s emphasized that schools and universities must change to foster judgment and critical thinking rather than just transmitting information.
  – **AI investments:** Companies should invest in AI to remain competitive. This includes developing their own AI-powered applications and using platforms like Cloud Code.

  **Final comment:**
  The video explicitly discusses models and providers such as Claude, OpenAI, Gemini, and open-source solutions. It’s aimed at an advanced audience already familiar with AI and technology seeking deeper insights.

• KI-News: Claude verbessert sich IM SCHLAF! Musk rettet Anthropic, Codex & GPT-Realtime Updates
  10.5.2026, 08:15:10

  The video covers several current developments and updates in the field of artificial intelligence and agentic coding. Here are the key points:

  1. **Elon Musk’s SpaceX deal with Anthropic**: The deal doubles the 5-hour usage limit in Cloud Code, while weekly limits remain unchanged. This is part of a partnership where Anthropic runs cloud models on SpaceX’s Colossus-1 supercomputer.

  2. **Gemma 4**: Google’s open-source model is now three times faster thanks to multitoken prediction (MTP drafter). It can be used in Cloud Code, but with performance limitations.

  3. **Cloud Managed Agents**: Updates include multisession, webhook integration, and the new “dreaming” feature, which addresses the memory problem of agents. This feature was discussed with Vincent Pistor from Cogni, a startup specializing in agent memory.

  4. **Gemini updates**: Google Gemini is getting new features like skills and scheduled actions. Additionally, Gemini 3.1 Flashlight was made publicly available, a cost-effective model for high-volume tasks.

  5. **Codex Cloud vs. Cloud Code**: Codex Cloud has dramatically caught up in downloads and offers new features like a Chrome extension for browser automation.

  6. **Comparison of Remotion and Hyper Frames**: Both tools enable creating videos and animations via text prompt. Hyper Frames offers a faster entry point and more reliable results, while Remotion is better suited for scaling.

  7. **OpenAI’s Realtime 2 Model**: This model is designed for agentic tasks and enables better integration into software and user interfaces. It’s the engine behind ChatGPT’s Advanced Voice Mode.

  8. **AI business idea of the week**: The App Store is seeing a dramatic increase in apps, attributed to the growing importance of agentic coding. Companies like Mercedes are using agents to streamline their IT infrastructure.

  The video explicitly discusses OpenAI, Gemini, Cloud Managed Agents, Codex, Remotion, Hyper Frames, and specific tools like Cogni. It’s aimed at intermediate and advanced users who want to stay updated on the latest developments in the AI world.

Fireship (2 new videos)

• I can’t believe this trial is real…
  15.5.2026, 18:51:21

  The video offers a humorous outlook on futuristic technologies of 2026, including an AI-powered smart toilet, intelligent hair-cutting machines, and an AI pocket pet. However, the focus is on OpenClaw, a personal AI assistant system that has caused a stir since its launch in January 2026. The creator shows how to host and configure OpenClaw on a virtual private server from Hostinger to solve specific problems, such as answering family inquiries in your own voice. The process includes setting up a Telegram bot, integrating 11 Labs for voice generation, and configuring tools to generate automated responses.

  The video explicitly covers OpenClaw and Hostinger and is intended more for intermediate users, as it includes technical steps for setting up and configuring an AI assistant system.

• A single PR just hijacked the NPM registry…
  14.5.2026, 17:39:11

  The video introduces Claude Design, a new AI-powered platform from Anthropic based on the Opus 4.7 model that transforms Figma designs into interactive prototypes, pitch decks, and production-ready UIs. The AI can process high-resolution images and offers advanced features like interactive animations, design variations, and even video editing. The video demonstrates creating an iOS onboarding flow for a fictional app called “Horse Tinder,” where the AI provides a good starting point but doesn’t always apply the design system correctly. Despite some shortcomings, Claude Design shows promising capabilities that could revolutionize traditional design tools and processes.

  The video explicitly covers Anthropic’s Claude Design and Opus 4.7 as well as Google Cloud Run and is intended more for intermediate and 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)

• Claude Code from your phone working for you (Remote Control)
  16.5.2026, 13:23:09

  The video demonstrates how to install Claude Code from Anthropic on a server and control it via the Remote Control function using your phone to work on projects from anywhere. The user shows how to create a website, deploy it on a server, and continue editing it from your phone without having to sit at your PC all the time. Various commands like `Remote Control`, `Set a Goal`, and `Loop` are explained, which allow you to assign Claude Code tasks that are executed automatically until a specific goal is reached or at regular intervals. The user uses a Hostinger server with Claude Code pre-installed and demonstrates how to set up DNS records for a domain to go live with the created website. At the end, two useful commands are introduced: `Slash Goal`, to instruct Claude Code to complete a task until a certain condition is met, and `Slash Loop`, to have Claude Code execute a specific action at regular intervals.

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

• I hired Claude Code as my video editor
  11.5.2026, 17:02:17

  The video shows how to edit videos using AI tools like Claude Code, Hyperframes, and Claude Design without using traditional editing software. The process involves three main tools: Videouse, Hyperframes, and Claude Design. Videouse cuts the video based on the transcript created via 11 Labs. Hyperframes generates animations and motion graphics based on HTML, CSS, and JavaScript. Claude Design enables you to create animations and designs in the browser without writing code and integrates your branding. The workflow is explained step by step, including setup and usage of the tools. It demonstrates how to cut a video, add animations, and render the final result in 4K. The video is suitable for intermediate users who want to use AI tools like Claude Code, Hyperframes, and Claude Design.

Kyle Balmer | AI with Kyle (2 new videos)

• The 4 AI Skills Anthropic Wants You to Have (Explained in 20 minutes)
  15.5.2026, 05:00:02

  The video provides a summary of Anthropic’s “AI Fluency” course, which aims to teach effective, efficient, ethical, and secure collaboration with AI systems. The course is divided into three main use cases: Automation (task execution per instructions), Augmentation (collaborating with AI as a sparring partner), and Agency (AI working independently toward a set goal). The core of the course centers on the four “Ds”: Delegation (deciding which tasks to delegate to AI), Description (clear communication of the task), Discernment (critically evaluating AI output), and Diligence (ethical and responsible use). Particular emphasis is placed on the “Description to Discernment” loop, where AI output is continuously reviewed and refined to achieve high-quality results. The course advises against automating AI too quickly, recommending instead to first test and refine in an augmented mode. Diligence includes checking AI usage for accuracy, transparency, and verification to avoid ethical and practical errors.

  The video explicitly features Claude from Anthropic and is intended more for intermediate and advanced users.

• How I Produce 20+ Pieces of Content Daily Using AI (Full System Revealed)
  13.5.2026, 05:00:50

  # Summary: AI-powered content production at scale

  The creator demonstrates his system for producing 20+ pieces of content daily (YouTube video, livestream, newsletter, 5-8 short-form videos, carousels, LinkedIn posts) with just two people and AI agents.

  **Core Principle:** Audience is a business’s most important asset. He recommends building authority through educational content rather than pure entertainment. In two years, he built 300,000 followers and generated over $500,000 in sales to his audience.

  **The Complete Workflow:**
  1. **Topic Selection:** An AI agent (currently Hermes, previously OpenClaw) scans daily for trends/news and suggests topics—a mix of timely hooks and evergreen content.
  2. **Prep:** The creator records voice memos with his hot takes; the agent creates an outline, he provides feedback, then ChatGPT generates slides/visuals.
  3. **Livestream:** Live on TikTok, YouTube, LinkedIn, X, Facebook simultaneously (20-60 min)—essential for authenticity. The 4K clean cut is recorded separately.
  4. **Transcript & Processing:** Descript creates the transcript, which goes into his “AI Brain” (GitHub + Obsidian). A custom workflow with Claude processes it into multiple outputs simultaneously.
  5. **Long-Form Output:** Newsletter (human-edited), LinkedIn article, carousels (ChatGPT generates infographics), PDFs as DM giveaways (via ManyChat—100-300 emails/day, ~$10 per email value).
  6. **Short-Form Output:** Claude finds 12 viral hooks from the newsletter, generates 30-60-second scripts + hashtags. The creator films these himself on his phone. ChatGPT generates reel covers.
  7. **Scheduling:** Blotato or manual. Feedback loop: Each week the system compares AI drafts with final published versions to self-improve. Optional: feed content analytics for topic optimization.

  **Minimum Viable Product (for beginners):**
  Don’t livestream daily. Instead: AI interviews you via voice mode → transcript → newsletter + short-form scripts → you film the videos yourself. Not 20+ assets initially, just newsletter + a few videos.

  **Critical Human Elements:** Voice memos (captures tonality), livestream/interview (authenticity), newsletter editing (not pure AI output), filming short-form videos (no deepfakes). AI does the repetitive work, but the creator remains visible everywhere.

  **Tool Stack:** Hermes/Claude (agents), ChatGPT Images 2.0 (slides/visuals—better than NotebookLM), Descript (transcripts), Claude Cowork/Code (newsletters/scripts), Blotato (scheduling), ManyChat (email giveaways). He emphasizes: the specific tool matters less than a well-thought-out workflow; wasting time tool-hopping is a common mistake.

  **Why It Works:** AI content will flood the market → liveness, public speaking, on-camera presence become valuable. Automating these would be a mistake. Instead, use AI to handle boring stuff (outlines, thumbnails, carousel design) so you have more time for authentic, human moments. Four months to develop a system of this complexity—don’t start from scratch.

  **Result:** One-person media company instead of needing 5+ full-time team members before.

  —

  Claude, ChatGPT (Images 2.0), Hermes, Descript, Blotato, ManyChat, NotebookLM, OpenClaw — deep dive/opinion hybrid, intended for experienced creators.

Leon van Zyl (2 new videos)

• Claude Code Agent View: Parallel Agents Are Here
  14.5.2026, 10:51:58

  The video reports on a complex supply chain attack on open-source packages in the npm ecosystem, affecting over 100 packages with over 50 million weekly downloads. The attack exploited a vulnerability in Tanstack’s release process by using a pull request in a repository fork to trigger the CI/CD workflow. By using the “pull request target” option, the attacker was able to inject malicious files into the CI server’s shared cache, which later stole a valid npm publish token and published compromised package versions. The malware spread further by searching for additional npm publish tokens and infecting more packages, including those from Mistral AI, UiPath, and Open Search. The malware deeply integrated into developer environments and even employed a “dead man switch” that deleted the root folder of the infected system upon detection.

  To defend against such attacks, the video recommends using PNPM 1 or higher, which offers features like “minimum release age”, “block exotic subdependencies”, and “approved builds” to prevent installation of malicious packages. Additionally, Sentry is presented as a tool for error handling in production, working with an AI-powered agent called Seir to automatically investigate and resolve issues.

  The video covers specific tools such as PNPM, Sentry, and Seir Agent and is aimed at intermediate to advanced users.

• Codex CLI Full Tutorial: Build Your First AI Image Generator from Scratch
  11.5.2026, 11:17:21

  This video walks through building an AI image studio for creating YouTube thumbnails, posters, banners, and other graphic content step by step. The process begins with setting up a Next.js project using the Codex CLI tool, which is based on GPT-5.5 and a high reasoning level. A database is set up with Docker and Postgres, and necessary tables for user authentication are migrated.

  The focus is on creating a user-friendly interface that allows users to upload reference images, write prompts, and generate images. The developer leverages Codex’s capabilities to design and test the user interface while adhering to a predefined design system. After designing the interface, the actual functionality is implemented using the OpenAI API key for the GPT-4 Image model. The developer demonstrates how to upload reference images and assets, and how to generate thumbnails that combine these elements.

  Finally, the user interface is adjusted to ensure image generation only occurs through the dashboard rather than the homepage. The homepage is redesigned with a generated image and marketing text. The video emphasizes the efficiency and time savings achieved through using Codex and OpenAI, while noting the token limitations of the ChatGPT Plus plan.

  The video explicitly covers OpenAI (GPT-5.5, GPT-4 Image model) and Codex. It is intended for intermediate to advanced users, as it addresses advanced concepts such as Docker, Postgres, Next.js, and API integration.

Liam Ottley

No new videos in this period.

Mark Kashef (1 new video)

• Why 90% of Your Claude Skills Are Dead Weight
  13.5.2026, 20:00:29

  The video covers the challenges and best practices when using Skills in Claude Code. The author criticizes the fact that many users accumulate too many unnecessary or even harmful Skills that impair performance. He explains that Skills consume tokens and can overload context windows, leading to inefficient and error-prone interactions. The author presents eight tips for optimizing Skills:

  1. **Run it cold**: Test whether a vague prompt triggers the correct Skill.
  2. **Check the budget**: Ensure that the description is precise but not too long.
  3. **Ask user input tool**: Integrate a questionnaire to ensure the Skill has all necessary information.
  4. **Copy-related rules**: Build rules for tonality and formatting if the Skill works with text.
  5. **Rate the skill**: Let the user rate the Skill and ask for improvements.
  6. **Wrong primitive**: Check whether the Skill is the right tool for the task or if it would be better implemented as a rule or CLI command.
  7. **Audit with Claude Code Guide Agent**: Use a predefined prompt to systematically audit and optimize Skills.

  The author emphasizes the importance of progressive disclosure and good user experience to make Skills efficient and user-friendly. He also shows how to consolidate and test Skills to avoid redundancies and improve performance.

  **Final note:** The video explicitly covers Claude by Anthropic and is aimed at intermediate users who already have experience with Skills in Claude Code.

Matt Pocock (3 new videos)

• I stopped using /grill-me for coding. Here’s what I use instead:
  14.5.2026, 14:22:56

  The author introduces a new tool called “Grill with Docs,” which represents an improved version of the previous “Grill Me” skill. The original “Grill Me” skill interviews the user to build shared understanding and make decisions step-by-step. However, the author noticed that “Grill Me” often took too much time and didn’t document all relevant information. That’s why he developed “Grill with Docs,” which uses a shared language (Ubiquitous Language) from Domain-Driven Design to improve communication between user and AI. This new tool searches for a “context.md” file that documents the shared language and updates it during the session. It helps sharpen unclear language, discuss concrete scenarios, and link them with code. The author demonstrates using “Grill with Docs” in a practical example and shows how it improves efficiency and clarity of communication with AI. He emphasizes that “Grill Me” remains useful, but “Grill with Docs” is better suited for projects with a codebase.

  The video explicitly covers Claude AI models and is geared toward intermediate or advanced users.

• Anthropic’s “dedicated monthly credit” is actually a huge cut
  13.5.2026, 20:12:51

  The video addresses the problem of AI-accelerated software entropy, where codebases decay faster because changes are often made without considering the entire codebase. The author presents his method for rescuing such codebases by drawing on software fundamentals and his improved codebase architecture capability. He explains key concepts like modules, interfaces, implementations, deep and shallow modules, dependency graphs, seams, and adapters, which come from John Ousterhout’s book “A Philosophy of Software Design.” Deep modules are particularly valuable because they hide significant implementation behind a simple interface, thereby increasing maintainability and value. The author demonstrates applying these concepts to his own codebase by using Claude’s “Improve Codebase Architecture” skill to identify and implement improvements. He emphasizes the importance of tests and the need for strategic decisions by the developer, as AI agents can be tactically sound but require strategic guidance. The video targets intermediate to advanced developers who want to improve their codebases and leverage the benefits of deep modules. (Claude, Intermediate/Advanced)

• New Skills! /handoff, /prototype, /review and /writing-* | Skills Changelog
  12.5.2026, 09:56:10

  The video showcases the latest updates and improvements to an AI Skills repository that recently reached 70,000 stars on GitHub. The key new additions are two new skills:

  1. **Handoff Skill**: This skill creates a handoff document that summarizes the current conversation, enabling a new agent to continue the work. The document is stored in a temporary file and contains suggestions for skills to use in the next step. The skill is particularly useful for switching between different context windows without losing important information. There are two main usage patterns: “Fire and Forget” and “DIY Sub Agent”.

  2. **Prototype Skill**: This skill assists in creating prototypes, particularly in UI development and for creating logic prototypes for stateful applications. The skill generates multiple UI variants and allows you to test and refine them interactively before implementation.

  Additionally, several bug fixes were presented, including an improvement to the “Grill with Docs” skill through the use of XML tags to control information prioritization, and adjustments to the “To PRD” and “To Issues” skills to use the correct labels.

  There’s also a preview of two skills in development: one for writing text (Tripart Skill) and one for code reviews. The Tripart Skill helps structure text fragments into a complete story or article, while the Code Review Skill verifies compliance with coding standards and correct implementation of requirements.

  Finally, a documentation website is mentioned that contains videos for each skill and a newsletter to stay informed about the latest developments.

  The video explicitly covers Claude usage and is geared toward intermediate to advanced users.

Melvynx (7 new videos)

• J’abandonne Claude Code pour Codex… je te présente ça
  16.5.2026, 16:00:40

  The video is a detailed review and demonstration of the Codex application, which the user presents after a week of intensive use. The user explains why he switched from Cloud Code to Codex, primarily due to cost-efficiency and improved features. He shows his usage of Codex, including the use of Work-Trees for multitasking and integration with various tools like Z and Convex. The user highlights the benefits of Codex, such as the ability to use multiple agents simultaneously, a better user interface, and the possibility of using tokens more cost-effectively. He also demonstrates how he migrated his Skills and agents from Cloud Code to Codex using symlinks. The user mentions that he is preparing a masterclass on using Codex and invites viewers to use his Skills and resources.

  Final comment: The video explicitly addresses Codex and is geared more toward intermediate or advanced users.

• Better-Auth : l’outil d’authentification ultime avec Convex (évite Clerk)
  15.5.2026, 16:31:38

  The video discusses the pros and cons of Clerk, an authentication service, and compares it with Better H, an alternative solution. The author argues that while Clerk offers fast and easy integration, it can be problematic due to vendor lock-in and high costs. Better H is presented as a more flexible and cost-effective alternative that offers greater control and customization options. The author shows how he implemented a comprehensive authentication solution with Better H, including an admin dashboard and organization management, supported by AI agents. He emphasizes the benefits of Better H, such as the ability to self-host and control everything, as well as integration into his own stack, which he offers as part of his Nostac training.

  The video explicitly addresses Clerk and Better H and is aimed at intermediate to advanced users who are dealing with authentication solutions for their applications.

• LES DEVS ARRÊTENT D’UTILISER CLAUDE : c’est de pire en pire…
  14.5.2026, 15:44:37

  The video discusses growing dissatisfaction with Claude, an AI model from Anthropic, and the increasing popularity of OpenAI’s Codex. The author describes how Claude initially gained the trust of developers but is now criticized for new pricing and restrictions on the use of API tokens for programmatic applications. Particularly problematic is the introduction of a monthly credit of 200 dollars, which only covers 200 dollars in API tokens, which is perceived as insufficient.

  The author compares his experiences with Claude and Codex, highlighting how Codex convinces through its efficiency and flexibility. He describes how he switched his workflows from Claude to Codex, which gave him better results and more tokens for his money. He also praises the Codex app for its user-friendliness and multitasking capabilities.

  The author concludes with a recommendation to try Codex and shares his current stack, which includes both Claude and Codex. He emphasizes that he continues to use both services but is currently relying more on Codex.

  **Final comment:** The video explicitly addresses Claude from Anthropic and Codex from OpenAI and is geared more toward intermediate to advanced users.

• Convex : la MEILLEUR Database pour tes SaaS (j’arrête Prisma)
  13.5.2026, 16:01:23

  # Summary

  The creator reports on his switch from Prisma to Convex as a backend solution and explains why he now uses Convex daily for new projects. Unlike Prisma (an ORM for databases and migration management), Convex offers a holistic backend system: database, real-time synchronization, functions, jobs, webhooks, and authentication – all expressed in TypeScript and without needing a separate backend server.

  The core promise: **Automatic data synchronization** across all open browser tabs and devices, without having to manually manage cache invalidation, webhooks, or web sockets. A concrete example is his chat app (codelinks.dev), where messages and live drafts are updated immediately everywhere. In a multi-tab situation (e.g., changing an organization name), the change is automatically reflected everywhere – with Prisma you’d have to coordinate APIs, cache layers, query invalidation, and potential Next.js page caches.

  Convex has pre-configured integrations (Stripe, Resend, Mux, etc.) and built-in webhook management with automatic retries and delivery tracking. Jobs and workflows (e.g., delayed email sends) work natively without external tools like Vercel Cron or Redis. For development, he uses Convex in the cloud (not locally), which eliminates setup issues; AI agents can read Convex logs and fix errors autonomously.

  For deployment, Convex offers multiple environments (production, dev, preview) for free. His chat app’s pricing was recently around 26 cents per month (for 900k function calls, 2GB storage) – well below the Professional plan limits (25$ + usage). The biggest cost trap was unoptimized code that caused database scans; after optimization, the price dropped dramatically.

  Disadvantages: stronger dependence on Convex as a service (alternative: self-hosting via Docker), and migrating complex legacy projects like his email application is too risky due to the large codebase. He recommends his boilerplate project (Nost Stack), which is pre-configured with Claude + Convex CLI, to avoid common issues.

  **Demo/opinion, provider Convex (explicit), with AI integration (Claude mentioned for code generation and log analysis).**

• Pourquoi j’ai arrêté d’utiliser Next.js (je sais… tu n’étais pas prêt)
  12.5.2026, 16:00:10

  # Summary: Why I Stopped Using Next.js

  The creator shares his journey of leaving Next.js for new projects and instead opting for **TanStack Start**. After a conversation with the creator of TanStack during an event in San Francisco, he realized fundamental problems with Next.js that he had previously overlooked.

  **Core problems with Next.js:**

  He demonstrates with real applications (Ciao app, email campaigns) that he was forced to migrate from Next.js to React Router to enable instant navigation. With Next.js Router, this was impossible – instead, he had to set `use client` everywhere and lost the benefits of server-side rendering.

  **Measurable advantages of TanStack Start:**
  – Build time: 44 seconds instead of 1:56 minutes (with Next.js)
  – Local dev startup: 2-3 seconds instead of 10-12 seconds
  – Navigation between pages: significantly faster (demonstrated through live comparison)

  **Complexity and AI issues:**
  Next.js’s server components, client components, and multi-layered caching mechanisms (Full Route Cache, Data Cache, Router Cache) are difficult for humans to debug – even more so for AI models (Lia). TanStack Start is based on standard web patterns (React Router), which enables AI models to implement features correctly in one shot instead of constantly making mistakes.

  **Cost issues:**
  He showed that server-side rendering for frequently visited pages (his training pages) cost around 5 dollars monthly – for static content. After migrating to a React app shell with client-side rendering, his Vercel bills dropped from ~100 to ~50 dollars because the computational effort per navigation dropped drastically.

  **Caching nightmare:**
  With Next.js every navigation = new server render. With React app shell: only once at initial load; afterward only data fetching. This reduces server compute time by a factor.

  **Convex fits better:**
  Convex with Next.js doesn’t make sense – you’re rendering client-side in a framework that demands server-side. TanStack Start + Convex is the clean setup.

  He plans a new boilerplate (NoStack) with TanStack Start, Convex, and Better Auth, offering beta access. He’s keeping old apps on Next.js out of laziness but wants to at least migrate Subfast.

  **Conclusion:** Client-side rendering is faster, cheaper, simpler, and less error-prone – especially for dashboard apps. Next.js was a good learning phase for him (he made money from training videos about it), but for AI-driven development today, it’s clearly obsolete.

  —

  Demo of TanStack Start and Next.js (Vercel deployment); code inspections; no specific AI tools mentioned by name except “Lia” and Claude — **opinion/reflection**.

• Je change complètement ma stack… introduction de Now-Stack (secret project)
  11.5.2026, 19:00:01

  # Summary: NStack – New Development Stack Announcement

  The creator presents **NStack**, a new technology stack for developing SaaS applications, which he will detail in a video series starting this week.

  The stack consists of three main tools:

  **TanStack Start** replaces Next.js as the framework. The creator cites concrete reasons: build time is approximately 2.2x faster than Next.js, less “magic” in the code (no complex distinction between server/client components), and routing runs client-side, resulting in instant navigation without loading indicators. The demo shows navigation between pages and admin areas without noticeable delay, while the Next.js comparison version requires 2 seconds of loading time for page transitions.

  **Convex** serves as database and backend: It manages data storage, real-time synchronization between clients, API requests, authentication, webhooks (e.g., from Stripe), and asynchronous jobs. The advantage is that all data updates are synchronized in real-time across all browser tabs and users – demonstrated by simultaneously updating organization data in two browsers. Convex thus replaces multiple separate tools that would otherwise be necessary.

  **Better Auth** is the third component (details remain vague as the creator says he’ll present them later).

  The boilerplate includes an extended admin dashboard with user management (filtering, roles, session overview), organization management, and Stripe integration.

  The creator emphasizes **automation through CLI commands**: `skill init project` initializes the complete project, `skill publish to production` deploys in one step to Vercel with Convex database, and `skill setup stripe` configures Stripe with just API keys – without manual clicking. Additional skills are planned, including one that walks through planning a SaaS feature for over an hour.

  The vision: developers should create SaaS applications from scratch to production primarily through commands. The stack is supported by Cursor, Codex, and Claude.

  A waitlist is available at mlv.sh/fn; details follow in the video series this week.

  **Announcement video for a new SaaS development stack (TanStack Start, Convex, Better Auth), focusing on automation through CLI commands; the creator works with LLM tools like Cursor and Claude.**

• Mon SaaS IMPOSSIBLE à vibe-coder : ce que personne ne vous dit sur le vibe-coding
  10.5.2026, 16:00:01

  # Summary

  The creator showcases the most complex application he has ever built: an email marketing SaaS called Luemail. The platform includes a dashboard with metrics on sent emails, subscribers, and engagement. The core functionality is creating and sending campaigns with extensive filtering options (by clicks, opens, workflow status, etc.), an email editor with templates and code output, plus snippets that automatically update everywhere when changed.

  The technical complexity lies in the infrastructure: The Next.js app sends emails via AWS SES, uses Upstash for rate limiting (50 emails per second) to respect AWS limits. To handle ~500 events per second (opens, clicks, bounces), Upstash Redis is used as a queue with batch processing workers to avoid overwhelming the database. Email HTML is stored in R2 with 30-day expiration. The platform offers an admin dashboard, user profiles with custom fields and tags, automated workflows, plus a full API and npm SDK with 60 tools. A CLI enables agent-based interactions; Claude can autonomously create and manage campaigns.

  The creator emphasizes: Claude (then called Lia) doesn’t think ahead and architecturally like a senior developer – it solves the immediate problem, not scalability. The standard UX is uninspiring and requires extensive rework. With ~28,000 emails, 477 campaigns, and 55,000 subscribers on the main account, the complex architecture has proven itself, but only through repeated production crashes during the learning process (e.g., in Bali). Customers are currently only personal acquaintances; broad marketing is avoided due to IP reputation risks in email sending.

  Claude (referred to as Lia back then) was started with Cursor, then switched to Claude — demo of a complex production SaaS application with technical deep dive into infrastructure and agent integration.

n8n (1 new video)

• n8n Desk: The Open Source Claude Desktop That Runs on top of n8n’s new MCP
  11.5.2026, 18:03:38

  The video showcases a demonstration by Marcel, who introduces two new tools: the official N8N MCP (Meta-Cognitive Protocol) and N8N Desk, an open-source desktop application that functions similarly to Claude Desktop, but with some important differences. The N8N MCP enables AI agents not only to execute N8N workflows, but also to create and edit them. This is a major advancement, as previously it was only possible to run workflows. Marcel demonstrates the creation of a support ticket agent that communicates via Slack. N8N Desk, on the other hand, doesn’t run code locally on your computer, but instead leverages N8N instances in the cloud, making it more secure and better suited for enterprise environments. The application supports various AI models, including local models, and offers features like Co-Work, Workflow, and Chat. Marcel emphasizes the importance of context and Skills for effective AI agent utilization and shows how N8N Desk integrates these aspects. The video concludes with an announcement that N8N Desk will be available as an open-source project, with the option to be notified via a waitlist.

  The video explicitly covers N8N, Claude, and open-source tools and is geared more toward intermediate and advanced users.

Nate Herk | AI Automation (5 new videos)

• How to Deploy Your Claude Automations (3 Methods)
  15.5.2026, 15:16:02

  The video explains three methods to deploy agents from Anthropic’s Cloud Code environment (Claude) so they run even when the user is not active. The methods are compared using a schema that answers the question of where the agent runs (locally or in the cloud) and how autonomous/agentic it operates.

  1. **Loops:**
  – Simple method where Cloud Code is instructed to create a loop that executes a specific task at regular intervals (e.g., every 10 minutes).
  – Uses internal tools like `cron create`, `cron list`, and `cron delete` for scheduling.
  – Loops are session-specific and run either in the desktop app or in the terminal.
  – **Advantages:** No additional setup needed, full agent functionality within the session.
  – **Disadvantages:** Session and computer must be running, maximum runtime of 7 days, fixed intervals with random delay (jitter).

  2. **Desktop Scheduled Tasks and Cloud Routines:**
  – **Desktop Scheduled Tasks:** Run locally on the computer and require the desktop app to stay open.
  – **Cloud Routines:** Run in Anthropic’s cloud and don’t require an active session or running computer.
  – Both methods inject a prompt into a Cloud Code session and execute the task.
  – **Advantages:** No additional infrastructure needed, full Cloud Code functionality, Cloud Routines can also be triggered via API or GitHub events.
  – **Disadvantages:** Cloud Routines have a minimum of 1 hour between executions, limited number of executions per day (depending on plan).

  3. **Deployment on Modal or Trigger.dev:**
  – Here, a script (Python for Modal, TypeScript for Trigger.dev) is deployed to the respective cloud platform and runs there on a schedule or as an API endpoint.
  – **Advantages:** No need to keep your own computer or session running, good for deterministic processes.
  – **Disadvantages:** No full agent functionality, AI processing occurs via API and is thus more expensive. With Claude’s Agent SDK you can also use agentic functions, but this is also more expensive.

  Additionally, **Managed Agents** from Anthropic and **Hooks** in Cloud Code are briefly mentioned, which can be used for event-driven automations.

  The video is better suited for **Intermediate** users who already have experience with Cloud Code and agents. It addresses specific tools and providers, including **Claude (Anthropic)**, **Modal**, **Trigger.dev**, and the **Claude Agent SDK**.

• Anthropic Just Dethroned OpenAI. Here’s What Happens Next.
  13.5.2026, 21:20:51

  The video discusses the current dynamics in AI-powered coding tools, particularly the competition between OpenAI (Codex) and Anthropic (Claude Code). It begins by noting that Anthropic has surpassed OpenAI in business usage, followed by OpenAI’s counter-move offering companies two months of free Codex usage. Claude Code responded with a 50% increase in weekly usage limits for the next two months. The author interprets these moves as part of a “free trial phase” where companies and users use the tools intensively while providers aim for adoption and data collection. The argument is made that current pricing is not sustainable and that users ultimately provide valuable training data for the AI models. The author advises leveraging current offers to avoid lock-in and remain flexible for possible future changes. Historical patterns are also referenced where similar dynamics in other industries led to price increases. The closing comment emphasizes that users should use the tools to avoid committing to a single provider.

  **AI tools/models/providers:** OpenAI (Codex), Anthropic (Claude Code); **Target audience:** Intermediate to Advanced.

• Every Level of Claude Explained in 21 Minutes
  12.5.2026, 13:59:35

  The video provides a detailed guide to using Claude across five ascending levels: Enthusiast, Beginner, Intermediate, Advanced, and Architect. Each level is defined by specific features and strategies that progressively guide the user from basic applications to complex automations.

  – **Level 1: Enthusiast** – Basic use of Claude for simple tasks like writing emails or explaining content. An important tip is using screenshots, as Claude can read images.
  – **Level 2: Beginner** – Introduction of Projects, which provide context and continuity. Key features are Memory, Connectors (integration with tools like Slack or Google Drive), File Creation (creating Excel files, PowerPoint presentations, etc.), Artifacts (interactive applications), Inline Visuals (visual representations within chats), and Office Add-ins (integration with Microsoft Office).
  – **Level 3: Intermediate** – Using Claude Co-work for tasks on your own computer. Key features are File System Access, Skills (reusable workflows), Scheduled Tasks, Mobile Control (smartphone control), Cloud Design (designing prototypes and presentations), as well as Plugins and Computer Use (navigating in apps).
  – **Level 4: Advanced** – Using Claude Code for complex automations and parallel work processes. Key features are using claude.md (configuration file), Plan Mode (planning and executing tasks), Sub Agents (specialized agents), Work Trees (isolated work areas), MCP (Model Context Protocol for tool integration), as well as various optimization techniques and Custom Commands.
  – **Level 5: Architect** – Creating fully autonomous systems that work without user interaction. Key features are Cloud Routines (scheduled tasks in the cloud), Hooks (security-relevant logic), Channels (controlling external platforms), Headless Mode (autonomous task execution), Agent SDK (creating your own products), and Remote Control (remote control of sessions).

  The video emphasizes that the transition to the highest level is less about technical skills and more about trust in the systems. It’s recommended to start with simple, low-stakes automations and gradually implement more complex systems.

  **Closing comment:** The video explicitly addresses Claude and is intended for Intermediate to Advanced users.

• Claude Code Just Got an Agent Dashboard
  12.5.2026, 01:06:06

  The video introduces the new “Agent View” feature of Cloud Code, which allows managing multiple agent sessions in a single terminal tab. Users can switch between sessions, monitor their status (e.g., “waiting for input”, “completed”), and navigate directly in and out of sessions. This is particularly helpful for developers who previously struggled with many open terminal tabs and confusing sessions. Control is via arrow keys or by clicking, and new sessions can be started with simple commands or direct task descriptions. The video also mentions the new “Goal” feature, which allows setting long-term goals for Claude Code to pursue autonomously. The video also shows how to delete sessions or provide input directly from the Agent View. At the end, a course is referenced that shows how to use Claude Code as your own operating system.

  The video explicitly addresses Claude Code and is intended for Intermediate or Advanced users who already have experience using Cloud Code or similar tools.

• Hermes Agent: Zero to Personal AI Assistant (1 Hour Course)
  10.5.2026, 02:42:25

  The video is a tutorial on setting up and using Hermes Agent, a powerful open-source AI agent from Noose Research. The creator walks viewers through the entire process, from installation on a VPS (Virtual Private Server) to configuration and usage of the agent. Here are the main steps and the end result:

  1. **Introduction to Hermes Agent**:
  – Hermes Agent is a self-improving AI agent that grows and adapts with the user.
  – The agent comes with 684 skills, of which 91 are pre-installed.
  – The agent can automate various tasks, such as generating Excel diagrams, transcriptions, and answering questions.

  2. **Comparison with Other Agents**:
  – Comparison of Hermes Agent with Cloud Code and OpenClaw.
  – Hermes Agent is lighter, faster, and focuses on self-improvement.
  – Cloud Code is primarily used for knowledge work and coding, while Hermes Agent is used for automating and managing tasks on the go.

  3. **Installation and Setup**:
  – The creator demonstrates how to install Hermes Agent on a VPS at Hostinger.
  – A step-by-step guide is provided, including setting up a Telegram bot for interacting with the agent.
  – The creator emphasizes the importance of security and shows how to manage API keys securely.

  4. **Using Hermes Agent**:
  – The creator demonstrates how to use the agent for various tasks, such as creating videos with Hyperframes and automating daily tasks.
  – It shows how to sync the agent with a GitHub repo to create backups and save work.

  5. **Self-Improvement Loop**:
  – The creator explains the self-improvement loop of Hermes Agent, where the agent learns from experiences and continuously improves.
  – It’s emphasized that users should correct the agent and provide feedback to improve accuracy and efficiency.

  6. **Advanced Features and Tips**:
  – The creator provides tips for using the agent, such as creating Skills and automating tasks with Cron jobs.
  – It shows how to use the agent for various tasks and projects, such as managing YouTube comments and creating reports.

  7. **Scaling and Managing Multiple Agents**:
  – The creator discusses scaling and managing multiple Hermes Agents on a VPS.
  – It explains how to host agents in separate containers and how to make them communicate with each other.

  8. **Resources and Community**:
  – The creator offers a free resource guide containing all steps and information from the video.
  – A reference is made to the creator’s free training community where additional resources and documents are available.

  **Closing comment**: The video explicitly addresses Hermes Agent, OpenAI, and specific tools like Hostinger and GitHub. It’s intended for Intermediate and Advanced users who already have experience using AI agents and managing servers.

NeuralNine (3 new videos)

• This is the most useful n8n workflow I have ever built…
  15.5.2026, 16:00:28

  The video demonstrates how to create a fully automated AI-coding agent using N8N, Claude Code (Anthropic), OpenAI, and GitHub. The agent can be triggered via Slack or GitHub Issues to make changes to a repository, commit them, and create a pull request. The process involves two N8N workflows: one for GitHub Issues and one for Slack messages. The first workflow is triggered by creating a GitHub Issue, executes a Python script that uses Claude Code in headless mode to implement the changes, and sends a success notification via Slack. The second workflow takes a Slack message, converts it to a GitHub Issue using an OpenAI agent, and triggers the first workflow. The video covers N8N setup, workflow creation, and integration of GitHub, Slack, and the AI agents.

  The video specifically addresses N8N, Claude Code (Anthropic), OpenAI, GitHub, and ngrok. It is better suited for intermediate and advanced users.

• JavaScript Crash Course: Full Beginner Tutorial
  13.5.2026, 16:01:18

  The video offers a comprehensive introductory course in JavaScript that focuses on the fundamentals of the programming language rather than its use in HTML for animations. It covers the following topics:

  1. **Introduction and Setup**:
  – Installing Node.js to run JavaScript files.
  – Creating a simple “Hello World” application with `console.log`.

  2. **Variables and Data Types**:
  – Defining variables with `let`, `var`, and `const`.
  – Differences between `let` and `var` (block vs. global scope).
  – Basic data types such as numbers, strings, booleans, and special values like `undefined`, `null`, `Infinity`, and `NaN`.

  3. **Operators**:
  – Arithmetic operators (addition, subtraction, multiplication, division, modulo, exponentiation).
  – Assignment operators and increment/decrement operators.
  – Comparison operators and logical operators.

  4. **Conditional Statements**:
  – `if-else` statements and `switch-case` statements.
  – Processing user input with the `readline` library.

  5. **Loops**:
  – `while` and `do-while` loops.
  – `for` loops, including iterating over arrays.
  – Loop control statements like `continue` and `break`.

  6. **Functions**:
  – Defining and calling functions.
  – Parameters and return values.
  – Anonymous functions, lambda expressions, and functions with variable numbers of parameters.
  – Functions as first-class objects and callbacks.

  7. **Data Structures**:
  – Arrays: creating, accessing elements, and iteration.
  – Maps (dictionaries): key-value pairs and accessing values.
  – Sets: simple collection of unique values.

  8. **Objects and Classes**:
  – Defining objects using shorthand notation and classes.
  – Using constructors and methods.

  9. **JavaScript in HTML**:
  – Embedding JavaScript in HTML files and interacting with the DOM.

  The video is suitable for beginners with basic programming knowledge who want a quick overview of JavaScript as a programming language. No specific AI tools or models are discussed.

• ONNX: The PDF Format For Neural Networks
  11.5.2026, 16:00:44

  The video introduces the ONNX format (Open Neural Network Exchange), which serves as a PDF-like format for neural networks. The creator shows how to train models in PyTorch and TensorFlow, export them to ONNX format, and then use them with ONNX Runtime without the original frameworks. Both simple and more complex models, such as an MNIST classifier, are demonstrated. The video also covers how to download and use models from Hugging Face in ONNX format. The focus is on reducing dependency on specific frameworks and increasing interoperability.

  **Final Comment:** The video addresses ONNX, PyTorch, TensorFlow, Hugging Face, and ONNX Runtime and is suitable for intermediate users.

Nic Conley

No new videos in this period.

Nick Saraev (1 new video)

• How to Build Mobile Apps with Claude Code: Full Course (2026)
  11.5.2026, 17:51:20

  **AI Tools/Providers covered:** Claude, OpenAI, Gemini, Open-Source, specific tools like Cursor, Lovable, Cline, n8n

  **Target audience:** Intermediate

Niklas Steenfatt

No new videos in this period.

No Priors: AI, Machine Learning, Tech, & Startups (2 new videos)

• Pax Silica: Inside the Trump Administration’s Tech Strategy with Jacob Helberg
  14.5.2026, 10:01:32

  The video is an interview with Jacob Helberg, Under Secretary of State for Economic Affairs, discussing the “Pax Silica” initiative. This initiative is an economic security coalition aimed at securing AI supply chains for the US and its allies. Helberg explains that Pax Silica pursues an ecosystem-based approach and already includes 14 countries. A central project is establishing a “forward-deployed industrial base” in the Philippines, where 4,000 hectares of land will be used for industrial development. This zone is intended to offer a combination of the American legal system and Philippine industrial advantages.

  Helberg emphasizes the importance of collaboration with private industry, as the US does not want to rely on state-run supply chains. Instead, the initiative aims to create commercially viable platforms that can function independently from the state. The initiative seeks to reduce dependence on China and diversify supply chains, particularly in areas like robotics and rare earth elements.

  The video also addresses differences between the American strategy and China’s Belt and Road Initiative, with Helberg highlighting the advantages of the American approach, which emphasizes partnerships with private industry and ethical business practices. He stresses the importance of innovation and the role of private capital in solving supply chain issues.

  The video is more suitable for intermediate and advanced viewers and does not focus on specific AI tools or models, but rather on political and economic strategies.

• Amex Global Business Travel: The World’s First AI Take Private with Long Lake CEO Alexander Taubman
  11.5.2026, 10:01:04

  **Summary:**

  Long Lake Management, a company specializing in acquiring and transforming businesses through AI, recently announced its intention to acquire American Express Global Business Travel (AmEx GBT) for $6.3 billion. This would be the first AI-driven acquisition of a publicly traded company. Long Lake has already acquired 30 companies and transformed them using their Nexus AI platform. This platform automates workflows, improves employee productivity, and drives company growth.

  Alex Topman, CEO of Long Lake, explains that the Nexus platform shares approximately 80% of infrastructure across various industries and can be deployed within days of an acquisition. This leads to immediate margin increases and growth. Long Lake focuses not on cost-cutting but on growth and customer experience. Through AI-driven automation, employees can serve more customers, resulting in higher organic growth rates.

  Topman emphasizes that Long Lake does not merely sell software but acquires companies to build deeper business relationships and achieve better results. This enables closer collaboration between employees and engineers, leading to better solutions and faster innovation. Long Lake aims to be the best employer in every industry where it operates, attracting talented staff in the process.

  The AmEx GBT acquisition exemplifies Long Lake’s long-term strategy of acquiring companies and holding them long-term rather than selling them short-term. Topman sees AI as a positive-sum game that benefits both employees and customers.

  **Final Comment:**
  The video covers Long Lake Management and their Nexus AI platform, used for business transformation. It is more suitable for intermediate and advanced viewers, as it focuses on complex business strategies and AI applications.

Productive Dude

No new videos in this period.

Sebastien Dubois (1 new video)

• Levels of AI Context Management: From Generic Answers to Precision Results
  16.5.2026, 08:48:32

  The video discusses the various levels of context management when using AI, particularly when interacting with language models like GPT or Claude. The speaker distinguishes eight levels, ranging from simple queries without context (level 1) to a comprehensive knowledge system that can be fully utilized by AI (level 8). In the early stages (1-3), the focus is on providing basic context and leveraging the built-in memory functions of AI, which are described as unreliable and pseudo-random. From level 4 onwards, advanced prompt engineering techniques are introduced, which provide better results but also come with challenges such as organizing and managing prompts. From level 5 onwards, the importance of a central knowledge system is emphasized, containing information about projects, goals, and work processes. This enables AI to deliver more precise and relevant results. In the higher levels (6-8), the focus shifts to managing AI memory, implementing AI agent capabilities (Skills), and optimizing context management to improve consistency and quality of results. The speaker emphasizes that a well-structured knowledge system that can be fully understood and utilized by AI currently delivers the best results.

  The video explicitly addresses Claude, OpenAI, and specific tools such as Obsidian, Anthropic, Google, and OpenAI. It is aimed at intermediate and advanced users.

Tech With Tim (3 new videos)

• AI Web Scraping Is Insanely Good | Browserbase Full Tutorial
  15.5.2026, 15:19:07

  The video introduces Browserbase, a platform that enables AI agents to connect with remote browser sessions to perform web automation. Browserbase offers an alternative to traditional tools like Playwright or Selenium by providing an AI-powered interface that uses natural language to automate browser actions. The platform solves many typical web automation problems, such as IP bans, rate limits, and captchas, by using an agent identity that works with Cloudflare and other captcha providers. The video shows a step-by-step guide on how to integrate Browserbase with AI agents like Claude Code or Open Claw, as well as code examples in Python and TypeScript demonstrating the use of the Stagehand SDK. Various features are shown, such as extracting information, observing actions, and executing complex workflows. The advantages of Browserbase are highlighted through comparisons with traditional automation tools, particularly the ability to reliably automate dynamic and complex websites.

  The video explicitly focuses on Browserbase and Stagehand SDK tools and is geared more toward intermediate and advanced users.

• Claude Tutorial – How to Connect Claude to ANY Tool
  12.5.2026, 17:32:33

  # Summary

  The video shows how to connect AI agents like Claude or Copilot with external applications (Gmail, Google Drive, HubSpot, Salesforce, etc.) – not through native connectors, but through **Compose** (free).

  **Problem with native connectors:** When you enable multiple native connectors directly in Claude, “context bloat” occurs – hundreds of tools are passed to the LLM in the system prompt, leading to higher token costs and worse accuracy in tool selection. Additionally, you can only connect one account per tool.

  **Compose solution:** Uses “On-Demand Tool Discovery”. Instead of providing 500 tools, the agent is shown only 1-3 search tools. The agent uses these to semantically search for the right tool in Compose, keeping the context small and improving accuracy. Compose manages all authentications centrally.

  **Setup:** You connect apps in the Compose dashboard (Gmail, Twitter, HubSpot, etc.), receive a URL, and add it as a custom connector in Claude/Copilot. After authorization, the agent can immediately access all connected tools – and is even prompted to create new connections if needed.

  **Demo:** The agent summarizes the last three emails and then creates a Google Doc with the summary, saves it to Drive, and sends the link – all orchestrated through Compose tools.

  **Pricing:** 20,000 tool calls/month free; upgrades to 200,000 or 2 million available. The author uses the tool across multiple accounts without ever exceeding the free limits.

  **Practical:** You can create multiple organizations and selectively enable tools – one setup works for all AI tools where you integrate Compose.

  —

  *Compose is mentioned as a sponsor but is free; Claude and Copilot (as well as theoretically other AI agents) discussed as platforms; format: tutorial/demo.*

• Claude Just Got a Superpower No One’s Talking About
  11.5.2026, 14:01:32

  # Higgs Field MCP Server in Claude: Complete AI Education and Video Generation

  The video shows how to install and use the Higgs Field MCP Server in Claude to generate images and videos directly within the AI tool – without having to connect external APIs individually.

  **Installation:**
  Setup is done through Claude settings under “Connectors”, where you add a custom connector with the Higgs Field MCP Server URL and authenticate. Alternatively, you can install the server via CLI in Claude Code by installing the Higgs Field CLI tools, logging in, and adding the skill.

  **Practical Application:**
  The demo creates an ad campaign for a fictional coffee subscription brand called “Focus Brew”. Claude first generates three product photos with GPT Image 2 (flat, lifestyle, close-up), edits one based on feedback, then creates five-second video ads with Seedance 2.0 from different angles (productivity, ritual, skeptic perspective) and generates counter-narrative videos that address common customer objections. The generated assets are then integrated into an HTML landing page.

  **Core Advantage:**
  Instead of connecting and paying for 30+ different tools individually, Higgs Field provides a unified interface with access to multiple image and video models. The MCP integration enables automated multi-step workflows: Claude can independently generate images, evaluate them, and create videos – or use Open Claw to automate recurring tasks (e.g., generating 10 new images daily).

  **Special Feature:**
  The AI agent automatically refines the original prompts with details like quality, resolution, and aspect ratio, which usually yields better results than manual creation.

  Claude and Higgs Field MCP Server as central integration; tutorial/demo with pragmatic workflow.

TheAIGRID (3 new videos)

• 10 Best AI Tools For Small Business Owners in 2026 – How To Start Your Business With AI
  16.5.2026, 21:15:02

  The video introduces 10 AI tools suitable for small businesses and beginners without technical background. Here are the key points:

  1. **ChatGPT**: A versatile tool for tasks like email writing, brainstorming, document summarization, social media posts, and more. It’s particularly useful for small businesses to save time and overcome creative blocks. The Plus plan costs $20 per month.

  2. **Claude Co-work**: A desktop agent from Anthropic that can autonomously work with local files and applications. It can handle tasks like organizing files, creating presentations, and managing Excel spreadsheets. The base price is $20 per month.

  3. **Microsoft 365 with Claude**: Claude can be integrated into Microsoft Word and Excel to automate tasks like creating financial analyses and managing documents. It’s particularly useful for creating financial forecasts and managing data.

  4. **Copilot in Microsoft 365**: A built-in AI in Microsoft 365 that offers similar features to Claude but with less capability. It’s a cost-effective alternative for businesses with limited budgets.

  5. **Gamma.app**: A tool for creating presentations and websites. It’s particularly useful for small businesses that want to create professional presentations and websites without design knowledge.

  6. **Google’s Pomelo**: A free marketing tool from Google that generates social media posts, advertising campaigns, and email banners based on a website URL. It’s particularly useful for businesses looking to strengthen their brand identity.

  7. **Flow by Google**: An image generation tool that’s particularly useful for businesses that need to create images regularly for marketing purposes. It offers 50 free credits to get started.

  8. **Canva AI**: A tool for creating marketing visuals that suggests layouts, generates images from prompts, and adapts designs for different social media platforms. It’s particularly useful for businesses that want to maintain consistent brand identity.

  9. **Zapier**: A tool for automating workflows between different applications. It can automate tasks like creating calendar entries and managing customer inquiries. It offers a free version and paid plans up to $20 per month.

  10. **Chatbase**: A tool for integrating AI support agents into websites. It allows businesses to integrate automated customer support agents on their websites that can respond based on PDFs or Word documents containing company knowledge. It offers a free version and paid plans starting at $32 per month.

  11. **Perplexity Computer**: An expensive but powerful AI agent tool that can automate tasks like creating reports, scraping websites, and managing data. It’s particularly useful for businesses looking to automate complex tasks. The price is $200 per month.

  12. **Lovable**: A tool for creating websites with AI support. It’s particularly useful for businesses that want to create a website quickly and easily without design knowledge.

  The video is suitable for small businesses and beginners without technical background. It covers various AI tools and models, including ChatGPT, Claude, Microsoft 365, Copilot, Gamma.app, Google’s Pomelo, Flow by Google, Canva AI, Zapier, Chatbase, Perplexity Computer, and Lovable.

• Meta AI Tutorial – How To Use Meta AI
  13.5.2026, 21:15:04

  The video presents the features and tips for using Meta AI. It begins by explaining the three chat modes: “Instant” for quick, casual questions, “Thinking” for complex problems requiring extended reasoning, and the secret “Contemplating” function that deploys up to 16 independent agents for parallel analysis. This function is particularly useful for research, strategy development, and difficult decisions. Next, Meta AI’s search function is demonstrated, which pulls information from the open web and Meta’s social graph, including Instagram Reels and Facebook posts. Meta AI’s multimodality is highlighted, particularly image analysis and visual anchoring, which can provide detailed analyses and personalized health information from images. Document analysis including PDFs, Excel, and Word files is also mentioned. Additionally, image and video generation with Meta’s “Vibes” is introduced, currently powered by Midjourney, as well as the Meta Vibes social platform where users can share and remix AI-generated content.

  The video explicitly covers Meta AI and is geared more toward intermediate users.

• Google’s New Gemini Omni Just Shocked Everyone – Leaked Demo, Pricing, and what comes next
  12.5.2026, 19:45:04

  The video covers the leak of a new Google AI video model called Gemini Omni, which was accidentally made accessible to some users. The discovery began with a tweet pointing to an unusual text line “powered by Omni” in the Gemini app. Users generated initial videos that demonstrated impressive quality comparable to current top models like Sora 2. The videos showcase complex scenes such as a professor explaining a mathematical proof on a blackboard and a detailed restaurant scene with two people. Using the model is highly resource-intensive, with just two generations consuming 86% of a Pro subscription’s monthly quota ($20/month). The name “Omni” suggests a multimodal model supporting various input and output modalities, similar to the never officially released GPT-40. The video speculates about Omni’s potential launch at the upcoming Google IO and compares it with current competitors like Sora 2, Alibaba’s W2.7, and Cline 3.0. It discusses whether Omni could be a standalone product or an evolution of VO4, with the potential to significantly change the AI video landscape.

  The video explicitly covers Google’s Gemini Omni and compares it with other models like Sora 2 and GPT-40, targeting an intermediate to advanced audience.

Theo – t3․gg (5 new videos)

• AI Is Hacking Everything Now…
  15.5.2026, 20:09:07

  The video discusses a range of serious security vulnerabilities and exploits in various software and operating systems, particularly Linux. It highlights several critical security issues, including “Copy Fail,” “Dirty Frag,” and an attack on 84 Tanstack npm packages. The speaker emphasizes that the speed and frequency of these exploits are alarming and that traditional security practices, such as the 90-day disclosure deadline, are no longer sufficient. He argues that the use of AI has accelerated and simplified the discovery and exploitation of security vulnerabilities, leading to a new era of software security.

  The speaker proposes that the industry must make drastic changes to address these new challenges. This includes introducing a new tier of trusted actors who receive early notification of security vulnerabilities, as well as fundamental changes to open-source culture to control and delay code disclosure. He also emphasizes the need to build software more securely from the ground up by adopting safer programming languages like Rust and developing better architectures.

  The speaker also shares personal measures he is taking to protect himself from these threats, such as regularly creating backups and educating his family about security risks. He concludes with the hope that his warnings will be taken seriously and that the industry will work together to find solutions to address growing security challenges.

  AI Tools/Models/Providers: Claude, OpenAI, Gemini, Open Source (Linux, npm packages, Tanstack), Mythos, OpenAI Daybreak, GPT-4.7, Claude Opus 4.7, Gemini 31 Pro, GPT-5.5

  Target Audience: Intermediate to Advanced.

• I’m done.
  14.5.2026, 10:08:59

  The video covers Anthropic’s announcement that, starting June 15, paid Claude plans will receive dedicated monthly credits for programmatic use. These credits cover use of the Agent SDK, Claude-P, and third-party apps based on the Agent SDK. The speaker is disappointed because this change severely restricts Claude usage in third-party tools like T3 Code, OpenClaude, and others, reducing usage limits by 25 to 40 times. He criticizes Anthropic for failing to provide clear communication despite months of requests and presenting a solution that feels more like a restriction than an improvement. The speaker emphasizes that this is an attack on open-source projects and that Anthropic is severely limiting open and programmatic use. He announces that T3 Code will introduce an option to choose between the enhanced user interface and the original, but limited, usage. He advises viewers to use alternative solutions like Codex and expresses his disappointment with Anthropic’s actions.

  **AI Tools/Models/Providers:** Claude (Anthropic), OpenAI, Codex, OpenClaude, T3 Code, Agent SDK, Claude-P, Open Source
  **Target Audience:** Intermediate to Advanced

• Stop letting your agents write Markdown.
  13.5.2026, 06:46:31

  The video discusses the pros and cons of Markdown compared to HTML, particularly in the context of use by AI agents. The author argues that while Markdown is simple and portable, it is often too restrictive and should be replaced with HTML, which enables richer and more interactive content. He references articles by Thoric and Andrej Karpathy, which share similar views, and shows examples of how HTML can be used for various tasks such as planning, code reviews, and designs. The author emphasizes that HTML offers higher information density, is easier to read and share, and enables interactive elements. He concludes with the thesis that the future of human-machine interaction could lie in interactive videos and simulations, and encourages viewers to use HTML as a starting point and continue exploring.

  The video explicitly addresses Claude Code, Anthropic, OpenAI, and specific tools like Copilot Kit. It is geared toward intermediate to advanced users who engage with AI agents and their output formats.

• I wish this was clickbait
  12.5.2026, 10:52:14

  The video discusses the future of Bun, a JavaScript toolkit originally written in the Zig programming language. The speaker, a long-time Bun supporter, expresses concerns about Bun’s stability and future, particularly on Windows, and mentions that some developers like Dax from Open Code have switched to Node.js due to these issues. A central theme is the Bun team’s decision to rewrite Bun in Rust to address stability issues and improve maintainability. The speaker highlights that while this reimplementation is promising, it also brings new challenges, particularly due to extensive use of “unsafe” code in Rust, which points to potential security and stability risks. He also discusses the integration of Bun into Claude Code, a tool from Anthropic, and the possible impact of Claude Code’s current problems on Bun’s development. The speaker concludes with hope that the Rust reimplementation of Bun will be successful, though he remains skeptical about short-term benefits.

  The video explicitly addresses the AI models and tools Claude and Mythos and is geared toward intermediate to advanced users.

• We all fell for it…
  11.5.2026, 06:45:20

  The video discusses the impact of AI code tools on developers’ skills and working methods. The speaker shares personal experiences and reflects on the advantages and disadvantages of these tools, particularly regarding productivity, cognitive atrophy, and dependence on AI.

  Key points of the video:

  1. **Productivity gains**: AI code tools enable developers to work faster and more efficiently by automating repetitive tasks and solving complex problems.

  2. **Cognitive atrophy**: Excessive use of AI tools can cause developers to lose their skills as they engage less deeply with code details. This is referred to as “cognitive debt.”

  3. **Dependency and vendor lock-in**: Developers can become dependent on specific AI tools, which can cause problems if those tools become unavailable or costs increase.

  4. **Quality of work**: AI tools can improve code quality by reducing errors and suggesting better solutions. However, they can also lead to superficiality when developers don’t fully understand the proposed solutions.

  5. **Working method**: The speaker emphasizes the importance of using AI tools to improve one’s own work rather than relying entirely on them. He suggests using AI tools for planning, brainstorming, and research while handling implementation personally.

  6. **Vendor lock-in and reliability**: The speaker argues that developers should not be dependent on a single AI tool. He recommends using various tools and providers to increase reliability and flexibility.

  7. **Long-term impact**: The speaker warns that excessive use of AI tools can negatively impact developers’ careers in the long term if they don’t continue developing their skills.

  The video explicitly addresses AI tools and models such as Claude, OpenAI, Gemini, and specific tools like Cursor, Browserbase, and T3 Code. It is aimed at intermediate and advanced developers who want to understand the impact of AI on their work.

Tim Carambat

No new videos in this period.

Unsupervised Learning (1 new video)

• A Conversation With Brian Vecci
  13.5.2026, 15:00:14

  The video is an interview with Brian, a data security expert at Veronis, focusing on the challenges and solutions related to AI security.

  Brian explains that conversations with customers and security teams have shifted significantly over recent years from general uncertainty about AI to concrete questions about the safe deployment of AI tools like ChatGPT, Claude, and Microsoft Copilot. Today, it’s about how companies can leverage AI to work more efficiently while ensuring appropriate “safe zones” are in place to minimize security risks.

  He emphasizes that the attack surface is substantially expanded by AI, as attackers can use AI to identify and exploit vulnerabilities. Veronis has developed a comprehensive solution built on three pillars: observability, automation, and behavior monitoring. This solution encompasses data collection and analysis, access rights monitoring, and automation of security measures.

  Brian highlights that many organizations still don’t fully understand the complexity and risk of AI security. He warns that attackers are increasingly using AI to rapidly and efficiently breach systems, emphasizing the need to secure data and underlying infrastructure.

  The video concludes with a discussion on implementing the Veronis solution, which is quick and effective, and an announcement that Veronis will present further AI security news at the Black Hat conference.

  Final commentary: The video explicitly addresses AI models such as Claude, ChatGPT, and Microsoft Copilot as well as specific tools like Veronis AI Atlas. It’s geared more toward intermediate and advanced viewers, as it provides in-depth insights into AI security and complex solution approaches.

WorldofAI (7 new videos)

• Gemini 3.5 Flash + Pro: Powerful, Cheap, & Fast NEW AI Model! (Fully Tested)
  16.5.2026, 06:11:06

  The video discusses the latest developments and leaks regarding Google’s Gemini models, particularly the checkpoints for Gemini 3.5, which are expected to be unveiled soon at Google I/O. Initially, the Gemini 3.2 variants (Fanta, Sprite, Cola) and the Pro version were presented, which were good but not revolutionary. However, the latest checkpoints, internally codenamed “Cappuccino,” show impressive progress. Notably, the model’s ability to generate complex applications such as a Minecraft clone or a Mac OS clone, with detailed and consistent outputs, stands out. The Gemini 3.5 Flash variant shows strong improvements in code generation, aesthetic design, and spatial consistency, though it sometimes remains repetitive and cluttered in UI generation. There are also hints of a secret testing phase for the Gemini 3.1 Pro variant, which also delivers promising results.

  The video explicitly addresses Google’s Gemini models and is geared more toward intermediate to advanced users interested in the latest developments in AI technology.

• Codex + Ollama = Free Unlimited Coding AI
  15.5.2026, 06:48:19

  The video explains the integration of Ollama into OpenAI’s Codex, which makes it possible to use open-source models locally on your own computer. Codex is an AI-coding agent that helps with developing, editing, reviewing, and deploying software. Ollama allows local execution of open-source models such as Deepseek, Gemma 4, or Qwen 3.6. The combination of both tools makes it possible to use these models for free in Codex. The video shows step by step how to install Ollama, locally host a model like Gemma 4, and connect it with Codex. It demonstrates how to complete tasks with the locally hosted model in Codex, such as creating a landing page. It also explains how to switch back to the original Codex experience.

  The video explicitly addresses OpenAI’s Codex and Ollama and is more suitable for intermediate users.

• Anthropic’s Downfall? GPT-5.6, Gemini 3.2, Robots Running A Full 8-hr Shift, & Qwen 3.6 Plus FREE!
  14.5.2026, 06:39:49

  The video covers current developments in AI technology, particularly from Google, OpenAI, and Anthropic.

  Google is intensively testing new variants of Gemini 3.2, including Flash and Pro versions, which have delivered rather disappointing results so far. Front-end generation proves weaker compared to earlier versions, and expectations for the new models have not been met. Additionally, an unreleased multimodal video model from Google was leaked, showing promising but still early-stage results.

  OpenAI is intensively working on GPT 5.6, with internal testing and potential releases in June. There are hints of improved security checks and possible new variants such as Ember Alpha and Beacon Alpha.

  Anthropic has increased weekly limits for Claude Code; however, this has faced criticism as third-party agents were simultaneously moved to a separate, paid API credit system. This results in higher costs for developers and is criticized as lacking transparency and failing to address infrastructure issues.

  Additional updates include the introduction of a fast mode for Claude Code, an upcoming major update to Google’s Jewels, and continuous improvements to Hermes Agent, a self-improving open-source AI agent. Additionally, Figure AI presented an impressive robot demo featuring humanoid robots working autonomously in a warehouse.

  The video explicitly addresses Google’s Gemini, OpenAI’s GPT, Anthropic’s Claude Code, and open-source tools such as Hermes Agent. It is aimed at intermediate and advanced users interested in the latest developments and technical details in the AI industry.

• Hermes Agent v2.0! Huge New Updates: WebUI, Qwen 3.6 Plus FREE, Computer Use, & More!
  13.5.2026, 05:08:59

  The video presents the latest updates for Hermes Agent, an open-source AI agent that continuously evolves and can be operated on your own infrastructure. The main new features include:

  1. **Computer Use Feature**: Hermes can now control the computer in the background without taking control of your PC. This currently works only on macOS, but support for Windows and Linux is planned. The feature is compatible with various AI models such as Claude, GPT-4, Gemini, and local open-source models.

  2. **Integration of Alibaba’s Qwen 3.6 Plus**: This powerful model is now integrated into Hermes and offers long context windows and multimodal capabilities. It is available for free for a limited time.

  3. **Light Panda Integration**: An integrated, open-source browser backend designed specifically for AI workflows. It offers improved browser automation and reliability.

  4. **Kanban Board**: A new Kanban board allows for the creation of unlimited boards and projects, orchestration of multiple agents, and task management through a dashboard.

  5. **/goal Command**: A long-term autonomous goal mode function that continuously plans, executes, verifies, and retries on errors until the goal is achieved.

  Additionally, minor updates such as voice cloning and multilingual support have been introduced. The video recommends Tiny Fish for web automation tasks and points to the open-source nature of Hermes, which makes it a versatile tool for AI workflows.

  The video explicitly addresses Hermes Agent, open-source models, and Tiny Fish and is geared more toward intermediate to advanced users.

• Claude Code Agent View IS INSANE! Huge New Update Introduces /goal, sessions, & More!
  12.5.2026, 07:03:39

  The video presents the latest updates for Anthropic’s Claude Code, a tool for agent-based development. The main new features include:

  1. **Agent View**: A central interface for managing all Claude Code sessions. This enables parallel execution and management of multiple coding sessions in a single dashboard. Users can switch between sessions without losing context and run background tasks.

  2. **/goal Feature**: A function that enables autonomous execution with a defined goal. The agent works independently until the goal is achieved and is usable in various workflows as well as on mobile devices.

  3. **/radio and Claude FM**: An integrated Lo-Fi radio station intended to enhance the user experience.

  4. **Improved System Prompt Compaction**: Prompt trimming has been optimized to better preserve sensitive instructions and user intentions without displaying explicit warnings when context is shortened.

  The video discusses the benefits of these updates, particularly for complex, agent-based coding workflows, and mentions the possibility of using Depot CI for a faster and more consistent CI/CD pipeline. However, it critically notes the limited token usage of Claude Code, which could restrict practical applicability.

  Final Comment: The video explicitly addresses Claude (Anthropic) and Depot CI and is aimed at intermediate to advanced users.

• Hermes Agentic OS is The Future
  11.5.2026, 07:07:11

  The video introduces Hermes Agent, an open-source project by Mistral AI designed as a persistent autonomous system that continuously learns and evolves over time. It can run 24/7 on your own infrastructure, build long-term memories, develop reusable skills, and gain deeper understanding of the user. The author demonstrates how Hermes Agent, in combination with ION UI, can be evolved into an agentic operating system (AIOS) capable of managing multiple autonomous AI agents that execute complex multi-step tasks without constant human supervision. ION UI is a free open-source co-work platform that enables AI agents like Hermes to work directly on the user’s computer while reading and managing files, writing and executing code, browsing the internet, automating workflows, and working alongside the user as a colleague. The video shows how to install and configure Hermes Agent and ION UI, as well as various use cases such as creating a financial dashboard in Excel, organizing files on the desktop, and creating a report on rural EV charging infrastructure. The author emphasizes the system’s ability to execute multiple tasks simultaneously and continuously improve, making it a powerful tool for automation and efficiency enhancement.

  The video explicitly addresses Hermes Agent and ION UI and is geared more toward intermediate and advanced users.

• Hermes Agent NEW Desktop App – The 24/7 Self-Evolving AI Agent!
  10.5.2026, 06:27:40

  The video introduces Hermes Agent, an open-source AI project that distinguishes itself from other tools such as OpenClaw and Claude Code through its ability for continuous self-improvement and long-term knowledge retention. Hermes can run around the clock on personal devices and evolves through the use of closed-loop learning and persistent storage systems. A key advantage is the ability to autonomously complete complex tasks, such as creating videos with Hyperframe. However, setup has been heavily CLI-based so far, which limited user-friendliness. The video presents the new Hermes Desktop App, which offers a user-friendly interface and facilitates management of multiple agents and workflows. The app supports various operating systems and enables integration of API providers such as OpenAI as well as the use of local models. It also offers features for managing Skills, Personas, Memory, and Tools, as well as the ability to create Cron jobs and connect with other platforms. The app also enables migration of OpenClaw configurations. Hermes can be used for diverse applications, from content creation to financial analysis. The video demonstrates the setup and use of the desktop app and emphasizes its advantages for user-friendliness and flexibility.

  The video explicitly addresses Hermes Agent, open-source models, and specific tools such as OpenAI, Hyperframe, and OpenClaw and is geared more toward intermediate and advanced users.

---

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