Blog

Opus 4.8, GPT-5.5, Gemini 3.5 Flash: Which Model Really Works for Everyday Use?

Tuesday, June 16, 2026

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

Model Releases & Benchmarks

Google’s Gemma 4 12B caused a stir in the local AI community: on a 16-GB M4 Mac Mini, according to Bart Slodyczka and WorldofAI, the model delivers performance surprisingly close to the 26B-MoE model — with multimodality (text, image, audio), a 250K-token context window, and multi-token prediction for faster generation. Data Curve’s DeepSWE Benchmark, which Theo (t3.gg) discusses, fundamentally questions SWEBench Pro: hand-verified validation and reformulated tasks show clear superiority of GPT-5.5 over Opus and Gemini 3.5 Flash. Meanwhile, leaks circulate about GPT-5.6 (codename “Jewel Alpha”) with strong SVG and frontend results, and Microsoft’s MAI Thinking One model in the reasoning category — both still not publicly verifiable.

Local & Open-Source AI

MiniMax M3 is the open-source surprise of the week: the model combines top-tier performance in coding, multimodal reasoning, and long-term workflows, and according to WorldofAI outperforms GPT-5.5 and Gemini 3.1 Pro in multiple categories — at a fraction of the cost of proprietary models. It’s available via the MiniMax API, the M-Code platform, and Open Router. Nvidia presented Nemotron 3 Ultra on GTC Taipei with 550 billion parameters on a hybrid Mamba-Transformer architecture: five times faster inference, 30 percent lower costs than comparable open models, including disclosed training data. Also from Nvidia: RTX Spark, a super-chip (Blackwell GPU plus Grace CPU) for local AI agents on consumer laptops. Additionally, Nvidia announced the N1 and N1X laptop platform with the GB10 superchip, which according to Tim Carambat should run models with 120 billion parameters and a million-token context window locally.

Claude Code & Anthropic Tooling

Dynamic Workflows in Claude Code is the most discussed new feature of the week. Mark Kashef systematically explains all six core patterns — from “classify and act” through “adversarial verification” to “loop until done” — and shows applications like email triage, due diligence checks, and codebase optimization. After over 500 hours of use, Nate Herk ranks the twelve most important Claude Code features: at the top are Skills, Sub Agents, /slashgoal, and Auto Memory, while he ranks Dynamic Workflows only in B-tier — useful, but not yet essential daily. The remote-control feature, which both NeuralNine and Niklas Steenfatt demonstrate, allows continuing active coding sessions from your smartphone. Melvynx introduces an open-source CLI tool that with `npx ai-blueprint unify` migrates Claude Code configurations to Cursor and keeps them available as symlinks on both platforms.

Coding Agents (non-Claude)

Cursor released Composer 2.5, a model that Tech With Tim tests in direct benchmarks against Opus 4.7 and GPT-5.5: it’s significantly faster (0.50 dollars per task versus 7 dollars for Opus 4.7) and delivered a working web application first in its live demo. Melvynx is more skeptical — Composer 2.5 works for small fixes but can’t match top-tier models in quality and reliability. Leon van Zyl presents OpenCode as a more efficient alternative to Claude Code for local models: OpenCode consumes significantly fewer tokens for system prompts and works seamlessly with Ollama models like Gemma 4 or Llama 3.6. The open-source tool CC Usage, which NeuralNine introduces, analyzes token consumption of various coding agent CLIs (Claude, Codex, OpenCode) and displays estimated API costs — even with active subscription.

Software Engineering & Dev Culture

Theo (t3.gg) delivers three essayistic contributions: he assesses Vite’s acquisition (via VoidZero) by Cloudflare as strategically significant — Cloudflare is thereby building an agent-optimized cloud platform that directly challenges Vercel. In a second video, he warns of “prompt technical debt”: growing prompt collections in CLAUDE.md or agents.md become maintenance problems just like bloated code, and he recommends minimal, specific prompts. Third, he argues that AI tools reduce the damage from weak engineers because they force better technical decisions — a thesis that guarantees pushback. Brian Casel addresses the “design drift” problem with AI-generated UIs: without a central design system, the agent reinvents colors, typography, and components at each session; his solution is a design system anchored in CLAUDE.md or agents.md, which he provides as Rails and React starter templates.

Personal AI OS & Agent Frameworks

Hermes was the most discussed agent platform of the week: the new desktop app for Windows, macOS, and Linux launched with native interface for sessions, skills, cron jobs, artifact management, and multi-profile support. Alex Finn, AI with Arnie, WorldofAI, and TheAIGRID discussed the app from different angles — consensus: installation is vastly improved compared to the earlier WSL workflow, and features like the Kanban orchestrator, background sessions, and the Curator daemon (which automatically archives outdated skills) set Hermes distinctly apart from simple chatbot wrappers. Julian Ivanov lists eight new features, including Session Search as token-free full-text search in past conversations and the Kanban Swarm feature, where multiple agents work in parallel and a verifier agent checks results. Alejandro AO shows two entry paths: the 15-dollar VPS setup via Hetzner plus MiniMax token subscription and an architecture explanation of the Pi agent as a minimalist alternative. Nate B. Jones builds his own token burn dashboard, visualizing his Codex consumption (800 million tokens in one day) and using it as a feedback loop for workflow optimization.

AI Automation & Workflows

Dave Ebbelaar shows in a four-hour tutorial how to build a complete GenAI RAG application: FastAPI backend, React frontend, Supabase database with PGVector for semantic search, Dockling for document conversion, and Railway for deployment. The n8n team presents OpenTelemetry support in version 2.22: workflow and node executions can now integrate into existing distributed tracing setups, and the feature is also available in the free community plan. Leon van Zyl demonstrates a different automation approach: he combines Claude Code with the Higgsfield AI CLI to embed animated avatars directly in landing pages — from image prompt to embedded video entirely within the agent workflow.

AI Business, Marketing & Freelancing

Kyle Balmer argues in two videos that the decisive differentiator in the AI age isn’t writing or design, but authentic public speaking: polished, AI-sounding content generates less trust; the path to paid enterprise workshops runs through raw TikTok videos, then livestreams, then structured webinars. Nate B. Jones discusses a Microsoft study showing that 86 percent of users only use AI output as a starting point — and warns that traditional work artifacts like memos or finished projects no longer demonstrate human judgment; as a solution, he proposes documented whiteboard sessions that make thinking processes visible in four dimensions (situation, decision, risk, change).

Prompting & AI Literacy

Kyle Balmer explains the mechanics of tokenization and context windows precisely in a fundamentals video: a token corresponds to about three-quarters of an English word, the “Raspberry problem” shows why LLMs fail at character counting, and larger context windows aren’t automatically better because they incur costs and precision losses. Nate Herk introduces the “Grill Me” technique, originally developed by Matt PCO: a structured prompt sequence systematically extracts user knowledge and converts it into reusable context — supplemented by a checkpoint system that prevents answers from getting lost in long dialogues. Ben AI shows three ways to build a personal AI dashboard: as a live artifact in Claude, as a customized Obsidian interface, or as a hosted HTML page.

AI Industry & Strategy

Satya Nadella speaks on the No-Priors podcast about Microsoft’s role as an AI platform: companies should be able to develop their own AI, Azure scales massively, and concrete benefits for communities need to be more visible. David Shapiro analyzes two warning signs: Microsoft’s internal cloud cancellations and Uber’s inflated AI budget — both indicate high costs, but not an end to AI investment. He simultaneously discusses Bernie Sanders’ proposal to create a state fund with 50-percent stakes in OpenAI, Anthropic, and XAI — Shapiro’s counter-model: passive acquisition of up to 10 percent across a broad portfolio without active company control. Nate Herk points to an internal Anthropic report showing that over 80 percent of delivered code is already written by Claude and the model achieves a 76 percent success rate on open tasks (six months ago: 26 percent) — which he interprets as a factual AGI threshold.

AI & Society / Future of Work

Historian Jürgen Kocka contextualizes AI disruption historically in the Everlast AI interview: technological upheavals have always created social disruptions and spawned new professions since industrialization — the Luddite movement as both a cautionary and inspiring example. Science theorist Klaus Mainzer criticizes in the same format that today’s machine learning merely performs statistical data optimization and doesn’t recognize causal relationships; as a way forward, he sketches “Physical AI” with integrated natural laws and neuromorphic chips (memristors, photonic systems) as more energy-efficient architecture. David Shapiro questions the viral rumor of “last 300 days of work”: AI demonstrably increases productivity, but the social adoption curve — from corporate decision-making to individual acceptance — takes far longer than short-term forecasts suggest.

Quick Notes

Meta’s planned AI companion (based on Limitless startup, testing starts earliest 2027) is assessed by TheAIGRID as a difficult-to-position wearable — data privacy and market acceptance remain unsolved. Microsoft’s quantum chip Majorana 2 should, according to TheAIGRID, keep qubits stable for up to a minute (compared to microseconds previously) and was partly developed by AI agents; commercial use earliest 2029. Odyssey’s Agora-1 world model, which models the entire world state rather than just the next token, is discussed in the Everlast AI interview with Dr. Jenny Sein Schwarz as a promising robotics foundation. Harmonic presents in the Unsupervised Learning podcast an MCP gateway solution that monitors and controls AI tool usage in enterprises without blocking innovation. ChatGPT’s revamped memory system (automatic context storage from chats, files, and connected apps) is explained by TheAIGRID in an updated guide. Alibaba’s Qwen 3.7 Plus as a multimodal coding model and OpenAI’s Codex update, which goes beyond pure code generation, appear in multiple news roundups without in-depth analysis.

AI Explained

No new videos in this period.

AI Foundations (1 new video)

• FULL Hermes Agent Tutorial For Beginners in 2026! (Become a PRO)
  3.6.2026, 14:00:39

  This video is a comprehensive tutorial on setting up and using Hermes Agent, an autonomous 24/7 AI agent. It starts with an introduction to Hermes Agent, including how it works and its three main layers: memory, skills, and cron jobs. The host explains how Hermes Agent learns and improves over time by completing tasks and receiving feedback.

  Next, the process of setting up Hermes Agent on a VPS (Virtual Private Server) is described in detail, using Hostinger as an example VPS provider. The host walks through the setup steps, including selecting a VPS plan, installing Docker, and configuring Hermes Agent.

  After setup, the video demonstrates how to connect Hermes Agent with Telegram to enable communication through this platform. The host shows how to create and configure a Telegram bot to interact with Hermes Agent.

  Another important step is connecting tools to Hermes Agent. The host demonstrates how to integrate Tavli, a web search and extraction tool, with Hermes Agent by setting API keys in the configuration. This enables Hermes Agent to access the internet and research information.

  Finally, the video demonstrates creating a scheduled automation process with Hermes Agent. The host shows how to create an automatic task that regularly researches and summarizes AI news and how to receive these updates in Telegram.

  This video is aimed at intermediate users who already have basic knowledge of AI tools and want to deepen their skills. It covers specific tools and providers, including Hostinger for VPS, Telegram for communication, and Tavli for web search.

AI with Arnie (1 new video)

• Hermes Desktop is dangerously good
  6.6.2026, 16:10:34

  The video introduces the Hermes Desktop app, which enables users to work with various AI models locally on their computer. Installation is now significantly simplified and no longer requires complex steps like the previous Windows Subsystem for Linux setup. The app provides a user-friendly interface that allows you to use different AI models like GPT-5.5 or Codex and switch between them. A special feature is Hermes’ ability to improve itself and automatically write Skills that can be used for various tasks such as frontend design, image generation, and integration with tools like Obsidian and ConfUI. The app also supports connection with various channels like Telegram, Discord, and Slack, as well as the creation and management of Subagents that can work in the same session. Additionally, users can use local models like Ollama and adjust the context window accordingly. The app is 100% open source and gives users the flexibility to use different providers and models without being locked into a specific vendor.

  The video explicitly focuses on the Hermes Desktop app, which is based on open-source models, and is more suitable for intermediate to advanced users.

AI News & Strategy Daily | Nate B Jones (3 new videos)

• My Codex Ran 800 Million Tokens in A Day. The Real Story Isn’t Cost.
  5.6.2026, 14:00:07

  The video author introduces a custom-built token-burn dashboard that helps him visualize and analyze his AI tool usage. The focus isn’t on boasting about burned tokens, but rather on how to optimize and expand your computing usage with AI. The dashboard demonstrates how the author uses AI tools like Codex and Claude to boost productivity and improve his workflows. He emphasizes the importance of feedback loops to understand and improve AI usage, and how analyzing token consumption can provide insights into your own working methods. The author also shows how he leverages AI tools like Slashworkflows and Subagents to automate and optimize complex tasks. He encourages viewers to build similar dashboards and share their experiences to collectively explore and expand AI possibilities.

  The author specifically covers the tools Codex, Claude, and Opus 4.8, as well as the open-source Skill Tufty. The video is geared toward Intermediate and Advanced users, as it assumes viewers already have experience with AI tools and are ready to analyze and optimize their usage.

• Microsoft Says 86% Treat AI Output as a Starting Point. Your Resume Just Stopped Working.
  31.5.2026, 17:00:39

  The video discusses challenges and solutions for handling AI-generated content in professional settings. Microsoft studies show that 86% of users treat AI output as a starting point rather than an end result, fundamentally changing how quality and productivity are evaluated. The core issue is that AI makes it possible to appear productive without actually making well-founded decisions. The speaker argues that traditional proof points like memos or projects are no longer sufficient to demonstrate human judgment. As a solution, he proposes using whiteboard sessions to make thinking processes visible. Four key areas should be covered: Situation, Decision, Risk, and Change. This method helps prove the quality of your work and is especially valuable in collaboration with others. The speaker emphasizes the importance of “Talent Boards” to document these thinking processes for job applications or promotions. He recommends forming a clear perspective early in a new role and refining it through whiteboard sessions with experienced colleagues. The video suits Intermediate and Advanced users and doesn’t cover specific AI tools or models.

• Opus 4.8 Scored 81. Your Workflow Doesn’t Care.
  3.6.2026, 14:00:38

  # Summary: Opus 4.8 – The Real Story Behind the Release

  The video explains why the common perception of Opus 4.8 is wrong: while the model performs well on certain metrics, it’s not the expected game-changing “Mythos” release, but rather a checkpoint release primarily accompanying a funding announcement.

  **The central issues with 4.8:**

  1. **Unpredictable scaling behavior**: Despite promises, higher reasoning levels don’t consistently deliver better results. On the Vending-Bench test, 4.8 on “high” outperforms “max” – even worse than 4.7. This is counterintuitive and problematic.

  2. **Overthinking problem**: The model over-thinks alignment and constitutional questions – reasoning traces show 4.8 max spending effort contemplating constitutional compliance rather than efficiently solving tasks. This reduces practical usability.

  3. **Harness deficits against Claude 5.5 + Codeex**: The speaker provides a concrete example – when building two websites, 5.5 fully deploys both, iterates, and finishes, while 4.8 fails twice. File access also works system-wide with 5.5 but is limited with 4.8. The “harness” (the product scaffolding around the model) is decisive.

  **What 4.8 does well:**

  The model excels at design, writing, and composition – a classic strength of the Opus line. The new `/workflows` command in Claude Code is innovative: the model composes dynamic multi-agent workflows, makes them transparent, and executes Sub-Agents accordingly. This is an interesting direction for agents in 2026.

  **The bigger message:**

  We’re at a new stage of competition in 2026. It’s no longer just about model drops and new high scores, but about **harness fit and practical workflows**. The video warns decision-makers against vendor lock-in – instead, architecture should remain flexible for API swaps based on outcomes.

  The episode also addresses how companies must learn to think about agents not just for individual productivity, but for complete agentic pipelines – to avoid overload from excessive generated downstream work.

  **Core message for different users:** Coding engineers should evaluate their harness (70% use Claudecode, 30% Codeex); knowledge workers must consider whether they’re creating downstream process overload; CTOs should architect systems that can integrate 10-trillion-parameter open-source models by year-end, rather than relying on single vendors.

  —

  **Mentioned AI tools/models:** Anthropic Opus 4.8, Claude Code, Claude Workflows, OpenAI 5.5 with Codeex, ChatGPT, Open-source models, expected “Mythos” model — **Opinion/reflection with deep analysis**, for advanced users.

Alejandro AO (2 new videos)

• PI Architecture EXPLAINED | Agent Loop, Tools, TUI and More
  5.6.2026, 14:25:55

  The video provides a detailed overview of Pi’s architecture, a minimalist AI agent. It starts by explaining the Agent Core, also known as the Agent Loop, which describes Pi’s fundamental operating mechanism. This process includes context initialization, context transformation, and invoking the Large Language Model (LLM), with the ability to utilize tools such as file operations or internet research.

  Subsequently, the structure of Pi’s sessions and storage system is explained. Sessions are stored in JSON-L files, which enable simple and efficient management of message history. Sessions are organized in a tree structure, making it easier to navigate and manage conversations.

  The video also covers Pi’s tools, which are limited to four tools by default but can be extended through extensions. Extensions make it possible to add additional features such as web search or specific commands.

  Another important aspect is the system prompt, which sets the foundation for the agent’s behavior. This can be customized through user-defined files or commands.

  Finally, the interaction between PyCore and Pi’s interactive user interface (TUI) is explained, including the handling of skills and custom prompts. The TUI is modularly structured and allows for dynamic adaptation of the user interface.

  The video explicitly addresses the AI agent Pi and is better suited for intermediate and advanced users interested in the technical implementation and architecture of AI agents.

• FULL Hermes Setup on $15/month in 10 minutes
  3.6.2026, 17:22:50

  The video demonstrates how to set up Hermes Agent on your own VPS for less than $15 per month. Three main steps are outlined: First, setting up a VPS at Hetzner for around $5 per month, with alternatives like OVH Cloud also mentioned. Second, setting up a Minimax subscription for $10 per month, which provides tokens for running Hermes Agent. Third, using an AI agent to configure the VPS without manual intervention, including installing dependencies and securing the system. The Minimax token plan is recommended as a cost-effective and efficient solution for running Hermes Agent, particularly for operating 1-2 concurrent agents.

  The video explicitly covers Hetzner, OVH Cloud, Minimax, Hugging Face, and Pi (Advanced).

Alex Finn (1 new video)

• Hermes Agent just WON (Hermes desktop app)
  3.6.2026, 21:47:44

  The video presents a detailed overview and review of the new Hermes desktop app, positioned as revolutionary for interacting with AI agents. The creator highlights several key features:

  1. **Sessions and Organization**: The app enables creating and organizing sessions for different life areas, providing better context management and structure compared to Telegram or Discord.
  2. **Artifacts**: A feature described as a “second brain” that stores all links, files, and images shared with the AI agent, making them easily accessible.
  3. **Messaging**: Simplified setup of messaging services without requiring CLI.
  4. **Skills and Toolsets**: The app displays all installed skills and allows disabling unnecessary ones to save tokens. Self-improving skills and toolsets for specific tasks like browser automation are also highlighted.
  5. **Cron Jobs**: Improved cron job management with a user-friendly interface that allows validating, testing, and pausing jobs.
  6. **Profiles**: Managing multiple Hermes agents with different roles, skills, and personalities.
  7. **Settings**: Customizable models, appearance options, and API key management.

  The creator demonstrates the app’s usage by creating a script and thumbnail for a YouTube video, emphasizing the app’s user-friendliness and efficiency.

  Final Note: The video explicitly focuses on the Hermes desktop app and is geared toward intermediate to advanced users.

Bart Slodyczka (2 new videos)

• Gemma 4 12B on a 16GB Mac Mini Is Surprisingly Capable
  4.6.2026, 12:00:09

  The video introduces Google’s new open-source model Gemma 4 12B, which can run on devices with 16 GB RAM with 12 billion parameters. The creator tests the model on a 16 GB M4 Mac Mini and compares it with larger models from the Gemma family. Despite its smaller size, the 12B model shows similar performance to the 26B model, as demonstrated through benchmarks in areas such as coding, agentic tool calling, and reasoning. The creator discusses technical aspects like RAM usage and context length, and shows how to load and use the model in LM Studio. The model’s multimodality and its ability to process images and audio are also mentioned. Additionally, the new multi-token prediction feature is explained, which accelerates response generation. The creator tests the model’s OCR capabilities with an invoice and demonstrates how to create a simple application to parse it. Despite some inaccuracies in OCR, the creator is excited about the model’s potential on a small device.

  **AI Tools/Models/Providers:** Google (Gemma 4 12B), LM Studio. **Target Audience:** Intermediate.

• Claude Code + LM Studio: FREE Unlimited AI Agents (Don’t Pay $200/month)
  31.5.2026, 12:00:10

  The video demonstrates how to use Claude’s new “Dynamic Workflows” feature to deploy up to a thousand agents for complex tasks without needing Anthropic’s paid API or an Anthropic account. Instead, Claude’s gateway version is used, which allows you to run local AI models directly on your computer. This involves installing Claude’s desktop app and switching it to developer mode to enable third-party inference. LM Studio is used as the local model, offering free local AI models for download and use. The process includes downloading a model, setting up the gateway URL, and configuring API keys in the Claude app. Additionally, the video shows how to set up web search capabilities by adding an MCP server (e.g., Brave Search). Finally, the video demonstrates how to use the “Dynamic Workflows” feature in Claude Code to deploy up to 16 simultaneous agents and a total of 1000 agents per task to work on complex projects.

  The video explicitly covers Claude, LM Studio, and Open Router and is intended for Intermediate and Advanced users.

Ben AI (1 new video)

• Stop Using Claude Without an Agentic OS
  3.6.2026, 08:11:18

  The video explains the benefits and possibilities of a personalized dashboard or “Command Center” for managing information and automating workflows with the help of AI. It highlights four main advantages: a personalized user interface with access to live data from various software programs, the ability to execute actions directly with AI, the integration of different AI models, and easy sharing of the dashboard with team members or clients. The creator demonstrates three options for setting up such a dashboard: a live artifact within Claude, a custom interface in Obsidian, and a custom HTML page that can be hosted on a website. Each option has its own pros and cons regarding simplicity, flexibility, and costs. The creator recommends starting with a simple version and gradually improving it based on actual needs and usage.

  The video explicitly focuses on Claude and is geared toward intermediate and advanced users.

Brian Casel (1 new video)

• Why apps built with AI look a little… OFF
  5.6.2026, 12:00:17

  The video addresses the problem of “design drift” when creating apps with AI agents. This often results in inconsistent designs because agents reinvent the UI during each session without a shared source of truth for the design. The creator presents a solution: a design system that centrally defines colors, typography, spacing, and components and integrates them into the code. This prevents design drift and ensures consistency.

  The creator demonstrates how he implements such a design system in his apps, comprised of various components like colors, typography, buttons, forms, and lists. He uses a free agent capability he developed to integrate the design system into his codebases. By defining the design system in the `claude.md` or `agents.md` file, the AI agent is ensured to consider design guidelines for every UI request.

  The creator emphasizes the importance of understanding such tools rather than using them blindly. He recommends integrating the design system into new projects from the start, as retrofitting it later is more challenging. Additionally, he presents Rails and React starter templates that already have his design system integrated.

  The video explicitly covers Claude and Cloud Code and is aimed at intermediate and advanced builders working with AI agents.

Coding with Lewis

No new videos in this period.

Cole Medin (2 new videos)

• Claude Plans, Gemini Designs: The Workflow to Build BEAUTIFUL Frontends
  4.6.2026, 00:00:04

  This video demonstrates a workflow that combines two different AI models, Gemini 3.5 Flash and Opus 4.8, to create full-stack web applications. Gemini 3.5 Flash is used for user interface design, while Opus 4.8 handles planning, page content creation, and integrations. The workflow consists of multiple steps, each running as separate AI sessions that communicate through Markdown documents. The process starts with repository and specification exploration, followed by planning, UI design, integrations, validation, and finally deployment with a smoke test. The creator emphasizes the advantages of this division, as it leverages the strengths of both models and optimizes costs. The workflow is designed to work for more complex applications, though it doesn’t work for every type of web application in a single step.

  The creator explicitly covers Gemini 3.5 Flash and Opus 4.8, showing how to use them in combination with tools like Pi, Anti-Gravity, and Claude Code. The video is best suited for intermediate to advanced users who already have experience with AI models and coding agents.

• Claude Plans, Gemini Designs: One Workflow for Beautiful Frontends (LIVE)
  31.5.2026, 04:37:29

  # Claude Plans, Gemini Designs: One Workflow for Beautiful Frontends (LIVE)

  This stream showcases the live development of an Archon workflow that combines multiple AI models to build full-stack web applications. The central thesis: Gemini 3.5 Flash creates aesthetically appealing UIs but hallucinates with content and integrations. Claude Opus, on the other hand, plans precisely but is less creative with design. The solution is a specialized workflow pipeline.

  **The workflow in eight steps:**
  1. **Exploration** (Sonnet): Analyze repository and spec
  2. **Planning** (Opus): Create detailed plan with three sections (UI scope, integration scope, deployment)
  3. **Build UI** (Gemini 3.5 Flash via Open Router): Design only the user interface
  4. **Integration** (Opus): Implement backend, APIs, and authentication (in this case: Clerk)
  5. **Validation** (Sonnet): Tests and quality checks
  6. **Fix Validation** (Opus): Fix issues
  7. **Deploy** (Clerk CLI): Deploy to Vercel and promote Clerk instance from dev to prod
  8. **Finalize**: Wrap up

  **Practical example:** A benchmarking dashboard for Archon Workflows with Clerk authentication. The stream shows how Gemini implements the UI well but forgets the Markdown documentation – exactly the problem this workflow solves. Opus then handles the complex parts (authentication, API integration). The local application runs at the end with functional Clerk auth; the Vercel deployment is pending (DNS propagation).

  **Key insights:** Provider mixing saves tokens and costs (Gemini cheaper than Opus for UI, Sonnet for simple tasks). Specialized workflows with focused prompts deliver better results than monolithic prompts. Archon enables vendor lock-in freedom, while Claude Code remains restricted to Anthropic.

  **Side topics:** Discussion of Anthropic’s June change (Agent SDK costs extra), comparison with OpenAI/Codeex, YouTube content trends, the creator’s journey since July 2024.

  **Explicitly mentioned:** Claude (Opus, Sonnet) / Gemini 3.5 Flash / Open Router / Archon / Clerk / Vercel / Codeex – for intermediate to advanced users, as workflow orchestration and multi-provider strategies are assumed.

Dave Ebbelaar (1 new video)

• Build a Full-Stack GenAI Project in 4 Hours (FastAPI, React, Supabase)
  6.6.2026, 12:44:23

  ### YouTube Video Summary

  The video demonstrates a comprehensive process for building a full-stack GenAI application that enables users to ask questions based on hundreds of documents stored in a database. The focus is on the entire architecture, including frontend, backend, database connection, ingestion and ranking pipeline, and how everything works together.

  **Key Steps and Final Result:**
  1. **Project Setup:**
  – Creating a GitHub repository with agent files (e.g., `agents.md`) and dependency management (e.g., `pyproject.toml`).
  – Setting up a virtual Python environment manager (UV) for managing dependencies.

  2. **Data Download and Processing:**
  – Using a Python script (`download.py`) to download financial reports (10-K, 10-Q, annual reports) from companies like Apple, Microsoft, Nvidia, Amazon, and Google from the SEC website.
  – Converting downloaded HTML files to Markdown using the Dockling tool to prepare data for further processing.

  3. **Database Modeling and Migration:**
  – Defining database models (e.g., `users`, `source_documents`, `document_chunks`, `chats`, `messages`, `citations`) with SQLAlchemy.
  – Performing database migrations with Alembic to synchronize models with the Supabase database.

  4. **Backend Development:**
  – Setting up a FastAPI backend with endpoints for user authentication, chat processing, and document queries.
  – Integrating OpenAI for creating embeddings and processing language models.

  5. **Frontend Development:**
  – Setting up a React frontend with TypeScript and Tailwind CSS for the user interface.
  – Implementing authentication mechanisms (e.g., user login and logout) with Supabase.

  6. **Ranking Pipeline:**
  – Implementing a hybrid ranking pipeline that combines both semantic search (with PGVector) and full-text search (with Supabase).
  – Using Pydantic models to structure queries and responses.

  7. **Deployment:**
  – Deploying the application on Railway, a platform-as-a-service provider, using Docker containers for backend and frontend.
  – Configuring environment variables and database connections for production.

  **AI Tools/Models/Providers:**
  – **Claude / OpenAI / Gemini / Open-Source:** The video uses various AI tools such as OpenAI for embeddings and language models, Pydantic for data structuring, and Dockling for document processing.
  – **Specific Tools:** Cursor, Lovable, Cline, n8n.

  **Target Audience:**
  – **Intermediate to Advanced:** The video is aimed at developers with advanced knowledge of Python, React, databases, and deployment processes who are interested in AI engineering and full-stack development.

  **Final Comment:**
  The video provides deep insights into building a full-stack GenAI application and demonstrates how various components work together to create a powerful and user-friendly solution. It is a valuable resource for developers looking to expand their skills in AI engineering and full-stack development.

David Shapiro (3 new videos)

• Bernie’s plan sucks, actually
  5.6.2026, 11:56:20

  The video discusses Bernie Sanders’ proposal to create a state fund for artificial intelligence (AI) that would acquire 50% of the stakes in leading AI companies like OpenAI, Anthropic, and XAI. The speaker analyzes the pros and cons of this plan and contrasts it with an alternative, more realistic model.

  Sanders’ argument is based on the idea that AI companies profit from publicly funded research and collective knowledge without providing appropriate compensation. He proposes creating a state fund to control these companies and distribute profits to the population. The speaker’s main criticisms are:

  1. **Expropriation and Constitution**: The planned seizure of 50% of corporate shares is constitutionally problematic and could cause companies to relocate abroad.
  2. **Governance and Control**: The proposed active state control is risky and could lead to political interference.
  3. **Objectives and Implementation**: Focusing on a few large companies isn’t the right approach. Instead, a broader spectrum of tech companies and infrastructure should be considered.

  As an alternative, the speaker proposes a passive governance model where the state acquires up to 10% of shares in various companies without exercising active control. Revenue should be paid directly to citizens or invested in public infrastructure.

  The video explicitly addresses OpenAI, Anthropic, XAI, Microsoft, Google, Amazon, Nvidia, Coreweave, TSMC, and ASML. It is geared more toward intermediate and advanced viewers, as it contains detailed economic and political analysis.

• Microsoft and Uber slam on the brakes of AI
  3.6.2026, 11:49:02

  The video discusses the current debate about a possible “AI bubble” and analyzes two relevant news items: Microsoft’s internal cancellation of cloud services and Uber’s excessive AI budget. The speaker argues that while these incidents point to high costs, they don’t necessarily signal the end of AI development. He compares the situation to SpaceX’s valuation, which appears inflated due to its AI involvement but doesn’t mean the end of technology. Despite some data center delays, he sees no signs of AI development stopping as long as investors and companies continue to invest in AI. Additionally, the speaker reports on the demonetization of his YouTube channel due to YouTube’s strict policies and shares his plans for a book and new online courses.

  The video doesn’t address specific AI tools or models and is geared more toward intermediate or advanced viewers.

• The last 300 days of work? (No, but…)
  31.5.2026, 13:07:59

  The video discusses the rumor that we’re in the last 300 days of work, spread by tech journalist Kevin Roose on Twitter. The speaker questions the credibility of this claim and analyzes its possible origins and meanings. He argues that although AI technologies like Claude and Gemini have already made significant progress and are boosting productivity in many areas, the full integration and acceptance of these technologies in society will take much longer than 300 days. The speaker emphasizes that there are multiple adoption curves, both at the level of AI models themselves and among end users and companies implementing these technologies. He references his own experience in large corporations and the challenges associated with introducing new technologies into established systems. He also discusses the fear and resistance many workers have toward AI and the slow decision-making processes in large companies. The speaker concludes by noting that while the long-term impacts of AI could be revolutionary, short-term expectations are often overstated.

  The video explicitly addresses the AI models Claude and Gemini as well as AI integration in tools like Canva and Slack. It is geared more toward intermediate and advanced viewers, as it provides detailed insights into the technology and its implementation.

Everlast AI (4 new videos)

• Nr.1 Historian: This time it could be DIFFERENT! Luddites, Capitalism & AI (Jürgen Kocka)
  4.6.2026, 14:15:11

  The video is a conversation between Leonard Schweding and Professor Dr. Jürgen Kocka, a renowned historian and capitalism researcher. The main topics of the discussion include the definition and controversies surrounding capitalism, historical changes in the world of work, the role of technology and automation, and the future of capitalism in the age of artificial intelligence (AI).

  Kocka emphasizes that capitalism is a complex economic system encompassing property rights, markets, and investments. He discusses historical developments, particularly industrialization and its associated social and economic changes. Kocka mentions Schumpeter’s creative destruction, which brings both opportunities and risks, and emphasizes the importance of protest and criticism for the further development of capitalism.

  A central theme is the role of technology and automation in the world of work. Kocka refers to historical examples such as the Luddite movement and the Weavers’ Revolt to demonstrate that technological advances always bring social and economic upheaval. He argues that fears of job loss through automation are often exaggerated, as new jobs and professions emerge.

  Kocka also discusses the current debate around AI and its impact on the world of work. He emphasizes the importance of adaptability and the need to limit the spread of capitalist principles into areas where they don’t belong. He warns against technocratic utopias and emphasizes the importance of criticism and protest for the further development of capitalism.

  The video is more suitable for intermediate and advanced viewers, as it addresses complex historical and economic concepts. It doesn’t focus on specific AI tools or models but rather offers a broad historical and social perspective on current developments.

  Final comment: The video does not focus on specific AI tools or models and is more suitable for intermediate and advanced viewers.

• “The REAL AI Revolution is still to come!” Neuromorphic Chips & Quantum Computers (Prof. Mainzer)
  3.6.2026, 14:45:39

  The video is an extensive interview with Professor Dr. Klaus Mainzer, a renowned science theorist and pioneer of complexity research. Mainzer discusses the development of artificial intelligence (AI) and its future challenges and opportunities. He begins with a historical perspective by explaining Pierre-Simon Laplace’s thesis and the limits of computability through Alan Turing’s halting problem. Mainzer emphasizes that complex systems, from galaxies to brains, follow the same mathematical principles, and explains the significance of symmetry breaking and phase transitions in complexity research.

  A central point of the conversation is current AI, which is primarily based on machine learning, which he describes as statistical data optimization. Mainzer criticizes that this form of AI does not fully represent human intelligence, as it is unable to recognize causal relationships. He uses the example of Newton and Einstein, who formulated major laws of nature with minimal data. As a solution, he proposes “Physical AI,” in which physical laws are integrated into AI to improve data optimization.

  Another focus is on the energy challenges of today’s AI systems. Mainzer explains that neuromorphic systems, modeled after the human brain, could offer a more sustainable alternative. These systems use memristors, which integrate memory and processor functions, and could significantly reduce energy consumption. He also mentions photonic neuromorphic systems, which work with light and could be even more efficient.

  Mainzer also discusses the role of quantum computers, which, through their ability to utilize superposition states, can solve certain problems faster. However, he emphasizes that quantum computers do not overcome fundamental limits of computability but merely overcome practical hurdles. He mentions the threat posed by quantum computers to current encryption techniques and the necessity to develop new encryption methods.

  In conclusion, Mainzer discusses the political and social challenges that come with AI development. He refers to China’s long-term AI planning and the effectiveness of entrepreneurial structures in technological development. He emphasizes the necessity to preserve democratic traditions while simultaneously developing effective political structures to address the complex challenges of the future.

  The video explicitly addresses models and providers Claude, OpenAI, Gemini, Open-Source, as well as specific tools such as neuromorphic systems and quantum computers. It is aimed at intermediate and advanced audiences, as it addresses in-depth technical and philosophical aspects of AI and related technologies.

• “It’s highly dangerous what’s beginning here!” The new AI job lie, Claude Opus 4.8 & China’s robots
  2.6.2026, 15:15:07

  The video covers a wide range of topics related to artificial intelligence (AI) and its impacts on society, the economy, and individual lifestyles. Here are the key points summarized:

  1. **AI and Society**: The discussion begins with the question of how AI will change society and what social upheavals might result. It’s noted that AI could lead to a kind of “industrial revolution on steroids,” bringing both enormous productivity gains and social tensions.

  2. **Technological Developments**: Recent developments in the AI industry are discussed, including the introduction of new models such as Anthropic Opus 4.8, which is described as more honest and error-resistant. The price reductions for Chinese AI models and their strategic implications are also addressed.

  3. **Job Market and Job Losses**: The discussion about potential job losses due to AI is conducted intensively. It’s noted that many jobs could be replaced by AI, but new opportunities will also emerge. Participants discuss how to prepare for these changes by developing skills such as creativity, empathy, and systems thinking.

  4. **Education and Learning**: Another focus is on how education and learning should look in a world with AI. It’s emphasized that traditional examination methods such as papers and exams could become obsolete due to AI, but there remains a need to convey practical skills and knowledge.

  5. **Investments in AI**: It’s discussed how one can invest in AI as an investor. It’s recommended to invest in tokens, as these promise high returns and the costs for AI applications are continuously declining.

  6. **Ethics and Regulation**: The ethical implications of AI are also addressed, particularly the question of how AI should be regulated to avoid negative impacts. It’s noted that AI should not be blindly trusted and that it’s important to understand the limits and possibilities of AI.

  7. **Future of Work**: Participants discuss how the future of work might look in a world with AI. It’s emphasized that it’s important to focus on the impact of one’s work rather than just the doing. It’s recommended to become self-employed and have multiple clients to remain flexible.

  8. **AI and Health**: It’s noted that AI could make enormous advances in medicine, particularly in the diagnosis and treatment of diseases. It’s emphasized that AI can enhance human creativity and independently find solutions.

  9. **AI and Education**: It’s discussed how AI can be used in education to promote individual learning. It’s emphasized that AI can play an important role in helping children and students develop their skills and knowledge.

  10. **AI and Investments**: It’s noted that AI requires enormous investments and that it’s important to reduce the costs of AI applications to ensure economic viability.

  The video explicitly addresses several AI tools and models, including Anthropic (Claude), OpenAI (GPT-5.5, Codex), Gemini, as well as specific tools such as Cloud Code and Cursor. It is aimed at an audience with advanced knowledge of AI and technology.

• Claude Opus 4.8: THESE 7 Things change EVERYTHING now! (+ AI World Model Interview)
  31.5.2026, 08:15:31

  The video covers several recent developments in the field of AI and robotics. First, seven important updates of the new Opus 4.8 model from Anthropic are presented, including improved performance, lower costs, and new features such as Dynamic Workflows and Ultra Code Mode. These enable complex tasks to be processed autonomously over longer periods. Additionally, the introduction of digital IDs for humanoid robots in China is mentioned, as well as curious incidents with household robots in San Francisco.

  Another focus is the discussion of world models, particularly the Agora 1 multi-agent world model published by Odyssey. Dr. Jenny Sein Schwarz explains that world models don’t just predict the next token but model the entire state of the world, which can be crucial for applications in robotics and other fields. The advantages of world models over conventional language models are highlighted, as well as the challenges and competitive dynamics in this area.

  In conclusion, various business ideas and trends in the emerging agent economy are discussed, including the development of software for agents and the integration of private MCP servers into OpenAI. The video is aimed at advanced users who view AI not just as a novelty but as a tool for increasing efficiency and saving costs.

  Featured tools/models/providers: Opus 4.8 (Anthropic), Codex (OpenAI), Odyssey (Agora 1), 11 Labs (Dubbing Version 2), Runway, agentmail, Miao. For intermediate and advanced.

Fireship

No new videos in this period.

Greg Baugues

No new videos in this period.

AI and Strategy | Le SamourAI

No new videos in this period.

Julian Ivanov | AI Automation (1 new video)

• Hermes Agent: 8 Features That Make OpenClaw Look Outdated
  1.6.2026, 14:41:32

  The video introduces Hermes, an AI agent, and compares it with OpenClaw, another AI agent. The speaker emphasizes Hermes’ reliability and efficiency, particularly due to its self-improving memory and ability to self-optimize. Hermes is highlighted as the most-used application on Open Router, a platform for language models.

  The video covers eight new features of Hermes that were added in recent weeks:

  1. **Hermes Dashboard**: A user-friendly dashboard that displays and controls various settings and functions of Hermes. It enables management of sessions, models, cronjobs, skills, and plugins. An integrated Kanban board documents and tracks tasks.

  2. **Slashgoal Command**: Allows Hermes to work on a task until a specific goal is reached. A separate model verifies whether the goal has been achieved, and Hermes continues working until it’s confirmed.

  3. **Background Sessions**: Hermes can complete tasks in the background while the user continues communicating with it. This enables simultaneous processing of multiple tasks.

  4. **Session Search**: A full-text search through past conversations that allows Hermes to quickly access previous information without overloading its memory. This search is token-free and has been massively accelerated.

  5. **Curator**: A maintenance function that runs in the background and archives outdated skills and information to keep the system lean and fast. This prevents Hermes from slowing down over time.

  6. **Profiles**: The ability to create multiple Hermes agents with different specializations and memories. This keeps the system organized and efficient.

  7. **Kanban Orchestrator**: Breaks down large tasks into smaller subtasks and distributes them to the appropriate agents. This facilitates handling complex projects.

  8. **Kanban Swarm Feature**: Allows multiple agents to work in parallel on the same task, with a verifier checking the results and a synthesizer creating the final product. This leads to deeper and more thorough outcomes.

  The video also explains how to install Hermes on your own server and make the dashboard accessible via Cloud Code and Tails. It’s recommended to try out Hermes and share your experiences in the comments.

  **Final Note**: The video explicitly addresses Hermes and OpenClaw and is geared more toward intermediate and advanced users.

Kyle Balmer | AI with Kyle (4 new videos)

• Learn how to teach AI to businesses and earn $1000/hour
  5.6.2026, 11:45:07
• How To Get Paid For AI
  5.6.2026, 05:00:01

  The video covers the importance of public speaking (or “yapping,” as Gen Z calls it) in an era where artificial intelligence increasingly takes over tasks like writing and design. The speaker argues that AI pushes people into specific areas where they excel, such as authentic speaking and community building. He emphasizes that polished, professionally-looking content is often perceived as AI-generated and therefore generates less engagement. Instead, people should create raw, authentic content to build trust and authority.

  The speaker proposes a step-by-step approach to developing public speaking skills: starting with short videos on platforms like TikTok and Instagram, then live streams, followed by structured talks like webinars, and finally stage appearances. He warns against using AI tools like avatars or voice clones to fully replace yourself, as this undermines authenticity and floods the market with such content.

  The speaker also shares his personal strategy for building a business through content creation and public appearances, starting with free workshops and then conducting paid workshops for companies. He emphasizes the importance of focusing on a specific niche and continuously practicing to improve.

  Towards the end, he encourages viewers to simply start with their smartphones and regularly shoot videos to develop the habit of public speaking. He directs them to his webinars and newsletter for further information and resources.

  The video covers no specific AI tools or models and is best suited for Intermediate and Advanced users.

• AI 101: Tokens, Context Windows & Why AI Makes Dumb Mistakes Explained
  3.6.2026, 05:00:18

  The video provides an introduction to the concept of tokenization, which is fundamental to understanding Large Language Models (LLMs). It explains that tokens are the smallest units that LLMs work with, and that these tokens don’t always correspond exactly to words. The host demonstrates this using tools like Tiktokenizer and OpenAI’s Tokenizer to show how sentences are broken down into tokens. It emphasizes that one token roughly equals three-quarters of an English word and that tokenization depends on language, with English being particularly token-efficient.

  A central topic is the “Raspberry Problem,” where the model splits the word “raspberry” into three tokens, leading to errors because it doesn’t recognize the entire word as a single unit. The host also explains the significance of context windows, which represent a model’s working memory capacity, and how context compaction works to improve efficiency. It’s noted that larger context windows aren’t always better since they can lead to higher costs and less precise results.

  The video also addresses practical aspects, such as token costs when using APIs and the importance of selecting the right model and the appropriate amount of context for specific tasks. It’s recommended to use the smallest useful context window and feed models with relevant information rather than all available data.

  Finally, rules for efficient use of LLMs are summarized, and references are made to Andrej Karpathy’s resources, which provide in-depth explanations of LLMs.

  The video explicitly covers OpenAI, Claude, Gemini, and specific tools like Tiktokenizer and Artificial Analysis, and is suited for Intermediate and Advanced users.

• How to Use Codex Pulses (and Delete Your To-Do App)
  1.6.2026, 05:00:14

  The video “This is AI with Kyle” introduces the concept of “Codex Pulses,” inspired by Dan Shipper. Codex Pulses are specific threads within Codex used for organizing tasks, ideas, and notes. Kyle demonstrates how he uses different pulse threads like “Content,” “To-dos,” “Things to buy,” and “Questions” to boost his productivity. Each thread has specific functions, such as capturing ideas through voice recording, prioritizing tasks, or automatically researching and shopping. The pulse threads sync across the Codex mobile app with a central computer and provide daily summaries. Kyle emphasizes that this system is simpler and more focused than OpenClaw, as it handles specific tasks and requires less maintenance.

  The video explicitly covers Codex and the Codex mobile app, which are used to implement the pulse threads. It’s aimed at Intermediate users who are already familiar with AI tools and want to optimize their productivity.

Leon van Zyl (2 new videos)

• OpenCode + Ollama: I Replaced Claude Code With This (Full Setup)
  4.6.2026, 12:06:58

  The video demonstrates how to set up OpenCode with Ollama for a free, local coding agent. The creator describes issues with Claude Code, which overwhelms local models due to its high token consumption for system prompts and tools. As an alternative, OpenCode is introduced, which operates more efficiently and is better compatible with local models like those from Ollama. The setup process includes installing Ollama, downloading a suitable model (e.g., Llama 2 or Mistral 3.6), and connecting OpenCode with Ollama. The creator demonstrates using OpenCode to create a Next.js project and implement a local chat app, emphasizing the importance of small, focused instructions. The video also shows how to debug and test the app by granting the agent browser access.

  The video explicitly covers OpenCode, Ollama, and specific models like Llama 2 and Mistral 3.6. It is better suited for intermediate and advanced users who already have experience with local models and coding agents.

• Build Animated Websites with Claude Code + Higgsfield
  2.6.2026, 11:57:56

  The video shows how to combine Claude Code with Higgsfield AI to create and integrate an animated avatar on a landing page. Leon starts with a simple, static website layout and guides viewers through the steps of generating an avatar image with Higgsfield, animating it, and finally integrating it into the website. The Higgsfield CLI is installed and authenticated via an agent capability in Claude Code. Using a detailed prompt, Claude Code generates four variants of a friendly robot, and one is selected and inserted into the website. The avatar is then animated with Higgsfield AI and the C-Tons 2 model, adding subtle breathing and head movements as well as occasional blinking. After several adjustments, such as correcting the aspect ratio and background color, the video is embedded into the website and the audio is removed. Finally, small improvements are made to the website to better integrate the avatar and verify responsiveness.

  The video explicitly covers Claude Code and Higgsfield AI and is better suited for intermediate users.

Liam Ottley

No new videos in this period.

Mark Kashef (2 new videos)

• Master All 6 Claude Code Dynamic Workflows
  3.6.2026, 22:00:06

  The video explains six core patterns for effectively using dynamic workflows in Claude Code, helping to solve common issues like agent laziness, self-preference, and goal drift. The patterns include “Classify and Act” for task distribution, “Fan Out and Synthesis” for breaking down complex tasks, “Adversarial Verification” for checking results, “Generate and Filter” for idea generation, “Tournament” for decision-making, and “Loop Until Done” for iterative processes. The author demonstrates the practical application of these patterns through concrete examples and prompts, such as triaging emails, conducting due diligence checks, fact-checking blog posts, and optimizing codebases. The video also explains how workflows can be shared with other users and when it makes sense to use workflows. The video is geared toward intermediate to advanced users and specifically addresses Claude Code and its dynamic workflow features.

• 3 AMAZING Claude Code Dynamic Workflows (Opus 4.8)
  1.6.2026, 16:30:08

  The video showcases three concrete use cases for Claude’s new dynamic workflows, aimed at optimizing and personalizing the use of AI models like Claude Opus 4.8. The first use case demonstrates how dynamic workflows can be used to analyze all of a user’s previous conversations with Claude and create a customized report as well as a tutorial for optimal use of the new model. This enables individual adaptation of prompting strategies and avoids generic tutorials. The second use case shows how dynamic workflows can be used for in-depth research to verify and validate claims about new features. Numerous agents are deployed to check sources and create well-founded analysis. The third use case involves auditing your own Claude ecosystem to identify and optimize inefficient or redundant skills and rules. The video emphasizes that dynamic workflows are particularly suitable for complex tasks that justify high token usage.

  The video specifically addresses Claude’s dynamic workflows and is geared toward intermediate to advanced users who already have experience with AI models and their applications.

Matt Pocock

No new videos in this period.

Melvynx (2 new videos)

• MIGRATION de Claude à Codex : 1 seul ligne à faire (TUTO SIMPLE)
  4.6.2026, 16:00:38

  The video shows how to migrate configuration from Cloud to Codex to simplify the use of Skills and Agents and standardize configuration across different platforms like Cursor and Codex. The creator introduces an open-source CLI tool that migrates the entire configuration in less than 5 seconds using the command `npx ai-blueprint unify`. The tool creates symlinks for Skills and Agents so they’re available on both platforms. Additionally, a backup of the original configuration is created to prevent data loss. The creator emphasizes the importance of keeping configuration easily portable since tools can change over time. The video also introduces Lalal.ai, a tool that separates audio tracks and offers various audio editing functions.

  The video covers specific tools like Codex, Cursor, and the open-source CLI tool from ai-blueprint and is more geared toward intermediate users.

• Composer 2.5 : le modèle le plus intelligent et pas chère ?
  31.5.2026, 16:38:50

  The video covers the introduction of Cursor’s new model **Composer 2.5**, which is advertised as faster and more cost-effective yet more powerful than GPT 5.5 High and Opus 4.7 Extra High. The creator tests the model in various scenarios, including thumbnail creation and code generation. The tests show that while Composer 2.5 is fast, it doesn’t match leading models in terms of quality and reliability. The high RAM usage of Cursor is particularly criticized as it puts significant strain on the system. The creator concludes that Composer 2.5 is suitable for small tasks and quick fixes, but cannot compete with the best models like GPT 5.5 or Opus 4.7.

  Final note: The video specifically covers the AI model **Composer 2.5** from Cursor and is geared toward **intermediate** to **advanced** users.

n8n (1 new video)

• n8n Now Speaks OpenTelemetry: Production Tracing for Workflows
  4.6.2026, 12:58:23

  The video introduces the new OpenTelemetry support in n8n 2.22, which enables integrating workflow traces into existing distributed tracing setups. With just two configuration steps, workflow and node executions are included in traces, making it easier to visualize latency, errors, and anomalies. n8n uses OpenTelemetry (OTEL) and the W3C Trace Context standard, enabling seamless integration into existing tracing setups. The feature is also available in the free Community plan and offers customization options such as configurable sampling rates and the option to disable node spans. Users can also add their own metadata via telemetry tags. For setup from scratch, a trace collector and visualizer are required, with n8n forwarding the trace ID to downstream services, which are then responsible for sending the traces themselves.

  The video explicitly focuses on n8n and OpenTelemetry (OTEL) and targets intermediate users who are already familiar with distributed tracing or want to implement it.

Nate Herk | AI Automation (5 new videos)

• Is Claude Mythos Coming?
  6.6.2026, 21:26:20

  The video discusses recent developments surrounding Anthropic’s AI model Mythos, which briefly appeared on the company’s API and then disappeared. Mythos is a highly potent model specifically developed for cybersecurity and is considered the successor to Claude Opus. To date, it has only been made available to select partners as part of Project Glasswing to identify and fix security vulnerabilities in critical software.

  Mythos’s appearance on the API has sparked speculation that the model could soon become publicly available. However, Anthropic has already made it clear early on that Mythos should not be made generally accessible for now. The video creator is skeptical about whether Mythos will be released to the public in the near future and suspects that recent leaks might be deliberate marketing strategies to generate attention.

  Additional factors fueling speculation include Anthropic’s upcoming IPO and competition with OpenAI, which could also be bringing a new model, GPT-5.6, to market. The creator believes it is more likely that Mythos’s capabilities will be gradually integrated into future versions of Opus rather than Mythos being released as a standalone model.

  The video explicitly addresses Anthropic and OpenAI and is better suited for intermediate and advanced users.

• AGI is Here. Anthropic Just Proved It.
  5.6.2026, 20:42:32

  The video discusses current AI developments, particularly based on a report from Anthropic showing that over 80% of code shipped by Anthropic is now written by their own AI model Claude. The speaker argues that AGI (Artificial General Intelligence) is already here, defined as an AI’s ability to independently solve open, unclear problems. Anthropic categorizes tasks into four categories, from trivial to open, and shows that Claude has achieved a 76% success rate on open problems, a massive increase compared to 26% six months ago. AI can now handle tasks that take 12 hours and has even made better decisions than human researchers in some cases. The speaker discusses three possible future scenarios for AI: stagnation, continuous improvement under human control, and full AI autonomy. He emphasizes the importance of AI alignment to ensure it serves humanity and the difficulties in guaranteeing this, as progress is hard to monitor. The speaker warns against ignoring AI development and stresses that the ability to identify the right problems and deploy AI effectively is increasingly valuable.

  The video explicitly addresses Claude AI model from Anthropic and is better suited for intermediate or advanced users.

• The Skill That 10x’d My Claude Code Projects
  4.6.2026, 17:55:54

  The video addresses the challenge of transferring knowledge and context from your own mind into an AI system to achieve better and more personalized results. The focus is on a technique called “Grill Me,” which is designed to extract comprehensive information from the user and transform it into reusable context for the AI. This technique was originally developed by Matt PCO and consists of a simple but effective prompt structure that guides the user through a series of questions to reach a shared understanding. The video author has further developed this technique by adding a checkpoint system that ensures all answers are captured in a document to prevent information loss. This leads to better skills, context, and projects, as the AI develops a deeper understanding of the user and their processes. The author emphasizes the importance of this technique to improve the efficiency and quality of AI results.

  The video explicitly addresses Claude Opus 4.8 AI model and is better suited for intermediate or advanced users.

• I Tested Every Claude Code Feature, These 12 Are the Best
  3.6.2026, 16:16:11

  The video provides a detailed assessment and ranking of Claude’s features based on over 500 hours of use in the Claude ecosystem. The author rates the features based on their impact on daily workflows, particularly in the areas of knowledge work and automation. The rating ranges from D-Tier (useful but not outstanding) to S-Tier (essential for daily work).

  **D-Tier:**
  – Basic features like `clawmd`, context lifecycle management (`/clear`, `/compact`), web search, file uploads, fast mode, permissions, and custom themes.

  **C-Tier:**
  – Features that are useful but not used daily, such as voice mode, co-work, interactive connectors, and local file access.

  **B-Tier:**
  – Useful features not included in the top 30, such as dynamic workflows, deep research, Git worktrees, ultra review, interactive diagrams, recap feature, cloud for Microsoft 365, and Chrome.

  **A-Tier (Honorable Mentions):**
  – Features that barely missed the top 12, such as Google Workspace CLI, Dispatch, Claude Desktop App, Cloud Design, Auto Mode, Ultra Code Hooks, Effort Levels, Agent View, `/context`, Prompt Caching.

  **Top 12 Features:**
  1. **Skills:** Recipe-like instructions for consistent and repeatable tasks.
  2. **Status Line:** Displays model, effort level, and token usage.
  3. **Routines:** Planning and execution of tasks by agents.
  4. **Remote Control:** Control your local Claude session from phone or web.
  5. **/loop:** Repeat prompts at fixed intervals.
  6. **Sub Agents:** Background agents working in parallel.
  7. **/slashre:** Undo code and conversation to earlier checkpoints.
  8. **Agent Teams:** Groups of agents that debate and collaborate.
  9. **Auto Memory:** Automatic memory enhancement without manual prompts.
  10. **/slashinsights:** Generates reports on Claude usage.
  11. **Ultra Plan:** Offloads planning to various planning agents.
  12. **/slashgoal:** Sets goals for Claude that need to be achieved.

  The author emphasizes that the rating is subjective and based on his specific use of Claude. The video is better suited for intermediate to advanced users who already have experience with Claude and want to optimize their usage. It explicitly addresses Claude and its features.

• 100 Years of Artificial Intelligence Explained
  2.6.2026, 12:57:25

  The video tells the 100-year history of artificial intelligence (AI), starting with Alan Turing’s Bombe in World War II, which cracked the Enigma code, to modern AI models like Claude, ChatGPT, and Gemini. It describes the two main approaches to AI development: Marvin Minsky’s symbolic method and Frank Rosenblatt’s neural network method. After several AI winters marked by lack of progress and funding problems, the combination of powerful hardware (Nvidia graphics cards) and large datasets (ImageNet) led to breakthroughs in neural network research. Alex Krizhevsky’s AlexNet revolutionized image recognition, and DeepMind’s AlphaGo demonstrated that AI can make creative decisions. The introduction of the Transformer model and the development of language models like ChatGPT finally brought AI into the mainstream. Today, companies like OpenAI, Google, and Anthropic dominate the market, with Anthropic’s Cloud Code particularly popular among developers. The video concludes with the observation that the history of AI is far from over and new tools and features are being developed daily.

  The video explicitly addresses Claude (Anthropic), OpenAI (ChatGPT, Codex), Gemini (Google), and specific tools like Cloud Code and Anti-Gravity. It is better suited for intermediate and advanced users as it delves deeply into the history and technology of AI.

NeuralNine (3 new videos)

• This Is My Favorite Claude Code Feature…
  5.6.2026, 16:00:18

  The video demonstrates the use of the “Remote Control” feature of Claude Code, which allows you to control code sessions from a smartphone. Users can continue an active session on their computer from their phone via a QR code or link connection. This enables coding on the go, such as on the subway or in the bathroom. The process includes starting a session, activating remote control, and connecting the phone to the session. Users can then send commands and make changes from their phone and see results in real-time. The feature is particularly suitable for simple tasks like changing styles or brainstorming features, but is less suitable for complex programming work. The user emphasizes that the session should run in an isolated environment, as ending the session sometimes doesn’t work reliably. He recommends using multiple sessions in parallel to work on different projects.

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

• PostgreSQL Crash Course – Beginner Tutorial
  3.6.2026, 16:00:33

  The video is a crash course on PostgreSQL, a relational database management system. It begins with an introduction to PostgreSQL, highlighting its benefits such as being open-source, user-friendly, and extensible. The course covers PostgreSQL installation, both locally and via Docker, and demonstrates the use of pgAdmin, a web interface for managing PostgreSQL.

  The main part of the video covers fundamental SQL commands:
  – Creating and deleting tables (CREATE TABLE, DROP TABLE)
  – Inserting, updating, and deleting data (INSERT, UPDATE, DELETE)
  – Querying data (SELECT) with filtering (WHERE), sorting (ORDER BY), grouping (GROUP BY), and aggregation (COUNT, AVG, MIN, MAX)
  – Modifying tables (ALTER TABLE) and creating relationships between tables (JOINs, Foreign Keys)
  – Modeling many-to-many relationships through a separate junction table

  The video also shows how to manage the database from the command line using psql and how to use constraints like NOT NULL, UNIQUE, and CHECK.

  At the end, it is emphasized that this is a crash course for beginners and that intermediate and advanced courses could follow based on positive reception.

  **AI Tools/Models/Providers:** PostgreSQL, Docker, pgAdmin
  **Target Audience:** Beginners

• This tools shows you how much you spend on tokens…
  1.6.2026, 16:00:36

  The video introduces the open-source tool **CC usage**, which analyzes the usage of coding agent CLI tools on the system. It shows how much money you would theoretically spend on API costs, even if you’re using a subscription. Installation is done via various package managers like Bun X, Nix, PNPM, or AUR (for Arch Linux). The tool provides an overview of input and output tokens, total tokens, and estimated costs in US dollars. You can filter the data by provider (e.g., Claude, Codex, OpenCode), time period (daily, monthly), or session. Additionally, the output can be exported as a JSON object, which is useful for automations. An important note is that the data is only available locally on the system where the tool is running.

  The video explicitly addresses the open-source tools **CC usage**, **Claude**, **Codex**, and **OpenCode** and is aimed at **Intermediate** users.

Nic Conley

No new videos in this period.

Nick Saraev

No new videos in this period.

Niklas Steenfatt (2 new videos)

• I participated in the world’s biggest chess tournament!!!
  4.6.2026, 15:56:55

  The YouTuber reports on his participation in the Grankle Chess Open in Karlsruhe, one of the world’s largest chess tournaments. He describes his experiences over five days, playing against opponents with higher Elo ratings, including some strong players. He won some games, lost others, and experienced dramatic moments, such as defeating a stronger opponent who ran out of time. The YouTuber emphasizes the mental and physical strain of the tournament and the special atmosphere among chess enthusiasts. He analyzes some of his games and shows how he became more confident over time despite initial uncertainty. In the end, he achieved a total score of 5.5 out of nine points and reflects on the unique experiences and challenges of tournament chess.

  The video does not address any specific AI tools or models and is intended more for intermediate and advanced chess players.

• THE BEST WORKFLOW!!
  31.5.2026, 08:00:34

  The video shows how the user leverages Claude Code Remote to access and manage his server from anywhere. The user explains how he installs and configures Claude Code Remote on his Hostinger VPS to have access to his server at any time, even after a restart. He demonstrates how he can manage various agents in Paperclip with Claude Code Remote and query server information. The user emphasizes the benefits of this remote control, particularly the ability to monitor and repair the server even when facing issues like storage overflow or hacking attempts. He also mentions that he recommends Hostinger as a server provider and provides a discount code for its offerings.

  The user explicitly addresses Claude Code Remote and Hostinger, with the video being more suitable for intermediate users who already have basic knowledge of server management.

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

• We Need An Ecosystem in AI, And Every Company Can Win A Place In It
  4.6.2026, 10:00:31

  This video is a podcast interview with Satya Nadella, CEO of Microsoft, in which he discusses the future of AI, Microsoft’s strategy, and the societal impacts of AI. Nadella emphasizes the importance of tangible value and the need to earn society’s trust rather than just making promises. He talks about rapid scaling of Azure and the necessity to rethink work processes to harness AI’s seemingly impossible possibilities. Nadella also mentions the importance of education and the potential for new startups to develop new educational models. He discusses Microsoft’s role as a platform enabling companies to develop their own AI, and stresses the importance of private evaluations and the ability to improve models. Nadella also addresses AI’s impact on society and the need for AI’s benefits to be felt in communities. The video explicitly focuses on Microsoft’s AI strategy and is aimed more at intermediate and advanced audiences.

Productive Dude

No new videos in this period.

Sebastien Dubois

No new videos in this period.

Tech With Tim (2 new videos)

• Hermes Agent – Full Course & Setup Guide – For COMPLETE Beginners
  5.6.2026, 14:00:09

  This video is a comprehensive guide for beginners to install, configure, and use Hermes Agent. The creator walks through step-by-step how to set up Hermes Agent on a virtual private server (VPS), using Hostinger as an example. He emphasizes the importance of security measures and demonstrates how to securely manage API keys and access permissions.

  The creator explains the five core concepts of Hermes Agent: memory, skills, Chronos (scheduled tasks), personality, and self-improvement. He discusses the significance of Hermes Agent’s self-learning capability, which sets it apart from other agent frameworks like Open Claw.

  The creator shows how to connect Hermes Agent with various tools and services such as Gmail, Google Calendar, and Composio to create automated workflows. He demonstrates the creation of daily email triage and security audits, as well as other useful use cases.

  At the end of the video, the creator encourages viewers to develop creative and useful use cases for Hermes Agent, emphasizing that the agent’s effectiveness depends on proper configuration and integrations.

  The video explicitly covers Hermes Agent and is designed for beginners.

• Cursor just crushed Claude Code
  2.6.2026, 11:58:44

  This video presents a comparison between Cursor’s new Composer 2.5 model and other high-quality models like Opus 4.7 and GPT 5.5. The author finds that Composer 2.5 is significantly faster and more cost-effective, while delivering results of similar or even better quality. The price difference is substantial: Composer 2.5 costs $0.50 per task, while Opus 4.7 costs $7 per task. The author also emphasizes the importance of coding harnesses, which enhance model performance, and highlights that Cursor offers a particularly effective harness. A live demo shows that Composer 2.5 creates a working web application in minutes, while Opus 4.7 and GPT 5.5 take significantly longer and sometimes produce faulty results. The author also uses Here.dot.now for simple and free deployment of the created websites.

  The video explicitly covers the models and tools Cursor (Composer 2.5, Opus 4.7), OpenAI (GPT 5.5), and Here.dot.now, and is geared more toward intermediate or advanced users.

TheAIGRID (5 new videos)

• How To Use ChatGPT Memory (Updated Memory Guide)
  6.6.2026, 16:30:21

  The video explains OpenAI’s new Memory system for ChatGPT, which enables relevant and personalized responses. The system automatically remembers context from chats, files, and connected apps, reducing the need to repeat yourself. Users can manage, edit, and update their Memory summary in Settings under Personalization. The video demonstrates how to make specific corrections or exclude certain information for future responses. Additionally, users can view the sources of personalization and adjust them as needed. The video also addresses common questions, such as handling sensitive information, using temporary chats, and the differences between Memory and custom instructions.

  The video specifically covers OpenAI’s ChatGPT and is geared more toward intermediate users.

• Meta’s New AI Pendant Set To Shock AI World – Metas AI Pendant Explained
  4.6.2026, 23:00:31

  The video discusses Meta’s plans for a new AI wearable device, a small pendant that continuously records conversations and meetings and organizes them with AI. The technology is based on the acquired startup Limitless and is set to be tested starting in 2027. The device fits Meta’s strategy to sell hardware and attract users to AI services, but faces significant challenges. On one hand, it could increase productivity; on the other, there are major concerns regarding privacy and data protection. Meta must prove that the device is both useful and trustworthy, particularly given past data scandals and regulatory hurdles. The market for always-on listening wearables is difficult, as examples like Humane’s AI Pin and Friend AI demonstrate. Meta could be more successful with a gradual rollout in the enterprise sector before targeting the consumer market.

  **Final Note:** The video covers Meta’s AI models and specific tools such as the Limitless Pendant and Humane’s AI Pin, and is aimed at intermediate users.

• How to Install the Hermes Desktop App (Complete Setup Guide)
  4.6.2026, 02:30:27

  The video shows step-by-step how to install and set up the new Hermes desktop app. First, the app must be downloaded and installed from the Hermes website, which takes approximately 10 to 15 minutes. The installation does not replace the existing Hermes installation but rather merges with it.

  After installation, several settings must be configured:
  1. **Gateway**: Enter the session token and remote URL here to connect the desktop shell to a remote Hermes backend.
  2. **API Keys**: Various API keys can be configured here to use agents for different applications.
  3. **LLM Provider**: At least some LLM providers (such as DeepSeek) must be configured to interact with the model.
  4. **Model Selection**: Choose the model to use, e.g., DeepSeek V4 Pro.
  5. **Appearance**: Choose between different themes such as Dark Mode or Cyberpunk.
  6. **Memory and Context**: These options should be enabled to ensure consistency across chats.
  7. **Voice Chat**: Configure settings for voice interaction here.
  8. **Profiles**: Create and manage different profiles with different personalities.
  9. **Skills**: Equip agents with various capabilities that work depending on the model.
  10. **Cron Jobs**: Create and manage automated tasks.
  11. **Messages**: Configure interaction methods with the Hermes agent, including various chat apps and email services.
  12. **Usage Statistics**: Monitor API calls, estimated API costs, and token usage.

  The video specifically covers the Hermes desktop app and is geared more toward intermediate users already familiar with Hermes.

• Microsofts New Quantum Breakthrough is 1000X BETTER! – Majorana 2 Explained
  3.6.2026, 17:50:03

  The video reports on Microsoft’s announcement of its new quantum chip “Majorana 2,” which is reportedly 1,000 times more reliable than previous generations. The chip uses lead instead of aluminum as a superconductor, significantly improving qubit stability. Microsoft claims that qubits now remain stable for up to a minute, representing an enormous advance over microseconds. A key aspect of development was the use of AI agents to analyze data, automate measurements, and uncover design weaknesses. Microsoft sees this as evidence that AI can accelerate scientific research. The chip could become commercially viable by 2029 and solve problems that are unsolvable for classical supercomputers. Microsoft also introduced the AI platform “Microsoft Discovery,” which is now generally available. However, critics warn against accepting announcements uncritically, as reliability and scalability still need external verification.

  The video specifically covers Microsoft’s AI agents and the “Microsoft Discovery” platform and is geared more toward intermediate and advanced users.

• Nvidia Just Introduced 4 New Stunning AI Updates
  2.6.2026, 11:45:10

  The video summarizes Nvidia’s major announcements at GTC Taipei. First, Neotron 3 Ultra was unveiled, an open AI model with 550 billion parameters based on a hybrid Mamba-Transformer architecture. It is five times faster and 30% cheaper than other leading open models. Nvidia emphasizes the model’s openness, including training data and scripts, to enable developers to adapt and improve it.

  Nvidia also presented the new Vera CPU, specifically developed for the era of AI agents. Vera offers up to 3.6 terabytes per second of internal throughput and is designed to maximize GPU efficiency. The CPU is equipped with the Olympus core and uses LPDDR5X memory for lower latency and higher bandwidth.

  For the physical AI domain, Cosmos 3 was introduced, an open-world foundation model that helps robots and autonomous systems understand and predict the real world. Cosmos 3 is based on a new blend of Transformer architectures and offers both Nano and Super models for various use cases.

  Finally, Nvidia announced RTX Spark in collaboration with Microsoft, a super-chip that combines a Blackwell-RTX GPU with 6,144 CUDA cores and a 20-core Grace CPU. This chip aims to revolutionize personal computing by running AI agents locally and ushering in a new era of personal computing.

  The video specifically covers Nvidia’s Neotron 3 Ultra, Vera CPU, Cosmos 3, and RTX Spark and is geared more toward intermediate and advanced users.

Theo – t3․gg (5 new videos)

• Cloudflare bought Vite to destroy Vercel
  6.6.2026, 20:21:44

  The video discusses Cloudflare’s acquisition of Void Zero, the company behind the popular JavaScript tool Vite (short: Vit). The speaker emphasizes Vite’s importance for web development and the strengths of the ecosystem around Vite, such as Vite Test, Rollup, Oxe, and Vite Plus. The acquisition by Cloudflare is presented as strategically sound, as Cloudflare is now able to build a new cloud platform specifically tailored to the needs of developers and agents (AI-driven tools). The speaker highlights how integrating Vite into Cloudflare’s infrastructure significantly simplifies and accelerates application development and deployment. He compares this with the current situation at other platforms like Vercel and emphasizes the advantages of seamless integration between code and infrastructure. Additionally, the importance of AI agents for the future of web development is highlighted, as they can automate complex tasks and accelerate application development. The speaker also showcases an example project called Lakebed, which provides a new abstraction layer for application development and deployment specifically tailored to the needs of AI agents. He emphasizes that Cloudflare’s acquisition of Void Zero strengthens Cloudflare’s position in the cloud platform market and will significantly influence the future of web development.

  The video explicitly covers Vite, Cloudflare, and specific tools such as Browserbase, Infinite Red, and Lakebed. It’s aimed at intermediate and advanced developers.

• I miss when programmers were lazy.
  4.6.2026, 23:15:49

  The video discusses the three classic virtues of great programmers: laziness, impatience, and hubris, and how these change in the age of AI. Laziness drives programmers to create efficient abstractions that save time in the long run. Impatience leads them to want to solve problems quickly and correctly, while hubris drives them to work on challenging projects. The author emphasizes that these virtues lead to high-quality software, but warns that AI tools like Large Language Models (LLMs) can undermine these virtues. LLMs can generate large amounts of code without caring about quality or maintainability, leading to bloated and hard-to-maintain systems. The author argues that programmers must continue to take responsibility for creating simple and efficient abstractions, and that AI tools should only serve as aids to achieve these goals.

  The video explicitly covers Large Language Models (LLMs) and is aimed at intermediate or advanced programmers.

• More Prompts = Worse Code?
  3.6.2026, 00:06:00

  The video discusses “Technical Debt” and expands it to include the new concept of “Prompt Technical Debt”. The speaker emphasizes that technical debt has been a long-standing problem in software development and can be both worsened and alleviated by AI. He explains that prompts can also represent technical debt, as they are complex, difficult to maintain, and prone to becoming outdated, especially with model updates. The speaker mentions various tools and approaches, such as AgentMD, ClaudeMD, and T3 Code, and emphasizes the importance of keeping prompts minimal and specific to avoid technical debt. He recommends using pre-made AI tools from third-party vendors and leaving them unconfigured as much as possible to benefit from the work of the teams continuously adapting those prompts.

  The video explicitly covers tools such as Cursor, Claude, OpenAI, Gemini, T3 Code, and ArcJet, and is aimed at intermediate to advanced developers.

• This might be a Hot Take
  1.6.2026, 08:38:00

  **Summary:**

  The speaker discusses the impact of AI on engineers, particularly those who are considered “poor” engineers. He argues that AI makes these engineers less harmful by making better technical decisions and improving the quality of their work. He compares this to situations where experienced engineers are often frustrated when working with less competent colleagues. The speaker emphasizes that AI can significantly increase the productivity and quality of work of less experienced engineers while also shortening the learning curve.

  He also mentions the challenges of integrating AI into existing systems and emphasizes the importance of agents capable of logging into various services. He praises the company Work OS for its solution that enables agents to log into web services, emphasizing that this is an important step for the future of software development.

  The speaker also discusses different types of engineers who can benefit from AI, emphasizing that motivated and learning-oriented engineers will benefit the most, while those unwilling to further develop themselves will face challenges in the future. He concludes with an appeal to viewers to seize the opportunity and continually learn and improve.

  **Final Comment:**

  The video explicitly covers Claude, OpenAI, and specific tools such as Work OS. It’s aimed at intermediate and advanced developers.

• SWE-Bench is getting replaced???
  31.5.2026, 08:31:09

  The video discusses the reliability and relevance of benchmarks for AI models in software development. The speaker criticizes existing benchmarks like SWEBench Pro for their lack of realism, contamination from leaked solutions, and insufficient verification methods. He introduces a new benchmark called DeepSWE, developed by Data Curve, which uses more realistic, hand-written verifications and newly formulated tasks without pre-existing solutions. The results show a clear superiority of OpenAI models (particularly GPT-55) over other models like Anthropic’s Opus or Gemini 35 Flash. The speaker emphasizes the importance of realistic benchmarks for evaluating the actual performance of AI models in software development and encourages developers to create their own benchmarks to better understand and compare models.

  **AI Tools/Models/Providers:** OpenAI (GPT-55, GPT-54), Anthropic (Opus 47, Opus 46), Gemini 35 Flash, DeepSeek V4, GLM 5.1, Haiku, MiniAx, Cursor, Claude, Codeex, Gemini CLI, Cloud Code, Mini SWE Agent.
  **Target Audience:** Intermediate to Advanced.

Tim Carambat (1 new video)

• NVIDIA Just Put a 1-Petaflop Supercomputer In a Laptop?
  1.6.2026, 18:17:55

  The video covers the announcement of the Nvidia N1 and N1X platform, which were unveiled at the Computex event. Timothy Garenbat, founder of Anything LLM, shares his thoughts on these devices, which will launch as laptops featuring Nvidia’s GB10 Superchip. The devices will be available through OEMs like Surface and Dell, enabling users to load and run large language models locally. Garenbat discusses the technical specifications, particularly the ability to run models with 120 billion parameters and a million context tokens, and compares the N1 and N1X with the already available DGX Spark. He highlights the benefits of local AI usage but also warns about the challenges associated with Windows on ARM, especially for developers. Garenbat speculates about the device pricing and market positioning, but emphasizes they shouldn’t be seen as direct competition to Apple devices. He concludes with hope that the devices will be a good platform for local AI applications and asks the audience for their opinion.

  The video specifically covers Nvidia N1 and N1X, DGX Spark, Windows on ARM and is aimed more at intermediate and advanced users.

Unsupervised Learning (2 new videos)

• AI Predicts The Text Of Answers
  4.6.2026, 18:32:26

  The video discusses the thesis that AI systems don’t understand the world but merely predict the next token in text. The author argues this view is superficial and that AI actually possesses a form of understanding, albeit not in a human way. He explains that AI doesn’t complete random texts but generates specific answers to posed questions. To demonstrate this, he created a website called “AI understands.ai” where AI is confronted with new, never-before-seen whodunit mysteries, including fictional physical laws. The author shows that AI can solve these riddles by drawing logical conclusions, suggesting it possesses functional understanding of the world. However, it lacks human, experience-based understanding and emotions. The author proposes distinguishing between functional and experience-based understanding, with AI possessing the former but not the latter.

  The video addresses generic AI models like ChatGPT and is geared more towards intermediate and advanced users.

• A Conversation With Alastair Paterson
  1.6.2026, 15:00:09

  The video is a conversation between Alister and Daniel about the company Harmonic and its approach to safe and strategic AI adoption in enterprises. Harmonic aims to help companies deploy AI tools securely and effectively without hindering innovation. The conversation covers Harmonic’s evolution from monitoring chatbot usage to a comprehensive solution that promotes both security and productivity through AI tools.

  Harmonic now offers a solution that monitors both AI tool usage in browsers and on endpoints. They’ve introduced an MCP gateway solution that controls communication between different AI tools and data sources. They’ve also developed an endpoint agent that monitors activities on employee devices and ensures sensitive data isn’t shared uncontrollably.

  A central aspect of Harmonic’s approach is the ability to analyze AI tool usage in the enterprise and understand which tools are actually valuable and where the greatest productivity gains are achieved. They use small language models to understand and categorize AI tool usage, enabling companies to optimize AI tool deployment and increase productivity.

  Harmonic also offers a solution for monitoring and controlling AI agents created by employees. This enables enterprises to monitor AI agent usage and ensure they align with company goals and policies.

  The video explicitly addresses Harmonic, which is enterprise-focused and geared more towards intermediate or advanced users.

WorldofAI (6 new videos)

• Gemma 4 12B Is INCREDIBLE! BEST Local AI Coding Model! IS POWERFUL! (Fully Tested)
  6.6.2026, 07:18:22

  This video introduces Google’s new Gemma 4 12B model, designed as a powerful, multimodal model for local deployment on consumer hardware. It fills a gap in the Gemma family between small edge models and larger workstation models. The model is encoder-free, meaning it processes text, images, and audio directly without requiring separate encoders. This reduces memory requirements and latency while still delivering competitive performance. On a 24 GB GPU, the model runs approximately 75% faster than the larger 26B MoE model and supports a context window of up to 250K tokens. Despite some limitations compared to larger models like the Qwen 3.6 35B A3B, Gemma 4 12B offers an excellent balance between performance and practicality for consumer hardware.

  The video also demonstrates how the model is used in various applications such as creating frontend designs, Minecraft and operating system clones, and SVG and 3JS generation. It emphasizes that the model is surprisingly capable at code generation and logical reasoning, though it’s not the best model for these tasks. The video recommends the model for users with approximately 12 GB VRAM looking for a modern local AI for coding, vision, audio, and 3JS generation.

  **Final Comment:** The video explicitly addresses Google’s Gemma 4 12B model and is geared towards intermediate to advanced users.

• Claude Oceanus, Anthropic AGI Claims, GPT-5.6 Checkpoint, GLM 5.2, Nemotron 3 Ultra & More! AI NEWS!
  5.6.2026, 07:27:15

  This video provides a comprehensive overview of the latest developments in AI technology. It begins with leaks of Anthropic’s new Mythos model, which shows impressive outputs and potentially revolutionary capabilities. Anthropic is already working on a successor called Oceanus, which could be even more powerful. Meanwhile, OpenAI has unveiled a new GPT-5.6 checkpoint called Jewel Alpha, delivering impressive results in SVG generation and frontend development. Google has released Dream Beans, a personalized storytelling tool, and Gemma 4 12B, a powerful multimodal AI model. Nvidia has introduced Nemotron Ultra, a model for long-running AI agents that offers up to five times faster inference and 30% lower costs. Additionally, a new benchmarking tool called Agent Arena was introduced to measure AI model performance on real-world tasks. The video also addresses challenges in verifying AI-generated code and presents Test Sprite as a solution. It also highlights the potential recursivity of AI development, where AI systems are increasingly capable of developing their own successors.

  The video explicitly addresses Anthropic, OpenAI, Google, Nvidia, and specific tools like Test Sprite, Dream Beans, Nemotron Ultra, and Agent Arena. It’s geared towards intermediate and advanced users as it provides in-depth technical details and analysis.

• Hermes Agent Official NEW Desktop App – The 24/7 Self-Evolving AI Agent!
  4.6.2026, 04:54:16

  This video introduces the official desktop version of Hermes Agent, recently released by the News Research Team. This native application brings the full functionality of Hermes Agent to your computer, including configurations, API keys, sessions, skills, memory systems, and autonomous workflows. The desktop version is available for Windows, MacOS, and Linux and enables seamless integration with existing Hermes configurations. Additionally, Hermes’ web dashboard has been redesigned and now offers a complete management interface for browser management.

  The video demonstrates the installation and setup of the desktop app, including model and skill configuration. It shows how Hermes Agent can be used for tasks like creating a frontend app or summarizing Reddit discussions. The application also enables scheduling of cron jobs and managing multiple sessions simultaneously.

  Zapier is mentioned as a sponsor, focusing on their new MCP Plus SDK infrastructure, which enables AI models to interact with various applications and create complex workflows.

  The video is geared towards intermediate and advanced users, as it assumes viewers are familiar with AI agents and their configuration. It addresses specific tools like Hermes Agent and Zapier.

• GPT-5.6 Leaked, Mythos Benchmark Leaks, Hermes Desktop App, Qwen 3.7 Plus, & More! AI NEWS
  3.6.2026, 07:08:40

  This video provides an overview of the latest developments in the AI industry. Key topics include the upcoming release of OpenAI’s GPT 5.6, which could compete with Mythos Preview 1, and a massive update to Codeex that extends it beyond pure coding. World of AI has launched a new Vibe-Coding platform and benchmark to compare models across various use cases. Microsoft has introduced seven new AI models, including MAI Thinking One, which performs strongly in the reasoning category. Hermes Agent has released an official desktop app, and Alibaba has introduced Qwen 3.7 Plus, a multimodal model with strong coding capabilities. Anthropic has released updates for Cloud Code and a new CLI for the Claude platform. Google may be planning an update to Notebook LM based on the Gemini Omni model. Additionally, Microsoft has unveiled new hardware for interacting with AI agents, and the World Intelligence Expo in China showcased hyperrealistic humanoid robots.

  Final Comment: The video explicitly addresses OpenAI (GPT 5.6, Codeex), Microsoft (MAI Thinking One, MAI Code One Flash), Hermes Agent, Alibaba (Qwen 3.7 Plus), Anthropic (Cloud Code, CLI), Google (Notebook LM, Gemini Omni) and is geared towards intermediate and advanced users.

• GPT 5.5 vs Opus 4.8 vs Gemini 3.5 – Which Model Should You Use?
  2.6.2026, 07:35:39

  This video compares various AI models, particularly OpenAI’s GPT 5.5, Anthropic’s Claude Opus 4.8, and Google’s Gemini 3.5 Flash, as well as open models like MiniMax M3, based on their performance across different areas such as software development, frontend design, agentic workflows, and more. GPT 5.5 is highlighted as the most consistent and reliable model for complex coding tasks and agentic workflows, while Claude Opus 4.8 is praised for its design talent and Gemini 3.5 Flash for fast and cost-effective iterations. The creator also introduces their own benchmarking tool that allows users to evaluate models based on specific tasks and domains. The tool is available for $12 per month (or $10 with annual subscription) and provides access to an extensive prompt library and evaluation criteria.

  The video explicitly addresses OpenAI (GPT 5.5), Anthropic (Claude Opus 4.8), Google (Gemini 3.5 Flash), and open-source models (MiniMax M3) and is geared towards intermediate to advanced users.

• MiniMax M3 IS INSANE! BEST Opensource AI Model! Beats Opus 4.7 and 50x Cheaper! (Fully Tested)
  1.6.2026, 07:36:37

  This video introduces the latest version of MiniMax, the M3, positioned as the first open-source model combining three top-tier capabilities in one model. The M3 delivers peak performance in areas such as coding, multimodal reasoning, and long-running workflows, all at a fraction of the cost of proprietary models. The model supports up to one million tokens in its context window and was trained multimodally from the ground up, which translates into impressive visual reasoning capabilities. In benchmarks, M3 outperforms both GPT-5.5 and Gemini 3.1 Pro in several categories and comes close to Opus 4.7. Practical applications demonstrate its ability to autonomously handle complex tasks like optimizing CUDA kernels. M3 is accessible via the MiniMax API, the M-Code platform, and Open Router, with a current promotional pricing that significantly reduces input and output token costs. The video also shows practical tests of M3 in areas such as frontend development, 3D development, and SVG coding, with the model delivering excellent results across all domains. M3 is described as one of the most impressive model releases of the year, particularly due to its coding performance and cost-effective pricing.

  The video explicitly addresses MiniMax M3, an open-source model, and is geared towards intermediate to 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.