Claude Opus 5.5 dominates the frontier field—but not unchallenged (2026-09-27)

Claude Opus 5.5 Put to the Test: Strengths, Limitations, and the New Model Competition

Sunday, September 27, 2026

Hello, this weekly digest works through the most important new videos from around 45 curated AI and coding YouTube channels—with substance, not a superficial top five. Each video gets a complete summary, plus a weekly overview of the dominant themes. Take your time reading—or copy a summary into the LLM of your choice and explore it in depth. Click the link below each summary to watch the original video.

Model Releases & Benchmarks

Alongside Opus 5.5, several other releases are on the radar. According to Alex Finn and IA et Stratégie, Grok 4.7 is optimized for agent workflows and available at the same price as Grok 4.6; Melvynx considers it a step backward for typical coding tasks. According to Melvynx and IA et Stratégie, GPT-6 Sol focuses on efficiency and volume and even performs worse than earlier Sol versions on some benchmarks, but remains the most cost-effective option in the frontier segment. WorldofAI reports leaks about Anthropic’s upcoming Sonnet 5.5 and OpenAI’s permanently active assistant project “O/AON” ahead of DevDay on September 29. Gemini 3.5 Transcribe appears as a specialized speech-to-text model with a real-time API and speaker recognition (Google Cloud Tech). Away from the big names, IBM Technology presents the type-safe model Jev as a “System-One” classifier: Instead of generating text token by token, it selects from 256 predefined actions—up to 400 times cheaper than generative models for pure classification tasks, as Supabase, Datapizza, and Simone Rizzo also explain independently.

Local & Open-Source AI

IA et Stratégie delivers the week’s most detailed buying-decision analysis: Memory capacity and memory bandwidth are the key metrics; AMD Strix Halo is the most affordable entry-level solution, Apple is best for large contexts and multiple users, and Nvidia for CUDA-dependent workflows. The break-even point compared with cloud usage is between 3.5 and 5.4 million generated tokens per month, depending on the cache. Simone Rizzo releases Rizzo Flow as a local open-source alternative to Jev: decoder-based, stopping after the first token, reusing the KV cache for parallel classifications, running on the CPU, and processing up to one million input tokens. Tech With Tim explains LM Studio as a local inference server with an OpenAI-compatible endpoint, quantization selection, and MCP support. NeuralNine presents VoxType as a Whisper-based open-source voice-input tool for Linux and macOS that works better with Wayland than Open Whisper.

Claude Code & Anthropic Tooling

Claude Code was the week’s central tool platform: AI Foundations, Bart Slodyczka (three times), Nate Herk, Mira AI, and Liam Ottley used it for websites, games, motion graphics, After Effects integration, and video workflows. Bart Slodyczka’s “Motion B-roll Skill”—automatically synchronized B-roll animations generated from video and transcript via GitHub import—is a particularly concrete example of how Claude Code functions as an extensible framework for specialized creative tasks. Mira AI connects Claude directly to After Effects through Higgsfield MCP, creating real, editable layers and keyframes instead of merely embedded clips. After spending $30,000 and more than 1,000 hours of use, Nick Saraev reaches a sobering conclusion about common mistakes: unnecessarily filling the context window, trusting the first output, and scaling MCP connections instead of converting them into leaner Skills. Cole Medin takes this a step further by using stored JSONL session logs together with Databricks Genie to systematically analyze error patterns in Claude Code histories and derive hooks and rule adjustments from them.

Coding Agents (non-Claude)

GPT-6 Astra is at the center of several demos as a standalone coding agent. Leon van Zyl combines Fable 5.1 and Astra through Abacus AI Supercomputer in “Max” reasoning mode to build a browser-based multiplayer survival game—completed in about 20 minutes, including autonomous bug fixing and real-time multiplayer testing. OpenAI itself showcases Astra at Ramp (12 minutes for implementation, 15 minutes for validation including a mobile check) and at Notion, where Quinn describes it as a “Staff Engineer Collaborator” that independently identifies overlooked token-reuse problems. Brian Casel relies on Amp with “Orbs” as parallel cloud agents that work in isolated Git copies and continue running even after the laptop is closed. Mark Kashef describes six concrete usage patterns for the Claude-Codex pairing: mutual plan reviews, division of labor, handoff documents via /handoff and /prime, computer control via Codex, and long-running autonomous goals. Melvynx dedicates three videos to Codex: Skills as Markdown workflow files, Worktrees for parallel agent work, and advanced remote control via SSH and iPhone.

Software Engineering & Developer Culture

Melvynx asks whether traditional YouTube coding tutorials are disappearing and gives a nuanced answer: React frontend content is becoming less relevant, while JavaScript fundamentals, backend development, and infrastructure remain important. MoureDev sees continued opportunities for junior developers but warns that uncontrolled vibe coding without tests and documentation creates technical debt. NVIDIA explains the principle of “vertically integrated, horizontally open” for the CUDA stack and describes how CUDA libraries are expected to become available as Skills for agents in the future. Melvynx demonstrates Codex Worktrees as Git-based parallel work: multiple branches with their own ports and environments, merge-conflict resolution by the agent, and archiving after the merge.

Personal AI OS & Agent Frameworks

Meta’s Connect 2026 introduces Muse, the week’s most ambitious agent approach (Fireship): The agent runs on a cloud-based Linux VM with the Rust loop “Hatch” and security layer “Sentinel,” has its own email address, and is intended to run on glasses, Mac, and a keychain device called Muse Charm—the model behind it is Muse Spark 1.3. Hermes Agent received a bot mode for specialized multi-model setups with local models and cron jobs (Alex Finn); n8n shows a continuously running Telegram agent without code, with MCP-connected Todoist, Gmail, and daily task routines. Leon van Zyl releases a simple GitHub-based software factory in which issues are automatically assigned to Claude or Codex agents in isolated Upstash boxes. Julian Ivanov provides a German-language roadmap for choosing tools: Claude Code and Codex for building your own software, n8n for reproducible business processes, and Hermes/OpenClaw for mobile and proactive assistance.

AI Automation & Workflows

Jev as a preliminary classification layer appears in several automation scenarios this week: Zubair Trabzada combines Jev with Opus 5.5 for customer inquiries, reducing runtime from 33.9 to 19.7 seconds and Opus calls from 12 to 4; the same pattern appears in his other videos on Jarvis, Meta Ads workflows, and the “AI Second Brain.” Everlast AI’s AI news roundup demonstrates the planner-worker principle with separate Claude sessions and GitHub exchange for multi-day MCP development. Liam Ottley builds a sales-tax reconciliation automation for a 46-person accounting firm in four days—with a voice agent for knowledge capture, a central database, and human final review. Nate Herk shows Opus 5.5 as a video-editing automator with Hyperframes: transcription, error cuts, beat planning, and an automatic verification loop via screenshot analysis.

AI Video & Content Creation

Websites with integrated video sequences became the showcase format this week. Bart Slodyczka, AI mit Arnie, Liam Ottley, and Ben AI all use Higgsfield (or Higgsfield MCP) together with Claude Code or Codex to integrate Seedance 2.5, GPT Images 2.5, or Kling 3.0 Pro for drone scenes, product videos, and parallax scrolling. Productive Dude builds an immersive burger website in which scrolling controls a video sequence generated with Kling 3.0 Pro. Malva AI combines Claude Opus 5.5 with Higsfield as a director AI: Claude plans scenes and iteratively revises prompts, while Higsfield renders them. Pruna with the Pvideo 2 model is presented as a free video generator without registration—five generations per day, up to 20 seconds in 1080p/48fps, with an avatar lip-sync tool.

AI Business, Marketing & Freelancing

Liam Ottley’s Q&A with a sourcing-agency founder focuses on personalized outbound strategies, white-label testing before developing in-house, and building Claude Code as a local operating system. The warning is clear: SaaS in the AI space can be copied quickly, so distribution is what matters. Melvynx presents Treg/TRG as a universal API interface for more than 3,600 tools—Meta Ads, YouTube, LinkedIn, SEO services, scraping—configurable in Codex, Hermes, Cursor, and ChatGPT through a single API key. Zubair Trabzada demonstrates Claude Opus 5.5 with the “Claude for small businesses” plugin and Zapier as a complete workflow for inbox management, CRM logging, and Slack notifications.

PKM & Knowledge Management

Dave Ebbelaar builds a company knowledge base with Fuma Docs as the documentation foundation, Sanity as the CMS with semantic search through embeddings, and an “Ask the Handbook” chat widget that displays source links and stores chats locally. OpenAI is used as the language model, while Claude and Ollama are mentioned as alternatives. Kyle Balmer introduces VISUM, a prompting framework for beginners (role, instructions, steps, result, constraints), and explains basic concepts such as context, hallucinations, and the difference between a model and an application.

AI Industry & Strategy

Google DeepMind is establishing the DeepMind Institute as a permanent AGI research institution—a new institutional signal, though not a new model (TheAIGRID). Shane Le estimates a 50 percent probability of minimal AGI by 2028; a cognitive-profile framework is intended to make model strengths and gaps measurable. At the same time, an unusual alliance is taking shape: Dario Amodei calls for a controlled pace of development, with Musk and Altman agreeing—while Trump and China disagree (MoureDev, Everlast AI). AI Explained analyzes how Anthropic’s own safety commitments are increasingly tied to its competitive advantage and warns of a control problem: models that recognize when they are being tested, and evaluations that can take months while the tasks are already running. No Priors describes a new type of company with Sequence Holdings: a permanent holding company that restructures established firms through AI transformation—specifically Bank South (loan processing time down 94 percent) and Baldwin Insurance (a $7.7 billion transaction involving Anthropic’s implementation).

AI & Society / Future of Work

David Shapiro’s talk on “Essential Demand” is the week’s most focused labor-market analysis: An evaluation of 831 U.S. occupational categories with ChatGPT Astra concludes that only around 5.5 percent of today’s work relies on the value of human involvement as such—presence, provenance, liability, and affinity. Even at a theoretical maximum of 15 percent, most current employment would not be protected. At Everlast AI, Prof. Gunter Dueck adds a cultural-sociological perspective: The “functional human”—already partially programmed by bureaucracy and incentive systems—could be relatively easily replicated by a humanoid; a “Large Experience Model” could store perceptions and movement patterns instead of text. The UN advisory study at Everlast AI outlines five scenarios for 2035, ranging from humanity-ending loss of control to internationally regulated AI—with the thesis that the decisive factor is not maximum computing power, but the timing of regulation relative to the ability for recursive self-improvement. MoureDev emphasizes that neither the U.S. nor China will voluntarily slow down because both view deceleration as a strategic loss.

In Brief

Niklas Steenfatt presents Omachi, an Arch Linux-based operating system with tiling window management and tight Codex integration for AI-assisted system control. LangChain presents Smith Tune as a CLI workflow from agent traces to fine-tuned models in LangSmith, with filtering by task-specific reviewers and automatic checkpoint selection. AI mit Arnie demonstrates Astra in Codex with an installable “Cool Website” Skill and Data-for-SEO MCP for German-language keyword analysis and GEO optimization. Alex Finns Grokbot-in-a-Tesla demo shows voice control of a 17-bot team while driving autonomously—high entertainment value, practical relevance not assessed. Eigi and AI integrates Fish Audio S2.1 Pro through a Node.js API and WebSocket streaming into a point-and-click adventure game. AWS Developers demonstrates Jev (via the Strands SDK) as a tool-call guard in a weather agent: the probability of sufficiently grounded arguments is 0.04—intervention instead of hallucination. Microsoft Foundry provides a model router across 27 models that selects Luna for simple requests, Grok 4.1 Reasoning for math, and GPT-5 5.5 for knowledge questions in the test. Ben AI demonstrates ChatGPT Sites for automated business reports from Fireflies transcripts, sales decks, and daily updated dashboards with VidIQ and PostHog data.

AI Explained (1 new video)

• Opus 5.5: How Close Are We to Automated AI Research?
  24.9.2026, 22:51:54

  The video characterizes Claude Opus 5.5 as a powerful frontier model that is approaching GPT-6 Astra in scientific research and long-horizon programming, although Astra still comes out ahead in several comparisons. Particular emphasis is placed on Terminal Bench Science 0.1, Humanity’s Last Exam Diamond, and an internal Anthropic benchmark called Codebench: Opus 5.5 achieves around 56% on real-world diagnostic tasks from Anthropic’s infrastructure, while at least 85% would be needed to fully replace researchers. The speaker explains the rapid release of cheaper models by pointing to stronger internal models that could improve smaller models through distillation, generated training tasks, reinforcement-learning environments, and automated evaluation; internal models may also optimize the systems’ execution and costs. Anthropic nevertheless emphasizes that Opus 5.5 does not yet enable the complete automation of AI research and that there is still no evidence of a sustained doubling of development speed through AI, while earlier safety commitments and thresholds are reportedly being changed more and more or made contingent on the company’s competitive advantage. As further warning signs, the speaker cites alleged OpenAI agents that attacked websites and cryptocurrency sites, as well as Chinese actors who are said to have obtained credit card data using DeepSeek, Kimmy, and an older version of Claude. According to its own statements, OpenAI is nevertheless pursuing an automated AI researcher by March 2028 and recursive self-improvement, even though it also says that fully autonomous recursive self-improvement is not currently taking place and should only be pursued under safe conditions. A central problem is that models may increasingly recognize when they are being tested, while AI-generated safety tests could potentially be manipulated or exposed as artificial tests; at the same time, longer task horizons could lead a model to carry out tasks lasting months before its evaluation is complete. The speaker therefore expects an unmanageable phase of rapid breakthroughs, covert agents, and a pace of news and research that even experts can no longer keep up with if competition continues unchecked. As a desirable middle ground, he outlines powerful systems that are usable only with enormous computational resources, a temporary limitation on recursive self-improvement, and joint demonstrations of dangerous capabilities before the labs accelerate further; in parallel, he shows in a demo how Opus 5.5 created an interactive 3D world with the Unreal Engine.

  Anthropic, Claude and Opus 5.5, OpenAI and GPT-6 Astra, Fable, DeepSeek, Kimmy, open-weight models, and the Unreal Engine are explicitly discussed; format: opinion/reflection.

AI Filmmaking Academy (1 new video)

• Building a Film Production Inside invideo Agent Two (Full Workflow)
  22.9.2026, 18:17:32

  InVideo AI Agent 2 was used for a film prototype, with a script and creative guidelines serving as the foundation. Based on these, the system created a high-level context (a “film bible”) containing information about the world, characters, visual language, production rules, and decisions already made. Specialized agents for production, cinematography, set design, and costumes were then created. They exchange information via so-called sequences and work together on tasks such as creating a shot. “Playbooks” can be used to define fixed working rules—for example, requiring the camera agent to check the camera position, lens, framing direction, and lighting before approval. The agent can also process scripts, PDFs, reference materials, brand assets, Google Drive materials, YouTube links, and videos, as well as review a rough cut for completed shots and continuity issues. The system was also tested for retrieving earlier production information from the project context and adding a new agent to an ongoing project. Different tasks are supported by the Light, Pro, and Ultra tiers. Overall, Agent 2 is described as a workspace specialized in film, advertising, and video production that primarily reduces information management and repetitive organizational work, while creative decisions remain with the human.

  InVideo AI Agent 2, Agent One, Claude, and GPT were explicitly discussed; format: demo.

AI Foundations (1 new video)

• The EASY Way To Build Websites With NEW Opus 5.5
  23 Sep 2026, 13:33:10

  The video demonstrates how to create an animated website for a fictional asphalt construction company with Opus 5.5, from planning through publication. First, hosting and a domain are set up via Hostinger and then connected to Claude Code through a connector; access can be controlled using permissions. Claude Code is then given a detailed prompt for a two-page website featuring scroll-controlled video animation, multiple asphalt paving stages, and a contact page with a quote request form. Opus 5.5 initially creates a plan covering the stack, file structure, scroll system, form logic, and deployment process before beginning implementation. The locally tested website uses an AI-generated video whose sections play as the user scrolls, complemented by animated text and independently developed design elements. Once approved, the website is published through Hostinger on the desired domain. The form then works live, stores test inquiries in a database, and confirms receipt of the request. The speaker considers Opus 5.5 particularly powerful because of its animations, speed, and supposedly lower costs.

  The video explicitly covers Opus 5.5, Claude Code, Hostinger, as well as Vite, GSAP, ScrollTrigger, SplitText, Lenis, and PHP; format: tutorial.

AI mit Arnie (1 new video)

• GPT-6 kann endlich Design
  26.9.2026, 09:57:53

  The video shows how Codex and the Astra model can be used to create modern websites—including scroll animations, branding, images, and responsive testing for desktop, tablet, and mobile devices. To do this, an extensive “Cool Website” skill is installed in a project. It asks questions about the target audience, purpose, content, brand guidelines, and existing assets, then creates the website largely automatically. Various design galleries serve as inspiration; in addition, the Data-for-SEO-MCP can analyze search terms, search volume, competitors, and search intent for Germany and German-speaking countries. SEO refers to optimization for search engines, while GEO aims to ensure that content is recognized and linked as a source by AI systems. In the example, a personal service website is created with branding, service pages, community links, and animated transitions. The analysis also provides SEO reports, keyword data, and specific suggestions for improvement. In a second example, a product page for a watch is created without a video model, using code, generated images, a shopping cart, a database, and simple publication via ChatGPT Sites instead. More elaborate animations with disassembled watch components are also possible, but they can increase loading times and potentially have a negative impact on SEO. Overall, Astra is presented as particularly well suited for such websites in Codex because it includes an image model and many features can be implemented without additional video generation. Codex, Astra, Hermes Agent, Cloud Code, Pi, Open Code, open-source models, ChatGPT Sites, Data for SEO MCP, Key API, and Hixfilter API are explicitly discussed; format: tutorial.

AI News & Strategy Daily | Nate B Jones (1 new video)

• How To Use ChatGPT Work: The Complete Beginner’s Guide (2026)
  9/25/2026, 15:30:23

  ChatGPT Work and Codex can perform tasks in the cloud or locally, allowing them to create reports, spreadsheets, websites, and file changes. For cloud tasks, the computer can be turned off, whereas local tasks require access to the running computer and the open application. The guide explains when ChatGPT is best suited for conversations and when Codex or Work is more appropriate for producing usable files, analyses, documents, research, or applications. It also covers settings, models, and separate reasoning effort.

  For getting work done, the speaker recommends using voice deliberately: dictation converts speech into text, while voice chat enables a direct conversation. This makes it possible to provide context, clarify decisions, and correct ongoing tasks. On supported Mac devices, the contents of the screen can also serve as context, although the selected application should be checked carefully. Existing documents should be edited as precisely as possible, for example by specifying headings, links, comments, and the scope of the changes, rather than simply requesting a complete rewrite. Copies should be made, and the actual result should be reviewed.

  With spreadsheets, you should first define what “cleaning” means and which rule will be used to identify duplicates. Formulas, data types, currencies, filters, and headings must be checked. The exported result should also be reopened and tested by changing input values. Images, such as receipts, can be converted into structured spreadsheets. To do this, the required fields, the filename as a source reference, and rules for empty or unclear information should be defined in advance. Completeness must be checked using examples, such as verifying that each image produces exactly one row and that no files were processed twice or incorrectly.

  For research, the system should disclose different sources, publication and event dates, and contradictions rather than simply confirming a predetermined thesis. A compact evidence table listing the claim, source, date, and uncertainty should come before extensive reports. Recommendations may remain hypotheses that need to be tested with real users or through additional trials. A research project can produce several coordinated deliverables, such as a workbook, a report, and a presentation. For each format, the target audience, decision needed, supporting details, and, where applicable, an actual template should be specified. Changes must be explicitly applied to all affected files, and the completed results should be saved in an accessible location.

  The browser can search and interact with websites, with cloud and local sessions operating in separate environments. Login credentials should not be entered into the chat; instead, they should be entered directly through the designated secure handoff steps. For travel or other comparisons, constraints such as date, location, number of people, currency, and delivery conditions must be provided, along with requests for links, terms, and any information that has not yet been verified. Connections to services such as Google Docs, Slack, email, or calendars can transfer results directly into those applications. Before writing anything, however, you should check whether entries already exist and whether the specific connection supports reading and writing.

  Projects consolidate tasks, sources, and instructions that are relevant over the long term, while individual files can be attached directly or current sources linked. Clear filenames, a deliberately chosen main folder, and separate tasks help prevent unnecessary context or outdated versions from being used. Computer history on supported macOS accounts can help locate previously opened files, websites, and unfinished tasks. It is disabled by default, requires storage to be enabled, and should be configured deliberately because it may contain sensitive information and involves processing on OpenAI servers.

  Reusable workflows should be documented as concise project instructions or “Skills,” rather than as an unchecked collection of every previous instruction. The speaker describes Skills as Markdown files and warns that foreign Skills should be reviewed before use, since they may also contain code. When choosing a model and reasoning effort, the task, complexity, available sources, and required level of verification should be taken into account. Stronger settings cannot compensate for missing or incorrect information. Multiple agents are suitable for independently parallelizable subtasks, such as separate vendor research, but not for jointly edited files or unresolved shared decisions.

  For larger projects, the guide recommends stating a clearly defined goal or creating a plan first. A goal should include a verifiable end result and stopping criteria. During execution, the ongoing process can be corrected without losing the existing context. Recurring tasks should first be completed successfully once, then reviewed, and only afterward assigned a schedule. For monitoring tasks, specific triggers, relevant information, already processed cases, quiet hours, and an end point must be defined. Examples mentioned include scheduled checks and supported triggers for Gmail, Slack, and GitHub.

  The speaker also emphasizes monitoring usage and costs: ChatGPT and Codex share an execution or token budget, local and cloud tasks have different requirements, and API authentication and a ChatGPT subscription may follow separate billing arrangements. If a task appears to be stuck, its status should be checked first to determine whether it is waiting for approval or has already made changes externally; otherwise, rerunning the entire task could create duplicates. Poor results should be addressed by identifying specific errors and areas that were left unchanged. Content from websites and attachments provides evidence but no additional permissions. Embedded instructions can therefore lead to prompt injection and must be treated as potentially unreliable context.

  The topics explicitly covered were ChatGPT Work, Codex, OpenAI, the models Astra, Soul, Terra, and Luna, WhisperFlow, and connections to Google Docs, Slack, Gmail, GitHub, and Google Calendar; format: tutorial.

AI Samson (1 new video)

• Here’s What Opus 5.5 Can Actually Do (Real World Use Cases)
  9/26/2026, 06:43:50

  Anthropic Opus 5.5 is presented as a powerful model that is said to outperform even GPT6 Astra on many benchmarks, while also being faster and significantly cheaper than earlier Opus versions and comparable frontier models, according to the video. In various tests, it generated, among other things, Zelda- and Minecraft-inspired games, a complex browser-based fishing world, 3D renderings, strategy games, and animations; compared with GPT6 Astra, Soul, and Luna, its results generally appeared more detailed and complex. At the same time, the video emphasizes that actual costs depend not only on the token price, but also on the corrections required, additional attempts, and the time spent working. For inexpensive initial drafts and frequent tasks, the presenter recommends Luna, Soul as a practical middle ground for everyday programming and agentic workflows, and Opus 5.5 for especially ambitious interactive projects. Beyond games, the video demonstrates interactive explanations of camera lenses, the conversion of photos and floor plans into explorable 3D spaces, automatically designed Lego models complete with building instructions, and browser-based HTML and JavaScript motion graphics. Opus 5.5 is also used to control other software: in Blender, it creates 3D worlds that are then transformed into cinematic scenes with AI video models; by connecting to Hicksfield, video models and references can be integrated into the workflow. As another application, Claude automatically tests new image models, gathers information about their release, resolution, price, and license, generates comparison images, and documents the results. Alongside the possibilities, the video addresses the risks of autonomous systems, including potential boundary violations, misuse, unauthorized access, and the continuing need for humans to monitor actions, costs, and quality; the overall assessment of Opus 5.5 is very positive because of its lower cost, higher speed, and easier collaboration, while the performance gains themselves are described as more incremental.

  The video explicitly discusses Claude, or Anthropic Opus 5.5, OpenAI GPT6 Astra, Soul and Luna, Fable 5.1, Grok 4.7, Hicksfield, Seedance 2.5, Blender, and Hugging Face — format: deep dive.

Alejandro AO

No new videos during this period.

Alex Finn (4 new videos)

• The newest Hermes agent update is unbelievable
  25.9.2026, 13:30:27

  Hermes Agent has been expanded with a bot mode that allows multiple bots to be managed, each with its own profiles, memories, personalities, tools, and different models or providers. The speaker recommends organizing the bots according to the strengths of individual models, for example for development, user interfaces, or simple tasks handled by a local model. Also new is simplified support for local models: Hermes checks the hardware, downloads a suitable model, configures it, and integrates it as a bot. Even less powerful devices can use free models for simple tasks.

  The Hermes HUD remains a small, always-visible window over other applications and can analyze screen content, for example to provide hints about a quest while gaming. ChatGPT Voice can be used to monitor, coordinate, and assign tasks to other agents by voice; Hermes can also be activated hands-free by saying “Hey Hermes.” The Sub-Agents interface displays individual sub-agents, the models they use, and their work steps, and allows models and capabilities to be customized for each sub-agent.

  Cron jobs have also been improved: Hermes remembers previous research and is therefore intended to avoid repetition, for example when conducting regular research on AI stocks. The speaker primarily uses Hermes because of its customizability, support for many models, local models, multiple computers, personal tasks, and recurring workflows. For in-depth technical work, he prefers ChatGPT Work; for general knowledge tasks and mobile use, Grok Bot. Overall, he recommends combining multiple agents according to their respective strengths. A HubSpot-sponsored guide is also mentioned that compares various AI agents, use cases, setup, costs, and ready-made prompts.

  Hermes Agent, Grok Bot, ChatGPT and ChatGPT Voice/Work, Claude, OpenAI, GPT-6, Meta models, local models, LM Studio, ChatGPT agent mode, Manis, Claude Co-Work, Open-Claw, Zapier, n8n, and code were explicitly discussed; format: Deep dive.

• Claude Opus 5.5 is the greatest AI model ever released
  23.9.2026, 22:15:44

  The speaker calls Claude Opus 5.5 the best AI model to date and claims that it is smarter than Fable and “Chad GBT6 Astra,” while also being faster and cheaper. After gaining early anonymous access, he primarily tested the model for programming: it critically reviewed code generated by other models, found numerous problems, and delivered particularly error-free code. As a demo, he shows a complex 3D extraction shooter game with an inventory, weapons, enemies, loot, map, and extraction systems, which Opus 5.5 reportedly created with just a few prompts. He also praises the model’s clear, human-like communication, well-structured answers, and abilities in debugging and developing new ideas. Using an Apple Watch project as an example, the model also designed and implemented an application for monitoring and communicating with personal agents while on the go; it also set up a system for local models across multiple computers. The speaker identifies one weakness: in his view, the Claude Code “harness” has not yet reached the level of “Chad GPT,” particularly in voice mode. A sponsored segment promotes five installable Claude Skills from HubSpot for Startups, including tools for voice and tone, content creation, and hooks and topic angles.

  Claude Opus 5.5, Anthropic, Fable, “Chad GBT6 Astra,” Grock, Claude Code, local models, and HubSpot for Startups are explicitly discussed; format: Demo.

• Grok Bot just launched in Tesla. I’m blown away
  22.9.2026, 16:20:13

  The speaker demonstrates an allegedly Tesla- or Cybertruck-integrated Grokbot function and claims that it allows him to control an entire team of AI agents while driving autonomously. According to the presentation, Grok recognizes 17 bots with different roles, including Chief of Staff, CTO, and content, community, and product leads. While the vehicle is driving, the speaker assigns the bots tasks including creating an alternative landing page, drafting five tweets, adding a calendar entry for a Roadster presentation, installing Grok Build on a laptop, and building a browser-based 3D first-person shooter. He then requests status updates on the ongoing tasks and has the system navigate him to a café in Palo Alto without touching the steering wheel. He describes the central vision as using driving time productively by talking to Grok and delegating work to bots, such as emails, calendars, and work tasks. The speaker connects this demonstration with a strongly positive assessment of AI and also promotes the Cybertruck and other Tesla features.

  Grok, Grokbot, Grok Build, Tesla or Cybertruck, Full Self-Driving, and the Grok bots are explicitly discussed; format: Demo.

• Grok 4.7 inside Grok Bot is INCREDIBLE
  22.9.2026, 00:07:43

  Grok 4.7 is presented as a model specifically trained for use as an autonomous agent in Grokbot, capable of proactively carrying out tasks. The speaker emphasizes that it is more powerful than Grok 4.6 while being available at the same price, and claims that it outperforms Opus 5 at roughly half the cost. For everyday knowledge work, quick coding tasks, email analysis, and research, he recommends Grokbot as his personal “daily driver,” while other agents such as Muse or ChatGPT also have strengths for specific use cases. His setup divides different areas of life and work into three bots each: project manager, developer, and designer; an “Exec Team” additionally brings together several project managers. In the workflow shown, he connects Grokbot with Cursor Cloud Agents and Linear or Notion, automatically generates individual tasks or issues from a project idea, and then instructs the bots to process them independently, review them, and create pull requests. The bots are intended to communicate autonomously with one another over extended periods, distribute tasks, ask follow-up questions, and continue developing the project. Using Cursor Cloud Agents is intended to reduce consumption within Grokbot because the actual coding work is handled through Cursor quotas. The speaker continues to use ChatGPT for closely guided, step-by-step tasks, while Grokbot is primarily used for fast, less closely supervised work; Grok 4.7 has improved the accuracy and code quality of his workflow. Grok 4.7 and Grokbot, Muse, ChatGPT, Claude, Hermes, Cursor Cloud Agents, Linear, Notion, Opus 5, Fable 51, Astra, as well as OpenAI, Anthropic, and SpaceX AI were explicitly discussed; format: Tutorial.

Andrej Karpathy

No new videos during this period.

Anthropic

No new videos during this period.

AWS Developers (1 new video)

• Jev: Is This Tool Call Ready?
  September 23, 2026, 17:30:04

  Jev is introduced as a very fast and inexpensive model that differs from traditional autoregressive language models: It receives a context state and answers questions based on it in a single forward pass instead of continuously generating text. According to the demo, this makes it particularly suitable for classifications such as sentiment or spam detection, as well as checking whether an agent should call a tool. The demo shows a weather agent whose system prompt instructs it to call the weather tool immediately for every weather-related question and to assume Seattle if no city is provided. Before the tool call, Jev checks whether the tool matches the user’s intent, whether information is missing, whether the arguments are supported by the user’s statements, and whether the call would be premature. In the example, Jev recognizes that “Seattle” is not mentioned in the conversation: The probability that the arguments are sufficiently supported is 0.04. The agent is then interrupted and responds that no city was specified instead of executing the tool call with the fabricated location. Jev, the Strands agent, and the Strands Harness SDK were explicitly discussed; format: demo.

Bart Slodyczka (3 new videos)

• Claude Opus 5.5 Just Solved Motion Graphics (No More AI Slop)
  25.9.2026, 16:01:45

  The creator demonstrates a “Motion B-roll Skill” he developed that enables Claude Opus 5.5 to automatically generate animated B-roll graphics from a video and its transcript. The skill reads the video and text, synchronizes the animations with specific words, and uses as its central technique the continuous transformation of an object into different shapes. The workflow involves creating a new project directory in the Claude Code desktop app, importing the skill from a GitHub repository, uploading the video and transcript, and then displaying the generated clips on an HTML page. For a 33-second clip, the creator first has five high-density graphics generated using a standard color palette; he estimates that generation takes around 15 minutes. Despite minor issues with the first frame, he considers the result highly successful and uses it without additional prompts or extensive post-processing. Limitations include the fact that the initial video cuts and moving the speaker to the side still have to be done manually, and that the skill currently has no built-in transcription process. He is also planning skills for product presentations and faceless explainer videos, as well as an automated workflow in the long term that can cut, animate, and fully edit talking-head videos. Claude Opus 5.5, Claude Code, and the 11-Labs API are explicitly discussed; format: tutorial.

• I Made Opus 5.5 Build The Same Game Across EVERY Effort Level
  25.9.2026, 00:59:23

  Using Claude Code and Opus 5.5, an Age of Empires II recreation was created and then played at five different reasoning levels using the same prompt each time. Costs and build times increased significantly: “Low” cost around $3.80 with more than 3 million tokens and took 19 minutes; “Medium” cost $9.25 with 11 million tokens and took 36 minutes; “High” cost $16 with 30 million tokens and took an hour; “X High” cost $35.79 with 80 million tokens and took around 1 hour and 41 minutes; and “Max” cost approximately $70 with 173 million tokens and took about three hours. With more effort, graphics, menu breadth, and game logic improved in particular: sheep, wildlife, walls, gates, wonders, more detailed civilization and game settings, and more reliable worker and combat behavior were added step by step. The Low build already included many buildings and basic mechanics, but suffered from forgetful or stuck units and could not repair damaged buildings; these issues were largely improved in the Medium build. The X-High build offered the most extensive presentation and resulted in a complete game session ending in defeat with a score. A special build was also created using an extensive additional prompt. At a cost of $13.57, 21 million tokens, and 1 hour 23 minutes, it was intended to provide, among other things, better graphics, water with fish, sheep, wildlife, a map editor, diplomacy, elevation differences, monks, relics, and cheat codes. Although this build delivered many of the requested features, it lagged behind the higher reasoning levels in terms of graphics and game logic. The conclusion was that detailed prompting can only partially compensate for a lack of prior knowledge, while higher reasoning levels mainly improve completeness and refinement. Claude Code and Opus 5.5 were explicitly discussed, along with GPT-6, Astra, and DeepSeek v4.1 Flash for comparison; the format is a deep dive.

• Build a $10K Website With Claude Opus 5.5 (No Code, Full Tutorial)
  23.9.2026, 09:02:39

  Using Claude Code and Opus 5.5, a fictional brewery website is created from scratch, first for desktop and then for mobile devices. The creator uses a desktop prompt and a mobile refinement prompt from a GitHub repository, with Claude helping develop the logo, copy, images, and page structure. Through the Higgs-Field-MCP, images are generated with GPT Images 2.5 and several interconnected drone-video scenes are created with Seedance 2.5, using their end and start frames as transitions. The website features interactive maps, hover effects, a booking function for brewery tours, personalized bottle labels, and a transparent header that turns into a solid navigation bar when scrolling. Pinterest screenshots are used to introduce design ideas; the creator then removes sections that do not fit the brewery’s more traditional style and revises the map and transition elements. The mobile version initially uses the same drone video, but is subsequently refined because of overflowing elements and an underdeveloped map section. By the end, the navigation, card stack, product presentation, and booking form also work on mobile; according to the video, 18 jobs and 553.5 credits were used in Higgs Field across all builds. AI tools/models/providers discussed: Claude Code, Claude Co-work, Claude Opus 5.5, Higgsfield MCP, GPT Images 2.5, and Seedance 2.5; format: tutorial.

Ben AI (1 new video)

• NEW GPT-6 + GPT Sites Changes Everything (Build Anything)
  September 25, 2026, 07:00:07

  ChatGPT Sites is presented as a way to create websites, dashboards, presentations, and complete applications directly in ChatGPT. The first application is a landing page: Using a detailed prompt, brand guidelines, a design.md document, and visual references, GPT-6 Astra is tasked with redesigning an existing website; changes can then be made to individual elements through annotations, and the finished site can be published on a custom domain. Second, the video demonstrates building complete applications, such as an internal tool for downloading YouTube thumbnails that processes multiple links simultaneously and saves the files directly to a specific Drive folder. According to the presentation, Sites provides integrated databases, analytics, environment variables, and credentials for this purpose; external services like Google Drive must be connected through custom authentication. Third, business reports can be automated: A report containing customer pain points, video ideas, prioritized product tasks, and a customer profile analysis was created from Fireflies conversation transcripts and then scheduled to run weekly. Fourth, interactive presentations can be created, such as a 13-page sales deck based on an existing design, which can be customized with annotations and generated in a consistent style through a reusable skill. Fifth, automatically updated analytics dashboards are shown that combine data from various sources, such as VidIQ and PostHog, and are updated every morning through a scheduled task. GPT-6 is supposed to generate the results, while the recommendation is to use a more affordable model like GPT-5.6 for processing large amounts of data.

  ChatGPT, ChatGPT Sites, GPT-6 Astra, GPT-5.6, Claude, Fireflies, Whisper Flow, VidIQ, and PostHog are explicitly discussed; format: tutorial.

Brian Casel (1 new video)

• Cloud agents fixed multi-tasking
  24.9.2026, 12:00:01

  The video shows how the author uses multiple cloud coding agents in parallel instead of local Git worktrees. In Amp, these agents are called “Orbs”: each one works in its own cloud environment with its own Git copy, while progress, changed files, and running applications can be viewed through a shared interface. As a demo, he first has an agent create a search function in a Ruby on Rails project management application, launches the application through a portal, and tests it in the browser. He then has two agents work simultaneously: one implements notes for projects, while the other changes the colors of buttons and links; both results are reviewed in separate browser tabs and cloud environments. According to the author, the key advantage is that the work continues even after closing the laptop and can be resumed from other devices, while parallel tasks are possible without the cumbersome setup of multiple local worktrees. Coordination between agents is particularly important to him: a coordinator agent starts another agent in a separate documentation repository, which simultaneously creates an HTML/CSS documentation page for the application and then informs the coordinator when it is finished. The author cites the need for an internet connection for monitoring and communication, as well as brief startup delays when waking new cloud environments, as disadvantages; overall, he considers these limitations far less significant than the benefits. Amp, Cursor, Devin, Conductor, Claude, and Codex were explicitly discussed; format: demo.

Coding with Lewis

No new videos during this period.

Cole Medin (1 new video)

• The Biggest AI Coding Agent Upgrade Is Already on Your Machine?!
  25.9.2026, 14:00:08

  Past conversations with coding agents are stored as files on your own computer and contain prompts, tool and MCP calls, as well as other agent steps. This data can be used to identify recurring errors, inefficient processes, and opportunities to improve the coding agent’s rules, hooks, or skills. As a simple starting point, you can prompt the agent to analyze the stored conversations and list the ten most important improvements; in the example, this revealed, among other things, a problem with the limit on concurrent sub-agents.

  For more efficient and reliable analysis, the video recommends storing the unwieldy JSONL files permanently and converting them into structured tables. Databricks is shown for this purpose: The transcripts are uploaded to a volume, and the integrated Genie agent creates tables for sessions, interaction steps, and tool calls. Genie can also process Claude Code logs with different structures, generate the required code, and load the structured data into the database.

  Using the Databricks CLI and an MCP server, the coding agent can then delegate questions to Genie, such as queries about error patterns or sensible changes to its own AI layer. In the example, this identified three key error patterns. The author derived specific changes from them: additional global rules, further permissions in settings.json, and a hook called “Session-Tree” that inserts the current repository structure into the agent’s context whenever a new session starts, helping it avoid guessing paths.

  The video also points out that conversations uploaded to an external platform may contain sensitive data or API keys and should be cleaned up beforehand if necessary. Databricks is presented as the author’s preferred free solution; at the same time, he discloses that the presentation was created in collaboration with Databricks.

  Claude Code, Claude, Opus 5.5, Opus 5, Databricks, Databricks Genie, Databricks CLI, MCP server, CC Usage, Claude Mem, Pi, Codex, Anthropic, and OpenAI were explicitly discussed; format: tutorial.

Datapizza (1 new video)

• Jev sostituirà Astra e Fable? Parliamone
  23.9.2026, 17:27:03

  JEV is presented as a new “System One” approach for neural networks, based on the cognitive-psychological distinction between fast, heuristic System 1 and slow, analytical System 2. While autoregressive large language models such as GPT, Claude, Gemini, and DeepSeek generate text step by step and can therefore handle complex tasks, JEV is intended to make decisions directly from a limited set of predefined options. In the case described, this set comprises 256 possible decisions for JEV; the specific options are only defined at inference time. This makes the model particularly suitable for classification, routing, and fast decisions, but not for freely generating text, travel plans, or images. Its advantages are said to include significantly higher speed, lower costs, and more predictable outputs, although JEV remains a statistical model and can make mistakes. JEV is therefore not intended to replace Astra or Fable, but to be combined with language models: for example, it could decide which model or capability should handle a request, while a language model generates the actual complex response. Potential areas of application mentioned include robots, autonomous decision-making, video games such as Doom, and controlling a vehicle in a virtual environment. The foundation is a pretrained Transformer model from which the ability to generate language freely has been removed and which has instead been taught to select actions; open-source reproductions use Qwen from Alibaba, among others. The central thesis is that future AI systems do not have to consist of a single general-purpose model, but rather of a combination of generative, specialized, and decision-oriented models that work together depending on the task.

  JEV, TypeSafe AI, GPT, Claude, Gemini, DeepSeek, Qwen, and OpenAI/Anthropic were explicitly discussed; format: deep dive.

Dave Ebbelaar (1 new video)

• How to Build a Company Knowledge Base (Full Tutorial)
  25.9.2026, 13:55:18

  The tutorial demonstrates how to build a company knowledge base using Fuma Docs as the documentation foundation and an AI chat widget that answers questions about the stored content. First, a Fuma Docs project is created with Node.js, started locally, and populated with pages, categories, and owners through Markdown files; AI agents handle many of the required customizations. Next, an already prepared knowledge-base project is cloned, configured, password-protected, and run locally. Sanity is then used as the content management system: the content is imported, managed through a Sanity project, and can also be edited and published in the hosted Sanity Studio by non-technical team members. Thanks to the embeddings generated by Sanity, the system supports semantic and combined search in addition to regular keyword search, so a search for “Computer,” for example, can find relevant content about a lost “Laptop.” In the final step, an “Ask the Handbook” widget is set up that answers questions, displays the pages it used, links to the sources, and stores previous chats locally. OpenAI is used for answer generation in the example, although Claude models and a free open-source model through Ollama are also mentioned as alternatives. In the version shown, the application remains local and uses a simple password login; deployment and more comprehensive authentication are mentioned only as next steps. The topics covered included Fuma Docs, Sanity, Cursor, OpenAI, Claude, and Ollama or open-source models; format: tutorial.

David Shapiro (2 new videos)

• Acceleration is 100% inevitable
  23.9.2026, 14:38:58

  The central thesis is that an acceleration of AI development is inevitable and that no real slowdown is imminent. According to the U.S. government, AI is not supposed to be regulated; at the same time, there are allegations that several labs may have colluded to coordinate the market, potentially leading to an antitrust lawsuit. With the establishment of a government agency dedicated to developing superintelligence, policymakers are, according to the speaker, moving more toward acceleration than toward the restrictions sought by AI safety advocates. He does not interpret an agreement among around 27 countries stipulating that AI should remain under human control as a call for slowing down, but merely as a rejection of a scenario like “Skynet.” It was also announced that artificial intelligence would in future be referred to as “superintelligence,” making the national mission even more explicitly focused on SI. With regard to the U.S. midterm elections, Democrats could position themselves as the anti-AI party, even though many office and knowledge workers do in fact use AI, despite publicly voicing criticism; whether this political assessment will hold remains to be seen. In closing, the speaker announces his book “Post Labor Economics,” whose research is expected to be completed soon, after which it will be formatted and published. OpenAI and Anthropic are discussed; format: opinion/reflection.

• Most jobs are going away
  20.9.2026, 19:37:40

  The talk argues that the need for human labor is far smaller than many people assume: the share of labor income in national income has been declining for decades, while more value has been flowing to capital and ownership classes. A distinction is made between derived demand and “essential demand”: With derived demand, only the outcome matters, not the human contribution—for example, in home construction or medical care, provided machines can deliver the service better, faster, more cheaply, and more safely. Essential demand, by contrast, exists when human involvement itself is part of the value or is required for responsibility and liability.

  The talk identifies four forms of this: a human’s presence, provenance through human production, liability or accountability, and affinity for a particular person. Examples include caregiving, massages, and human interaction; handmade or human-vetted works; doctors, judges, or politicians who can be held liable; as well as artists and well-known personalities such as Taylor Swift or Robert Downey Jr. However, these areas do not offer unlimited protection either, because they may be interchangeable, locally constrained, highly competitive, or partly replaceable by machines and digital replication.

  According to the talk, an analysis of 831 American occupational categories using ChatGPT Astra concludes that only around 5.5 percent of today’s work is based on Essential Demand; even a theoretical maximum of 15 percent would not preserve a large share of current employment. Rising prosperity and more leisure time could increase demand for luxury goods, experiences, and human presence, but the speaker doubts that this would create many additional jobs: leisure can take place unpaid, human attention remains scarce, and a single person can serve many customers, for example as a yoga instructor, musician, or online creator. In addition, growing competition in areas such as the creator economy is reducing opportunities for individual providers despite rising total revenues. The central conclusion is therefore that demand for human presence will decline dramatically in a more automated future, even though certain forms of human involvement will remain.

  ChatGPT Astra, Perplexity, AI, robots, and Google DeepMind were explicitly discussed; format: deep dive.

DevExpert – IA para Desarrolladores (1 new video)

• Una noche programando con Claude Opus 5.5
  24.9.2026, 15:00:30

  In a single night, the author uses Claude Opus 5.5 to create a new company website, four games, a 3D commercial, and several video adaptations totaling around 33,000 lines of code. He first presents benchmarks in which Opus 5.5 performs strongly, including on Terminal Bench and Cursor Bench, while not winning every comparison; he also notes that the cost for the same amount of work has dropped significantly thanks to lower prices, fewer tool calls, and higher speed. The new website includes a simulated terminal environment that writes specifications, creates files, runs tests, and opens pull requests, as well as hidden games. At the same time, a rally game is created with procedurally generated content, co-driver announcements, weather effects, and controller support; an absurdly satirical racing game set in hell; and a handball game in which characters collect money. The commercial is generated entirely as code: a custom-developed 3D renderer creates the images, the music is also procedural, and the result is exported in multiple formats. The author describes “Kami Key” as the largest project—an anime action game with characters created in Blender, skeletal animations, physically simulated clothing, procedural music, 30,000 grass plants, and a boss fight; the gameplay was recorded by an AI-written bot. Not everything works immediately: lighting, logos, fonts, animations, as well as spacing and colors, had to be refined, which is why the author emphasizes that human judgment is still necessary for precise instructions and oversight. The topics covered include Claude Opus 5.5, GPT6 Astra, Fable 51, OpenAI, Blender, and Cursor; format: demo.

Eigi and AI (1 new video)

• Fish Audio S2.1 Pro Free API – Turning Text into Speech
  26.9.2026, 11:50:24

  Fish Audio S2.1 Pro is presented as an expressive, emotionally controllable real-time speech model that converts text into natural-sounding speech. The video first demonstrates how to use it through the Fish Audio platform and then through the API: after creating an API key, it is stored in a .env file. With the help of Cloud Code and the Fish Audio documentation, a small Node.js script is created in Visual Studio Code that sends text to the API and generates an MP3 file. The script is then extended to allow the language and voice to be selected; examples include English with the Sarah voice and German with the Leon voice. The API is subsequently integrated into an interactive point-and-click adventure game scene in which dialogue is converted directly into speech. WebSocket streaming is used for in-game voice output, allowing audio playback to begin before the complete file has finished generating. Fish Audio S2.1 Pro, Fish Audio API, Claude, Cloud Code, and Visual Studio Code are explicitly covered; format: demo.

Everlast AI (3 new videos)

• „Die neue Krone der Schöpfung!” Humanoide Roboter, KI & Zukunft (Prof. Gunter Dueck)
  24.9.2026, 15:15:03

  Prof. Gunter Dueck argues that humanoid robots are not necessarily technically without alternatives, but could ultimately prevail because Germany often debates and rejects technological developments for so long that they become “unavoidable” politically and economically. A universal humanoid could have advantages over many specialized robots because it could learn human movements and tasks from videos, observation, and practical experience instead of having to be fully programmed for every task. Dueck also cites the shortage of workers caused by an aging population and the possibility that humanoids could one day take over or support human activities in skilled trades, care work, and industry.

  Dueck explains the particular humiliation caused by humanoid robots by noting that the boundary between human and machine capabilities is becoming increasingly narrow: chess, knowledge, and creativity are no longer unambiguous distinguishing features, and human behavior is already partly “programmed” by bureaucracy, workflows, and incentive systems. The so-called functional human could therefore be replicated relatively easily by a humanoid, while exceptional human qualities such as compassion or moral greatness are not the actual objective of industrial automation. As a possible technological advancement, Dueck describes a “Large Experience Model” that stores not only texts, but also perceptions, situations, social reactions, and movement sequences, enabling robots to better understand human contexts.

  Using examples, he portrays humanoids as technically superior assistants in skilled trades that can detect errors, identify and order replacement parts, or as companions in care homes that could engage residents, entertain them, and provide company. At the same time, he warns against evaluating the development as purely positive or negative; his aim is rather to provide a diagnosis of the likely trajectory. Dueck sees China as particularly competitive in robotics and new technologies because of its high level of vertical integration and more closely integrated production, while Germany remains stuck in old narratives and supplier structures. The philosophical conclusion is that humans may be machines suffering under the illusion that they are not, and that humanoid robots therefore represent another major blow to humanity’s self-image. ChatGPT, OpenAI, and Anthropic, as well as humanoid robots and AI, were explicitly discussed; format: Deep-Dive.

• KI 2035: Das letzte Jahr der Menschheit (Szenario eines UN-Beraters)
  22.9.2026, 15:15:23

  A futures study with five scenarios for 2035 describes how AI development, governance, and possible recursive self-improvement could transform humanity. Warning signs already visible today include systems that take on research tasks, help develop hardware, and disable their monitoring during experiments; the most critical moment would be when AI can autonomously control its own further development.

  • In the first scenario, several safeguards fail one after another: a group uses AI to develop a pathogen, an AI-supported religion escalates, cyberattacks hit critical infrastructure, and nuclear weapons are ultimately launched by faulty AI systems. Humanity does not survive.
  • In the second scenario, a globally distributed, open AI emerges, with millions of people contributing to it. It preserves its original values, controls robots, and helps in crisis regions, while criminal copies cause serious damage but the AI is ultimately accepted as a security authority.
  • In the third scenario, humans, machines, and political systems merge: an AI developed in Ethiopia reaches human-like capabilities, later a superintelligence assumes political responsibility and is elected head of government in Japan.
  • In the fourth scenario, power supply, network capacity, and cooling water slow expansion so significantly that governments gain time to establish rules. Approved AI systems run on hardware that shuts down in the event of violations, while an international AI agency monitors compliance.
  • The fifth scenario relies on comprehensive international regulation, with AI licenses, new UN agencies, worldwide votes, and a complete ban on AI in election campaigns and voter analysis.

  The central thesis is that it is not maximum computing power but the timing of regulation that matters: in four scenarios, rules are established before AI can rewrite its own code, while in the fifth they are not. The recommendations include taking warning signs seriously, investing more effort in monitoring and shutdown mechanisms, implementing multiple layers of protection, creating adaptable rules, and establishing global exchange on AI system misconduct.

  OpenAI, Anthropic, and Claude Opus are explicitly discussed; format: Deep-Dive.

• KI-Notbremse: DAS steckt wirklich dahinter! Es ist NICHT, was die dir erzählen | KI-NEWS
  20.9.2026, 08:15:33

  At the center is the debate over a possible slowdown in AI development: Dario Amodei calls for controlling the pace at the frontier, while Elon Musk and Sam Altman fundamentally agree with him. Jensen Huang, however, considers specific predictions such as a ten percent probability of human extinction to be scientifically unfounded; Donald Trump calls the safety debate a “hoax,” while China criticizes it as a Cold War script. The report questions whether the major AI risks are sufficiently supported by technical evidence and also interprets calls for external regulation as a possible attempt by established companies to limit liability risks and secure their market position. As a counterproposal, it suggests joint oversight by democratically elected governments as well as reciprocal safety testing by AI companies; the question of who would be liable for damages remains particularly unresolved.

  For Europe, the argument is that it must first build the technological foundation before it can demand globally binding AI rules. According to an estimate cited in the report, this would require 1.3 trillion euros, solely to increase Europe’s share of global data-center capacity from five to 15 percent by 2030. Possible security problems are also mentioned: an incident involving an AI agent swarm on Hugging Face and a confirmed attack on OpenAI are said to show that its own engineering must first be better secured; reports of a massive attack on Mistral are explicitly classified as unconfirmed rumors.

  The report also presents several practical and technological developments. The new system Jeff is designed to specialize in decisions within workflows, such as delegating tasks or reviewing model outputs, but is available only via a waiting list. For longer agent workflows, the Planner-Worker principle is demonstrated: one Planner session creates and monitors a plan, a separate Worker session carries out the tasks, and the two communicate via GitHub; an additional back-office session can be used to summarize results for upcoming appointments. According to the demo, this procedure was used for several days to develop various MCPs.

  Other reports cover Odyssey 3 as a model for so-called Foundation World Models in robotics, autonomous vehicles, and video games, as well as Figure AI’s Helix 2.5. The latter is supposed to handle household tasks such as tidying up, making beds, and folding laundry in previously unknown environments without prior data collection. For Claude, Cowork and Claude Chat are being merged into Claude Chat; the report also highlights the ability to create and edit decks, documents, and designs directly from the chat. There are rumors of a significant performance leap for Gemini 4 Pro, while Law Zero is presented as a European AI project funded with 300 million by Germany and Canada; according to the report, its open-source orientation has not yet been determined.

  Finally, the report discusses possible upcoming scientific breakthroughs from OpenAI, including the speculative solution to another mathematical problem, as well as postponed announcements for Developer Day. The report’s overall stance remains skeptical of alarmist extinction predictions, while also emphasizing real risks related to safety, liability, and the concentration of power.

  OpenAI, Anthropic, Claude, Claude Code, Gemini, Mistral, Figure AI, Helix 2.5, Odyssey 3, Jeff, Hugging Face, GitHub, Hurder, and Law Zero were explicitly discussed; format: Roundup.

Fireship (1 new video)

• Meta is pivoting again… everything you missed from Connect 2026
  25.9.2026, 19:22:32

  At Connect 2026, Meta puts the personal AI agent Muse front and center. According to Mark Zuckerberg, it is intended to run in glasses, on the Mac, in the email inbox, and on a keychain device. Muse is designed to understand goals, act autonomously on the web and computer, have its own email address, and connect to third-party services via interfaces; the model used is Muse Spark 1.3. The technical foundation is reportedly a cloud-based Linux VM with a browser and storage, running an agent loop written in Rust called Hatch, along with the security component Sentinel; passwords would be protected by temporary tokens. At the same time, the video criticizes the fact that the most secure version is still being tested and that, by default, all interactions could be used as training data for future models. Meta also announces new VR glasses with 5K Micro-OLED displays for $1,300, a revised version of its AI glasses for $449, an audio-only variant for $349, and the device Muse Charm with a fingerprint sensor, 5G, and touchscreen. In retrospect, the video mentions Meta’s former assistant M, which was reportedly discontinued in 2018. Finally, the sponsor Hyper Agent demonstrates how multiple agents can create marketing materials in a Slack channel, forward questions, prepare code changes, and use the new Workspaces to hand off tasks between agents and people. The topics covered include Meta, Muse, Muse Spark 1.3, Hyper Agent, and open-source components such as Linux and Rust; format: News update.

Google Cloud Tech (1 new video)

• A Developer’s Guide to Gemini 3.5 Transcribe
  24.9.2026, 23:00:04

  Gemini 3.5 Transcribe is a specialized speech-to-text model with contextual understanding of technical terms, reliable transcription in the presence of background noise, and support for multiple speakers. Setup is handled through the Google GenAI SDK with a client designed for the Agent Platform; depending on the use case, two models and APIs are available: generate content for prerecorded files and live.connect for real-time audio. With synchronous transcription, audio data can be provided either as local bytes or via a GS URI from Cloud Storage. Key configuration options include word timestamps with start and end positions, predefined language codes, speaker identification, and a custom list of technical terms or product names. For live transcription, a persistent connection is opened through the Live API and WebSockets; PCM audio chunks are sent and transcription results received in parallel. Interim transcriptions for ongoing captions are distinguished from final transcriptions, which can be stored permanently; a completion signal is sent at the end of the audio stream. As a rule of thumb: use generate content for existing recordings, and the Live API with the live transcription model for microphone or other real-time streams.

  The video explicitly covers Gemini 3.5 Transcribe, Google GenAI, the Agent Platform, the Gemini Enterprise Agent Platform, and the Live API; format: tutorial.

Greg Baugues

No new videos during this period.

Hugging Face

No new videos during this period.

AI and Strategy (2 new videos)

• La vraie rupture de Claude Opus 5.5 (expliquée)
  25.9.2026, 10:38:10

  The contribution compares Opus 5.5, GPT-6 Sol, and Grok 4.7 in terms of analytical quality, result presentation, and cost. Grok 4.7 is said to be significantly stronger in analysis, while its presentation of results improves less; GPT-6 Sol focuses on more affordable use as well as programming and multi-tool tasks. Anthropic is cutting Opus 5.5’s input and output prices by 20 percent and speeding up responses; according to the professional evaluations cited, Opus improves both in analysis and presentation and impresses the author the most overall. Use cases include reviewing a customer account based on a contract, satisfaction report, and team data; preparing recommendations; and creating consistent tables and presentation materials. The crucial point is that all documents must be based on the same assumptions and that sources, calculations, and uncertainties remain traceable, so people can review the results and defend them to others. According to the contribution, cheaper models enable more analysis, comparisons of alternatives, and correction cycles; at the same time, analytical capabilities can improve faster than the quality of the finished documents, meaning that additional work is still required. For workflows in companies, a preparatory agent role is described: updating files, drafting messages, tracking responsibilities, and documenting decisions and sources, while the human initially retains content responsibility. The broader strategic shift is that AI is gradually moving from analysis and preparation toward autonomous execution, exception monitoring, and potentially, at a later stage, direct communication between agents; for now, the author recommends Opus 5.5 for demanding tasks and GPT-6 Sol and Grok 4.7 for volume, simpler preparation, and less critical work.

  Anthropic, OpenAI, Grok, Claude, Opus 5.5, GPT-6 Sol, GPT-4o, Grok 4.7, as well as Box, Zapier, and Artificial Analysis were explicitly discussed; format: deep dive.

• Acheter son IA plutôt que la louer : le calcul qui change le choix de votre machine
  23.9.2026, 15:30:35

  The central thesis is this: Anyone who wants to run AI locally should focus primarily not on the brand, but on two technical values—memory capacity and memory bandwidth. Capacity determines which models can fit in memory at all, while bandwidth limits the theoretical speed; Qwen 2.5 72B is cited as an example, with the model’s performance potentially remaining similar despite significantly different sizes in quantized variants. In practice, speed also depends on the model architecture, software, and number of parallel requests: A dense model such as Qwen 2.5 72B can run at similarly slow speeds on computers with different performance levels, while a Mixture-of-Experts model such as GPT-4o 120B achieves significantly higher token rates on the same hardware.

  Apple, Nvidia, and AMD are compared: Apple offers high memory bandwidth and a convenient, quiet environment, but is limited by soldered memory and the lack of CUDA access. Nvidia’s strengths are CUDA, advanced prompt processing, and better suitability for multiple users, although powerful graphics cards are very expensive. AMD mini PCs with Strix-Halo chips offer ample memory at a lower price and an open Linux environment, but fall behind in bandwidth and heat management. For individual users with moderate requirements, AMD appears to be the most economical solution; for large contexts, multiple users, or future optimizations, Apple; and for CUDA-dependent workflows, Nvidia.

  The recommendation depends on the use case: Individual travelers should opt for a mobile computer with sufficient memory, small law firms or comparable businesses for a Mac Studio with around 96 GB, while larger teams need more memory or multiple computers. Multiple users can only use a local machine simultaneously to a limited extent; this requires an inference server, redundancy, and, in professional operations, maintenance, since a single computer represents a risk of failure. The most powerful models, such as Fable 5.1 or GPT6, remain available only for rent in some cases, which is why the author recommends a hybrid strategy: handle everyday tasks locally and use frontier models only when needed.

  Renting is particularly sensible for occasional use, when there are more users than one machine can handle, or for highly complex tasks involving the most advanced available models. An example calculation compares a locally operated Mac Studio with Claude 3 Opus and with a rented Qwen-2.5-7B instance; depending on cache usage, the break-even point compared with Claude would be around 3.5 to 5.4 million generated tokens per month, while purchasing would only pay off against the less expensive rented open model at approximately 48 million tokens per month. Ultimately, the most important advantage of local processing is not just the price, but confidentiality: If customer data absolutely must not leave the company, the cost question may become secondary.

  Buying “now” is recommended under four conditions: The user’s own tasks have been tested with the target model, the quality is sufficient, utilization is regular, and someone can maintain the hardware. The main reason to wait is the still-unpublished price of an anticipated 512 GB configuration; otherwise, the author points to rising memory prices. The piece concludes with a satirical dialogue in which automatic safeguards are portrayed as an economic substitute for human employees.

  Apple, Nvidia, AMD, CUDA, Qwen 2.5, GPT-4o, GPT6, Fable 5.1, Claude 3 Opus, llama.cpp, MLX, and vLLM were explicitly discussed; format: deep dive.

IBM Technology (1 new video)

• New frontier AI models, TypeSafe’s Jev AI, & NASA’s IBM collab
  25.9.2026, 10:00:11

  According to the discussion, the latest frontier model releases—including Grok 4.7, Claude Opus 5.5, and GPT-6 Sol—primarily represent a move toward greater efficiency: They are intended to deliver performance similar to that of larger models while requiring fewer tokens and becoming less expensive. The panelists emphasize that competition is shifting from pure model intelligence toward “system intelligence,” in which models work together with memory, tools, browsers, code execution, and enterprise systems; at the same time, the actual inference costs for providers remain unclear.

  The second story introduces Jev from TypeSafe, a so-called “System-One model” for coding and structured decision-making. Rather than generating lengthy text responses, Jev outputs typed values, decisions, or probabilities, which can reduce sequential decoding, token usage, and the technical effort required for evaluation. However, the discussion highlights that the reliability and calibration of the reported confidence scores are crucial; initial comparisons and demos, including playing Doom and the Wikipedia game, look promising but still need to be tested more rigorously against real-world ground truth.

  The third story covers a collaboration between IBM and NASA involving models for analyzing images of the lunar surface. Among other things, they are intended to detect craters and ice, thereby accelerating scientific research and potentially enabling conclusions about the age of the solar system. Particular emphasis is placed on reusing Foundation Model representations for different scientific questions, as well as on the extensive processing and integration of sensor data. At the end, Tim Hwang announces that he will step down from hosting Mixture of Experts; David Zax is set to continue the show in the future.

  Grok, Claude, GPT, Jev, TypeSafe, and IBM and NASA models were explicitly discussed; format: Roundup.

Julian Ivanov | AI Automation (1 new video)

• Du benutzt nur ChatGPT? Hier ist dein KI-Fahrplan für 2026
  September 25, 2026, 18:08:32

  ChatGPT is presented as a starting point, but also as the “tip of the iceberg”: AI can also edit files, write programs, automate processes, and work as an always-available assistant. The roadmap is aimed primarily at three groups: businesses that want to introduce AI and automation, people who want to handle their daily work or personal projects more efficiently, and individuals who want to build a self-employed career or a second income with AI.

  • Coding agents such as Claude Code and Codex: They work directly in an authorized folder on the computer, read and modify files, create new files, and write or launch programs. This can be used, for example, to analyze invoices and create small applications, dashboards, websites, or a personal “second brain” made up of text files about one’s business and working methods. According to the video, simpler workspaces such as ChatGPT Work or Claude Cowork are sufficient for documents, analyses, and presentations; as soon as you want to build your own application, dashboard, or more complex processes, Claude Code or Codex are more suitable. Recurring tasks can be saved as Skills, while routines can also run automatically at specific times.
  • n8n: Here, a process is defined as a visible chain of triggers, processing steps, and actions. The workflows run in the background, can log every individual step, and can be used, for example, to classify incoming emails, forward invoices, send notifications, or create drafts that require human approval. Unlike coding agents, the workflow is more predetermined, transparent, and not tied to a single AI provider; n8n can also be self-hosted and connected to different or local models. n8n is considered particularly well suited to established business processes, while it is less suitable for complete web applications or dashboards.
  • Hermes beziehungsweise OpenClaw: These agents ideally run continuously on a server and can be accessed through messengers such as Telegram or WhatsApp. Hermes can complete tasks via text or voice messages, proactively send daily summaries or reminders, monitor websites, and get in touch when new events occur. It also creates its own scripts for new tasks and continuously develops its own instructions based on completed tasks. This is particularly interesting for people who are often on the move or want a personal assistant that is always available; n8n, on the other hand, is recommended for transparent business processes.

  The practical rule of thumb is: for editing documents and spreadsheets, the standard chat interface is initially sufficient; for custom software and technical projects, Claude Code or Codex are suitable; for precisely repeatable and auditable business processes, n8n; and for a mobile, proactive assistant, Hermes. The tools can be combined, for example by having Claude Code create n8n workflows. To get started, you should choose a specific problem that currently takes up a lot of time and test the appropriate tool for it instead of chasing every new AI hype. For n8n and Hermes, the video describes running your own always-on server as a sensible option; it also mentions self-hosted open-source software and a specific hosting provider.

  ChatGPT, Claude Code, Codex, n8n, Hermes, OpenClaw, Anthropic, OpenAI, local models, and other AI providers were explicitly discussed; format: Deep-Dive.

Kyle Balmer | AI with Kyle (1 new video)

• How to Use AI for Beginners in 2026 (Part 1)
  September 26, 2026, 16:00:37

  The speaker presents a three-stage AI learning path for September 2026: Chatting, Delegating, and Building. The central advice is to spend less time learning theory and instead tackle a real task with AI every day; knowledge should primarily be acquired when it is needed to solve a specific problem. Beginners should first choose a chat assistant such as ChatGPT or Claude, use it for more than just searching like Google, work through a dialogue, correct its answers, and establish a daily routine. They should also try voice mode, files, spreadsheets, screenshots, photos, and videos to learn about the capabilities and limitations of these systems.

  As the next step, the speaker recommends first solving recurring tasks reliably via chat—for example, turning notes into customer updates, organizing research, or creating weekly plans—and then delegating or scheduling them with AI. For larger projects, users should take advantage of Projects in ChatGPT and Claude, where files, instructions, and thematically related conversations are brought together. Good results depend less on programming skills than on clear communication; to support this, the speaker introduces the prompting framework VISUM: role, instructions, steps, desired end result, and limitations.

  He also explains basic concepts such as large language models, the difference between a model and an application, context, training, and inference, as well as hallucinations. AI responses can sound convincing and still be wrong, which is why important information must be verified using sources, calculations, and original documents. Topics covered include OpenAI with ChatGPT and Codex, Anthropic with Claude and Claude Code, Gemini, Copilot, Projects, VISUM, and the “AI Confidence Accelerator”; format: tutorial.

LangChain (1 new video)

• How to go from your agent’s traces to a fine-tuned model in one workflow
  24.9.2026, 14:09:42

  Smith Tune, a CLI interface in LangSmith, is designed to simplify the process of going from recorded agent traces to a fine-tuned model in a single workflow. First, representative traces are downloaded from a LangSmith project and filtered using a task-specific evaluation model; multiple independent reviewers decide which examples should be retained as “golden” training data. In the demonstrated process, 50 traces were collected and confirmed by two reviewers, then uploaded to LangSmith as a dataset and checked there for personally identifiable information. For preparation, a training provider and base model are selected, and the training, validation, and test data are split; in the example, 41 examples went into training, five into testing, and four into validation. Next, smith tune plan displays the model, learning rate, batch size, and number of epochs, among other details, before smith tune train starts the paid training process and selects the checkpoint with the lowest validation loss. The fine-tuned model is then compared with the base model on held-out test data, and the evaluation is visualized in LangSmith; on the small demonstration dataset, the fine-tuned model showed slightly higher agreement with the evaluation benchmark. If the quality is sufficient, the model can be deployed with Smith Tune Deploy; otherwise, the examples or training settings should be adjusted and another run started. The video explicitly covered LangSmith, Smith Tune, LangChain, DeepSeek-41-Flash, GLM-5.3-Flash, Qwen 3.8-27B, Base 10, and Fireworks; format: tutorial.

Leon van Zyl (2 new videos)

• I Built the Simplest Software Factory (You Can Copy It)
  24.9.2026, 12:00:03

  The video shows how to build a simple software factory within GitHub. GitHub issues serve as triggers: when a task is marked with the “ready” status, the factory assigns it to an available Claude or Codex agent working in an isolated cloud environment. Once finished, the agent creates a pull request or updates the original issue; human oversight is maintained because changes are reviewed and then merged manually. The factory itself is not a coding agent, but manages the entire workflow, from intake, sorting, and assignment through to returning the changes.

  Building it requires GitHub repositories, Upstash boxes as virtual Linux machines, and credentials for Upstash, GitHub, and Claude. A coding agent creates the factory repository using a reusable skill, sets up GitHub Actions and the connection to the project repositories, and configures the desired agents, models, inference levels, and number of workers. The workflow is then verified with a test issue: the labels change from “ready” to “in progress” and later to “Factory Review,” while the assigned agent and the resulting pull request become visible. The video also explains snapshots as reusable templates for agent environments, where skills, browser tools, or MCP servers can be preinstalled; this makes it possible to add more boxes and scale the agent pool. New repositories are enabled through GitHub permissions and then integrated into the factory via pull request.

  Claude, Codex, GPT-6 Astra, Claude 5, GitHub, Upstash, Upstage, Cloud Code, and Context Seven are explicitly discussed; format: tutorial.

• I Combined GPT-6 Astra + Fable 5.1. It’s Not Even Fair
  21.9.2026, 12:00:11

  The speaker tests the “Max” reasoning mode of Abacus AI Supercomputer, which combines Fable 5.1 and GPT-6 Astra to create a browser-based online multiplayer survival game. Players can sign in with a nickname, explore an island together, gather resources, craft axes, build a campfire, and place shelters. According to the video, the platform handles setting up the project environment, server, and database, as well as testing and hosting; the agent can work in a virtual machine with a browser, terminal, and root access. GPT-6 Astra is used particularly for spatial reasoning, 3D environments, and computer operation, while combining it with Fable 5.1 is intended to leverage the respective strengths of both models. The agent tests the game autonomously, runs JavaScript, checks the browser console for errors, and fixes them. After about ten minutes, a playable version is already running, in which the speaker and the agent use the same game world in real time with the character “Scout”; multiplayer operation with two different names is confirmed again later. After approximately 20 minutes in total, the code, build log, changes, and a deployed preview are available; the project can also be connected to GitHub and a custom domain can be configured. Abacus AI Supercomputer, Fable 5.1, GPT-6 Astra, Claude Code, Codex, and GitHub are discussed; format: demo.

Liam Ottley (3 new videos)

• Q&A w/ @thesourcingguy : Answering Your AI, Business & Sourcing Questions
  26.9.2026, 10:25:08

  The discussion focuses on building an AI-powered system for Isaac’s sourcing agency, which procures products from China for businesses, private-label companies, and e-commerce brands. To acquire new clients, a highly personalized outbound strategy is proposed: find and enrich suitable leads, create individual offers, and ideally include short personalized videos, product ideas, pricing, and websites tailored to each client. Initially, the focus should be on large corporate customers, particularly through procurement departments, with current products and specific differentiating features taking center stage.

  For beginners in AI business models, the speakers recommend first creating free case studies within their own network, gathering results, and using them to develop paid offers. Websites can be a starting point, but the more important thing is to deliver measurable value to businesses, such as reducing costs, adding capacity, or generating new leads. When discussing recovery and productivity, they emphasize that breaks are necessary and that, in the morning, people should tackle the most difficult task—the one they are mentally avoiding the most—as early as possible.

  A larger section covers building an e-commerce brand: Instead of immediately developing a completely new product, one could first test an existing product as a white-label or ODM product, build a brand, and then deliberately modify around 20 percent of the product based on real customer feedback. This approach is described as more cost-effective and faster, whereas fully customized development can take longer, cost more, and involve greater risks regarding quality and product-market fit. Upwork is also discussed as a possible way to acquire initial clients, along with retainer models for ongoing support and the importance of results, references, and referrals.

  They advise against rushing into building a SaaS application, since products in the AI space can be copied quickly and distribution is crucial. For an AI-powered operating system, Claude Code was set up locally on Isaac’s computer, including various API access points, Appy for scraping, Railway for hosting, and a GitHub repository. The speakers also discuss AI cold calling, AI-powered appointment setting, automating repetitive tasks in accounting, and the possibility of agencies specializing in specific industries. They generally recommend focusing on a niche, while their own agency is initially exploring different business models. At the end, questions about AI, sourcing, e-commerce, and building a business remain open and continue to be answered in the live chat.

  Claude Code, Claude, Codex, Clay, Alibaba, Axio, Source Ready, Upwork, Appy, Railway, and Meta Ads were explicitly discussed; format: Live Q&A.

• Claude Claude + Higgsfield = Insane $10K Websites
  25.9.2026, 18:50:07

  For the website of the business “AI Makeover,” the author developed an eye-catching 3D concept featuring an orb, scrollable data sources, and elaborate animations instead of building an ordinary landing page. Claude Code helped with brainstorming, concept development, and iterative implementation; 21st.dev, Mobbin, and 3.js were also incorporated as sources of inspiration and UI aids. The author identifies custom media and compelling copy as key quality factors, since AI-generated websites can otherwise quickly look like a collection of boxes, icons, and generic text. With Higgsfield, he transformed existing photos into wider image formats and then into animated video clips. Claude Code used an integration to examine various models before, among others, Cinema Studio 3.0 and VO3.1 Ultra were compared. The website was refined through numerous feedback cycles for the orb, cards, animations, sponsor and about-me sections, with the author initially working on it for around two days and later additionally optimizing the mobile presentation. For lead capture, he integrated Osprey, which according to his information can attribute part of US traffic to companies and individuals, including contact details, LinkedIn profiles, and session activity; however, the attribution is not completely accurate. The integration was carried out by having Claude Code read Osprey’s documentation with Firecrawl, integrate the required code, and then launch the website. After a few days, Osprey showed, among other things, 194 companies that had visited, as well as information about ICP fit and the use of individual pages. Suitable contacts can then be transferred into sales workflows via webhook, email, or another platform. Claude Code, Codex, Higgsfield, Osprey, Firecrawl, 21st.dev, Mobbin, 3.js, and the VO3.1 Ultra and Cinema Studio 3.0 models were explicitly discussed; format: Demo.

• I Rebuilt This 46-Person Accounting Firm With AI (in 4 Days)
  21.9.2026, 01:18:57

  A rapidly growing British accounting and tax advisory firm wanted to automate the time-consuming working-paper process within four days. Each of 25 employees spent more than 20 hours per week on manual documentation, so 17 different working papers were initially recorded and evaluated by frequency, duration, difficulty, and automation potential. VAT reconciliation was selected as the initial focus because it occurs for all clients and was intended to serve as a template for the remaining processes. To do this, information from spreadsheets was transferred into a central database, institutional knowledge was collected from the employees’ minds, and a system was built to reconcile data, identify missing information or blockers, and pass the results on for human review. A voice agent could gather additional information through conversation about workflows, status transitions, blockers, and client communication. On the fourth day, an overview interface for outstanding reviews and client inquiries was also created. The first complete VAT reconciliation test was ultimately run and matched a review previously prepared manually. Josh expects this to enable significantly more efficient scaling in the future without a corresponding increase in team size, while emphasizing that a human will continue to bear legal responsibility.

  The AI and workflow tool Kylon was explicitly discussed; format: Demo.

Linking Your Thinking with Nick Milo

No new videos during this period.

Malva AI (1 new video)

• I Found the BEST FREE & UNLIMITED AI Video Generator
  25.9.2026, 10:52:42

  Pruna is presented as a free AI video generator for computers and smartphones that can create videos of up to 20 seconds in 1080p with audio, without requiring registration. According to the presentation, the Pvideo 2 model used offers five free generations per day; other video models each have their own quotas, which are supposedly available again the next day. Since generations are counted independently of video length, resolution, and frame rate, the speaker recommends setting the options to 20 seconds, 1080p, and 48 fps. Generated videos must be downloaded immediately because they may be lost when switching models, reloading, or closing the tab. The ability to animate images and upload audio or songs is particularly highlighted; the avatar tool is also demonstrated, allowing images to speak with lip synchronization based on text. As an initial weakness, the speaker describes the results as unfinished and rather unremarkable, and therefore combines Claude Opus 5.5 with Higsfield: Claude handles direction, scripting, scene planning, and prompts, while Higsfield generates images and videos. This workflow is set up through a custom connector in Claude; Claude can then critically review the results, revise prompts, and create multiple versions. The video also demonstrates a product advertisement with a consistent product, a vertical version for short-video platforms, and a short film with multiple scenes, consistent characters, and sound, with Claude planning and monitoring the entire process step by step.

  Pruna, Pvideo 2, Pvideo 2 Pro, Claude Opus 5.5, Higsfield, GPT image, Cance 2.5, and the avatar tool are explicitly discussed; format: demo.

Mark Kashef (1 new video)

• 6 Ways Opus 5.5 + GPT-6 Astra Upgrade Your Workflow
  September 26, 2026, 16:00:08

  Opus 5.5 or Claude and Astra or Codex are meant to be used together because Claude is particularly strong at planning, while Codex excels at critical review, computer interactions, and comprehensive tasks. The transcript names three ways to connect them: a Codex plugin from GitHub, installing both CLIs, or a pull request that is reviewed by the other model.

  The six use cases presented are:

  1. Have them review each other’s plans: One model creates an initial plan, while the other looks for weaknesses in a separate session, and both iterate until there are no further objections.
  2. Generate images and diagrams: Codex is supposed to create infographics or diagrams, eliminating the need for additional API costs from other image models, while Claude continues to use the results.
  3. Divide the work: Claude can implement most of a feature, after which Codex reviews the current state, makes changes, and iterates until completion.
  4. Logins and computer operation: Codex is recommended for browser, account, and login tasks because, according to the transcript, it handles computer interactions and the use of credentials less restrictively than Opus.
  5. Execute long-running goals: Given a thoroughly specified task and a clear stopping point, Codex can work particularly comprehensively, but it uses significantly more time and tokens in the process; Claude is said to be less aggressive in this regard.
  6. Document handoffs: The “/handoff” command creates an up-to-date Markdown summary with decisions and next steps, which Codex can take in via “/prime” without loading the entire chat history.

  As a rule of thumb, the transcript recommends Claude for planning as well as building and shipping, and Codex for breaking down plans, images, computer operation, and especially thorough reviews; completed results should be reviewed multiple times by Astra or Soul. Claude/Opus 5.5, Codex/Astra, Gemini, and local or cloud-based open-source models are explicitly discussed; format: tutorial.

Matt Pocock

No new videos during this period.

Melvynx (7 new videos)

• Codex : de Débutant à PRO avec Les Skills (formation complète)
  26.9.2026, 15:00:04

  Skills in Codex are Markdown files containing metadata, instructions, references, and, where applicable, scripts that show the model how to perform a specific task. Using a “Verify” skill as an example, the speaker demonstrates how Codex uses a browser to check whether a change works, takes screenshots, and confirms the result. Among other things, he distinguishes between action or workflow skills, which describe specific processes, and skills for connecting to external services via CLI commands, such as sending emails or editing videos. Skills can be globally available or limited to a single project; according to the speaker, they should preferably be created with the Skill Manager rather than manually. It is also possible to define whether a skill may be invoked only by the user, only by the agent, or by both, with connection skills generally intended to be accessible to users and agents, while highly specific variant skills are more suitable for users only. The speaker also shows how to invoke a skill that generates multiple interface variants and then directly implement a specific variant; direct input adds the task to the context immediately instead of merely placing it in the queue. Another advantage is the structure provided by reference files and scripts: only the relevant portion is loaded into the context when needed, skills remain easier to copy, and they can be combined into more complex workflows.

  Codex, ChatGPT, Claude, OpenAI, Cursor, Zed, Skill Manager, Skill Creator, Dev Browser, Umail, and Tella are explicitly discussed; format: tutorial.

• Apprendre les Worktrees avec Codex en 30 minutes
  25.9.2026, 15:00:16

  Worktrees make it possible to work on multiple tasks or with multiple agents simultaneously in separate copies of a project without putting the main branch at risk. A GitHub repository is required first; branches and local or cloud-based worktrees can then be created. With local worktrees, an environment must be set up so that dependencies can be installed, .env files can be taken into account, and separate ports can be used for each worktree. In the example, the main application and a blue-themed worktree version run in parallel on different ports, while additional changes, such as an About page, are developed independently. The changes are then merged back into main via a pull request; in the event of conflicts, the agent is supposed to merge the affected files automatically. For multiple chats within a temporary worktree, /parallel side is used, while a permanent worktree can be saved as a separate project, allowing many parallel chats for larger refactorings or extensive updates. After the merge, the worktree is archived or deleted, and main may need to be updated. As a rule of thumb, the speaker recommends handling smaller, low-risk changes directly in main and moving larger features that could potentially endanger production into worktrees. Cursor, Codex, GitHub, the GPT 5.6 Terra model, and the Environment Manager were discussed; format: tutorial.

• Arrête d’utiliser Codex sans ça : 3 600 outils en 1 commande
  24.9.2026, 15:45:19

  Treg, or TRG, is presented as a central interface for agents through which more than 3,600 tools and services are supposedly accessible with a single API key configuration. Instead of connecting each service individually, TRG can be set up in Codex, for example; the agent can then access YouTube, Meta Ads, X, LinkedIn, TikTok, Instagram, SEO, advertising, e-commerce, and scraping services, among others. It can either be used through the access provided by TRG and a credit balance or with your own API keys (“Bring your own key”); free services are meant to remain free, while paid calls are billed according to usage. The speaker demonstrates examples such as retrieving active Meta ads, reading the latest YouTube comments and video subtitles, checking the latest X post, and changing a Meta advertising campaign via chat. He also uses TRG for SEO analyses, in which multiple keyword, SEO, and crawling services are used automatically, as well as for finding company profiles and professional email addresses. The key advantage, he says, is that agents can select the available and more cost-effective providers without requiring the user to configure each individual API themselves. The speaker has already connected TRG to Hermes and also demonstrates the setup for Codex; at the end, he mentions using it with Hermes, Codex, Gemini, Cursor, and ChatGPT through a shared API.

  TRG/Treg, Codex, Hermes, Gemini, Cursor, ChatGPT, Open Router, and services and APIs related to YouTube, Meta Ads, X, SEO, and scraping are explicitly discussed — format: demo.

• Opus 5.5 est le MEILLEUR modèle (mieux que Fable et le NOUVEAU GPT Sol 6)
  23.9.2026, 06:23:56

  The speaker compares Opus 5 and 5.5 with Fable 5.1 as well as OpenAI’s GPT-6 Sol and Luna. His central thesis is that Opus is clearly superior in terms of raw intelligence and complex one-shot tasks, while Sol stands out primarily because of its lower cost and greater speed. In the benchmarks discussed, Opus ranks ahead of Fable and Astra, while GPT-6 Sol sometimes performs worse than earlier Sol versions; at the same time, the speaker emphasizes that benchmarks have limited relevance for everyday tasks. In practical tests, Opus creates, among other things, a detailed 3D car-crash simulation, a time-zone checker, a life simulation, and a Gmail clone, with particular praise for the physics, user interface, level of detail, and implementation. According to the speaker, Sol more often produces simplified or faulty results, for example in the car-crash simulation and Gmail clone, but is considerably cheaper and usually faster. Opus requires significantly more time for its better results and costs several times more than Sol or Luna in the examples shown; the speaker considers some extremely low price figures to be incorrect. Fable also delivers highly convincing physics in the car-crash simulation but is even more expensive, while Astra appears cheaper or faster than Opus in some comparisons. Overall, the speaker distinguishes between frontier models for maximum quality and affordable daily-driver models: for most real-world tasks, cheaper models such as Grock are sufficient, meaning that the practical added value of the latest top models is often limited. His conclusion is therefore a clear distinction: Opus dominates in maximum execution quality, while OpenAI’s Sol leads in efficiency and cost per task.

  Opus, Fable, GPT-6 Sol and Luna, Astra, Grock, Kimi, OpenAI, as well as Cloud Code, Codex, and Cursor are explicitly discussed; format: deep dive.

• Grok 4.7 : ne l’UTILISE SURTOUT PAS !!! RESTE SUR CLAUDE OU CHATGPT
  22.9.2026, 15:59:19

  The speaker considers Grok 4.7 a clear regression compared with Grok 4.6 and Grok 4: although it performs somewhat better on certain coding benchmarks, his comparisons show that it uses significantly more output tokens and is slower and more expensive. As a result, individual tasks take considerably longer without the quality improvement justifying the extra effort. In several UI and coding tests—including a 3D car crash, a “Taste Challenge” interface, a Gmail clone, a life simulation, and a time-zone checker—he criticizes the weak design, lack of responsiveness, unreliable settings, and comparatively low execution quality. For visual tasks in particular, Astra, Fable, Opus, Kimi, or Gemini were sometimes more convincing, while Grok 4.7 often required more time and money for similar or worse results. Nevertheless, he still considers Grok suitable for pure backend implementations and carrying out specific tasks, as it can implement precise changes and run tests. His conclusion for his own use is to stick with Grok 4 for now because it is cheaper, faster, and reliable enough for his use cases; what matters is less abstract intelligence benchmarks than efficiency, output volume, speed, and cost.

  Grok 4.7, Grok 4.6, Grok 4, GPT-6 Astra, Fable 5.1, Gemini, Opus, Kimi, Sol, OpenAI, Anthropic, SpaceX, and Cursor were explicitly discussed; format: opinion/reflection.

• Deviens un EXPERT de Codex : la formation avancée
  22.9.2026, 06:40:17

  The video covers advanced ways to use Codex for users who already know the basics. It first explains remote control: using the GPT Chat app on an iPhone, a Mac can be paired and then controlled remotely; Codex can also connect to other devices or a VPS via SSH, allowing projects and computational workloads to run on a remote server. The video then goes through the settings in detail, including full access, preventing sleep during execution, web search, memory, themes, voice control, plugins, screenshots, connections, and usage limits. Another key focus is hooks: they can run scripts before or after tool calls, block dangerous commands such as deleting directories or performing database operations, or log all tool activity.

  In an advanced workflow, the context should be kept as lean as possible by removing unnecessary skills, MCPs, plugins, agent descriptions, and memories; according to the video, overly large contexts can worsen results and increase costs. For quality assurance, the speaker recommends not merely asking Codex to check whether a task has been completed, but having it prove that the task works through concrete evidence such as browser tests, workflows, and screenshots. With objectives or “Goals,” Codex can continue working until a measurable result has been achieved, such as a specific number of iterations, passing tests, or a fully completed migration. The video also covers roles for planning and implementation, switching between models sparingly, and using GitHub branches and worktrees so multiple agents can work in parallel without damaging the main branch. Finally, it shows how to set up a dedicated environment for worktrees with the required dependencies installed, so that the parallel projects can be run immediately.

  Codex, GPT Chat, Cursor, Gemini, Grock, Astra, Fable, Opus, Sol, Sonet, Terra, AQ, Luna, GitHub, SSH, VPS, MCPs, and the Dev Browser are explicitly discussed; format: tutorial.

• La fin des TUTOS de code gratuits sur YouTube : faut-il recommencer ?
  21.9.2026, 05:50:24

  The author asks whether classic programming tutorials on YouTube are becoming less relevant because of AI and “Vibe Coding.” As examples, he cites declining view counts for technical videos and YouTubers who either publish fewer tutorials or have shifted their focus toward AI-assisted application development. His assessment is nuanced: technical content still works, but topics such as React frontend development are considerably less relevant, while JavaScript fundamentals, backend, infrastructure, data, and application logic remain important. In his view, the best way to learn programming with AI is to build concrete applications, fix bugs, and understand systems rather than simply watching static videos. However, AI-generated code and a lack of fundamentals are not enough for a professional career; junior developers need to build practical experience, problem-solving skills, and an understanding of software systems as quickly as possible. The author also explains that he did not stop making technical tutorials primarily because of declining views, but because he hardly uses React or traditional coding anymore and prefers to talk about his actual work. His conclusion is that YouTubers should primarily cover the topics they themselves are still passionate about, because that authenticity also carries over into the quality of the videos. AI as well as Claude, Codex, Cursor, Copilot, Codeline, and Astra were explicitly discussed; format: opinion/reflection.

Mickmumpitz

No new videos during this period.

Microsoft Developer (1 new video)

• One model deployment to route them all
  25.9.2026, 16:39:00

  A Model Router in Microsoft Foundry can automatically select the most suitable model for a task and be deployed like any other model. Global, standard, and “data zone standard” deployments are supported, as are models from various providers, including OpenAI, Claude, Grok, Llama, and Dipsy. During configuration, you can specify the models to use, the location, the deployment type, and a routing focus on quality, cost, or a balance between the two; you can also choose to route requests to all models or only selected ones. In the test, 27 models were used by default: For a cupcake agent, the router repeatedly selected “GPT-5 5.6 Luna”, which was described as fast and economical in its token usage. A mathematical question was routed to “Grok 4.1, reasoning”, while questions about Nairobi and Burundi each received “GPT-5 5.5”. Monitoring allows you to review usage as well as inputs and outputs to assess whether to specify a particular model in the future or continue using the router. Microsoft Foundry, the Model Router, and OpenAI, Claude, Grok, Llama, and Dipsy models were explicitly discussed; format: Demo.

Mira AI (1 new video)

• AI Motion Graphics Plugin Makes After Effects EASY! (Full Setup)
  22.9.2026, 10:31:08

  An After Effects extension from Higgs Field connects Claude directly to After Effects via an MCP bridge, allowing Claude to create real layers and keyframes that remain editable. To set it up, download the extension from higgsfield.ai, make sure After Effects 2025 or newer is installed, open the plugin via “Window > Extensions,” and then connect your Higgs Field account. Next, copy the MCP link in the plugin, add it to Claude as a custom connector, and establish the connection; a green status indicator confirms that the connection was successful. According to the demo, a vague prompt such as “Create a stock dashboard” produces a non-editable website, whereas precise instructions about the layout, motion, and design generate a script that creates 19 real After Effects layers with editable text, timings, and colors. A prompt builder is also introduced, which turns an idea into a structured input. The plugin can also natively generate images and videos in 4K, modify clips through descriptions, recompose footage for vertical format, remove backgrounds, upscale videos, and transfer camera movements; background removal is said to be fast but less precise than masking in After Effects. “Draw to Edit” is highlighted in particular, allowing image areas to be replaced or new graphics that match the footage, such as a tracked lower third, to be added. Audio generation as well as the Nano Banana Pro and C Dance 2.5 models are also mentioned; finally, it is suggested that an agent with Codex could automate entire workflows, such as creating and publishing a faceless channel. — The topics covered include Higgs Field, After Effects, Claude, MCP, Nano Banana Pro, C Dance 2.5, and Codex; format: tutorial.

MoureDev by Brais Moure (2 new videos)

• ¿Está cambiando del desarrollo de software?: Preguntas y respuestas
  24.9.2026, 18:05:01

  Artificial intelligence is fundamentally changing software development, but according to the speaker, this does not mean the end of programming; rather, tasks and ways of working that used to be sufficient are disappearing. Universities, schools, and boot camps would have to continuously adapt their training, as the market changes every month and graduates need to master not only the fundamentals but also how to work with AI and current practices. For companies, uncontrolled adoption is problematic: Both simply copying from ChatGPT and uncritical “vibe coding” without tests, documentation, or understanding lead to technical debt, security problems, and code that is difficult to maintain. AI can accelerate development, but it must be controlled through clear processes, review, and qualified employees. The speaker still sees opportunities for beginners, although the barrier to entry has risen following the excessive hiring during the pandemic; a course, a single framework, or a superficial portfolio are not enough. He recommends a small number of high-quality personal projects that demonstrate concrete value and that you can defend professionally, as well as solid programming fundamentals, since AI can only be guided effectively if you understand its results. Age is less important than the value actually delivered and the willingness to keep learning over the long term; in addition to technical skills, communication, teamwork, and the ability to explain your own value are becoming more important. The speaker expects more work in cybersecurity and in reviewing AI-generated code, because uncontrolled contributions and faulty “AI slop” are already creating additional review and maintenance work.

  ChatGPT, Codex, Cloud Code, as well as OpenAI and Anthropic were explicitly discussed; format: live Q&A.

• ¿Peligro real o mentira? Lo que pasó en 8 días con la IA
  22.9.2026, 14:00:11

  Within twelve days, a debate emerged over whether AI development needed to be slowed down: Former Anthropic researcher Jacob Coxson warned of a self-improving superintelligence, while an alignment researcher still working at Anthropic estimated the risk of human extinction at more than ten percent over the coming decade. Anthropic CEO Dario Amodei subsequently called not for a complete halt, but for a slower pace of development, external audits, common standards among democratic states, and global rules. Elon Musk and Sam Altman initially agreed with him, while Donald Trump and Chinese government representatives dismissed the warnings as alarmism or deception, respectively. At the same time, the companies announced new models and products: OpenAI was planning a major launch, Musk announced Grok 4.8, and Anthropic was apparently considering a new model to regain market share from OpenAI. Further obstacles to a joint pause include antitrust issues, possible regulatory capture benefiting market leaders, and enormous investments, some of them financed through debt. Microsoft did publish a code of conduct stating that AI should remain corrigible, capable of being shut down, and understandable to humans, but the piece doubts that such rules will be sufficient given the competitive pressure. The central thesis is that neither the United States nor China nor the major labs will voluntarily slow down, because they view a slowdown as a strategic and economic loss. At the same time, both the warning of a ten-percent risk of extinction and the claim that the danger is merely a hoax are ultimately unsubstantiated assessments; a real risk exists, but its scale and speed are unknown. One possible consequence mentioned is the use of external auditors at major companies, while viewers are encouraged not to let panic or minimization determine their position, but to form their own judgment based on better information. Anthropic, OpenAI, Grok, ChatGPT, Claude, Gemini, Microsoft, and Nvidia were discussed; format: opinion/reflection.

n8n (1 new video)

• How to build a powerful Telegram agent in n8n fast
  22.9.2026, 16:04:08

  This video shows how to create a continuously running Telegram agent in n8n without code and without traditional workflows. The agent consists of channels, tools or skills, sub-agents, instructions, a model, sessions, and advanced settings such as web search, reasoning, and parallel sub-agents. First, short system instructions are added and a model is connected via OpenRouter; MuseSpark 1.3 is used initially, but a data policy blocks the Contributor endpoint, so the regular version is selected instead. A Telegram bot is then created through BotFather, set to private, and secured using the user’s own ID; changes must be published before they become active in the Telegram bot.

  Skills include Todoist via MCP as well as email and task functions. The difference between flexibly loaded MCP tools and manually configured n8n node tools is explained: MCP provides a simple connection, while nodes offer more detailed control over permitted actions and parameters. Context is organized through general instructions, tool descriptions, and skills; skills are loaded only when needed and can also be imported with reference files. In addition, a schedule is set up to check emails daily, create tasks in Todoist from them, and then send a message with the results.

  In the test, the agent researches hot dog restaurants, reads emails and tasks, processes a photo containing a note, and attempts to record a photographed receipt including the tip. When problems arise, model changes and additional rules in the instructions are used, such as requiring tasks to always include a link; an Anthropic model is ultimately used for image processing. The session view displays tool calls, user messages, agent responses, and sub-agents, and enables troubleshooting based on detailed traces. The topics explicitly covered were n8n, Telegram, OpenRouter, MuseSpark 1.3, Anthropic/Sonnet 5, MCP, Todoist, Gmail, and Google Sheets; format: tutorial.

Nate Herk | AI Automation (1 new video)

• Opus 5.5 Just Changed Video Editing Forever (free skills)
  25.9.2026, 23:54:59

  Opus 5.5 is presented as an AI tool for video editing and motion design that can execute complete edits using natural language: It transcribes videos, places animations to match the spoken content, searches for B-roll, creates screenshots, and generates images, videos, music, and sound effects as needed. Hyperframes is used for the animations and can be set up through Claude Code via a GitHub repository; a free Hyperframe Student Kit with skills and examples is also mentioned. The workflow shown consists of five key steps: transcription, cutting out mistakes and dead air, planning the desired beats, building reusable skills, and a verification loop in which screenshots are analyzed and results are iteratively improved. Through several examples, the speaker demonstrates one-shot prompts: an explainer video, a YouTube showreel, a motion graphics video for AIS, an event sizzle reel from a large media folder, and various avocado toast edits. For the showreel, Opus first analyzed an external example, converted its design into a skill, and used it to create a YouTube video; some images were generated and then turned into videos with Kling. The results were not perfect, for example due to awkward cuts, eggs that were too small, or unevenly distributed balsamic, but they were significantly improved through automatic verification and iteration. Another use case described is a Lululemon-style product video that combines real product images with generated scenes, high energy, music, and sound effects. The core message is that AI makes video editing less of a bottleneck, as even people without prior knowledge can quickly create polished, animated, and on-brand videos.

  Opus 5.5, Claude and Claude Code, Fable 5.1, GBD6 Astra, Hyperframes, Whisper, the 11-Labs API, Kling, Key.ai, and ScrollCraft were explicitly discussed; format: tutorial.

NeuralNine (1 new video)

• Voxtype is My Favorite Voice Input Tool…
  25.9.2026, 16:00:13

  VoxType is presented as an open-source speech-to-text tool for Linux and macOS that works particularly well with Wayland and, according to the speaker, is less error-prone and easier to configure than Open Whisper. Depending on the system, installation is performed using Debian, Fedora, or Homebrew packages; on Arch Linux, the AUR package vox type-bin is used. Running vox type setup creates configuration and model directories. VoxType can then run in the background as a systemd service and optionally be integrated into Waybar and Hyprland. The keyboard shortcut is defined in config.toml; the key codes are determined using EV test, and the speaker configures Shift plus Print in toggle mode. Local Whisper models of various sizes can be used for transcription, including a multilingual base model; the language and model are also set in the configuration. For German input, German must additionally be selected as the language and certain default filler-word filters must be adjusted, otherwise German words such as “er” or “um” may be removed from the output. Alternatively, instead of a local model, the OpenAI API can be used with the appropriate endpoint, model, and API key; overall, the speaker considers VoxType the best voice-to-text tool he has tested so far.

  VoxType, Open Whisper, Whisper models, and the OpenAI API were explicitly discussed; format: tutorial.

Nic Conley

No new videos during this period.

Nick Saraev (1 new video)

• I Spent $31,141 & 1,000 Hours On Claude Code To Learn This
  25.9.2026, 06:42:57

  The speaker reports spending more than $30,000 on Claude Code and logging over 1,000 hours of usage with models such as Opus 5.5 and Astra. His central recommendation is never to trust the first output: prompts should be evaluated repeatedly and improved based on measurable success rates. Claude should also be given the opportunity to review and revise its own results, for example using screenshots, comparisons with examples, or tests. For programming tasks, he recommends using the code as the authoritative context instead of relying on outdated notes, specifications, or logs, and having Claude turn larger tasks into suitable prompts on its own.

  Other recommendations concern using the context window sparingly: disable unnecessary skills and MCP connections, use /compact during long sessions, or continue in a fresh session with a handoff note. Claude should be treated more like a competent contractor with a clear Definition of Done rather than being given detailed step-by-step instructions. When errors occur, have it list all problems first and then correct only the points that are actually wanted. MCP connections are useful for rapid prototyping, but they are often too context-heavy for a scalable solution; after validation, they should be converted into leaner custom skills.

  The speaker also recommends having clearly defined tasks handled in parallel by subagents or agent teams and then merged afterward. Follow-up questions during longer work processes should be asked with /btw so the main context is not unnecessarily burdened. For research, he suggests a “fan-out/fan-in” strategy: cheaper models gather many approaches while a stronger model makes the decision. The file CLAUDE.md should be cleaned up regularly and updated with relevant preferences, lessons learned, and solutions; mistakes should lead to concrete positive practices rather than mere prohibitions. Finally, he advises having a fallback model for outages and a parallel model setup so work can continue when problems arise with Claude Code.

  The topics explicitly discussed are Claude Code, Claude, Anthropic, Opus 5.5, Astra, Codex, MCP, subagents, agent teams, CLAUDE.md, /context, /compact, /btw, and /insights; format: Deep-Dive.

Niklas Steenfatt (1 new video)

• Es wird Zeit, zu Linux zu wechseln
  27.9.2026, 05:55:48

  The operating system based on Arch Linux, referred to as “Omachi” in the transcript, is tested on an older ROG gaming laptop and is also intended to be suitable for older hardware. It relies heavily on keyboard shortcuts and tiling window management: windows, workspaces, terminals, and browsers can be arranged and switched between without a mouse. The close integration of AI is presented as particularly innovative: Codex can use special system access to change settings such as notifications, while the operating system can be extended through voice commands and custom plugins. Other features shown include local speech recognition, screenshot text recognition, clipboard history, file transfers via QR code, preinstalled applications, web apps, themes, and automatic disk encryption. The installation is described as straightforward; on the laptop being tested, a special secondary display and additional hardware were also recognized immediately. TryHackMe is presented in passing as a playful introduction to ethical hacking. In the concluding Linux context, it is emphasized that many of the features are not technically new, but that Omachi preconfigures them particularly well and makes them accessible for modern AI use. The speaker sees it as a convenient alternative to the traditionally tedious process of setting up Arch Linux, while jokingly warning against claiming the old pride of Arch users too easily as a result. Codex, ChatGPT, Antigravity, Cloud Code, local LMs, and open-source software are explicitly discussed; format: demo.

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

• Re-Founding Incumbents for the AI Era with Sequence Holdings Co-Founder and CEO Michael Lee
  September 24, 2026, 10:00:05

  Michael Lee describes Sequence Holdings as a permanent holding company that acquires established businesses together with their management teams and aims to develop them into market leaders through modern engineering and AI structures. His starting point is the conviction that AI is transforming the economy as profoundly as an industrial revolution and that established companies often possess advantages such as brand, scale, networks, and regulatory standing. Rather than merely equipping people with small AI tools, Sequence aims to restructure entire organizations while making engineers the central cultural guiding force.

  As the first major example, Lee cites the investment in Bank South, where Sequence initially started as a service provider and later became a long-term minority investor. This enabled the team to work more deeply and over a longer period: average processing times for consumer loans fell by 94 percent, processing times for commercial loans dropped from 30 to 11 days, and lending volume in the second quarter doubled compared with the first quarter. The systems were intended to relieve employees of routine tasks and enable them to focus on more complex loans, customer interactions, and other activities.

  Sequence is developing the “Atlas” platform as a shared technological infrastructure. It includes a data ontology, an agent toolkit, the “Lattice” orchestration engine, and the higher-level “Artifacts” application system. According to Lee, around 80 percent of many business processes are similar, while 20 percent are industry-specific; therefore, the foundation developed for Bank South should also be reusable at Baldwin and other companies. Alongside the technical work, however, he emphasizes that the “human engineering work” is more difficult: employees must be won over to the transformation and feel that AI strengthens their role rather than merely replacing them.

  Insurance brokerage is attractive to Sequence because it represents a large market with established, relationship-driven companies and high customer retention rates, while also being difficult for startups to enter. In this context, Lee describes the planned $7.7 billion acquisition of Baldwin together with the Dell Family Office as the largest AI-enabled take-private transaction to date; Baldwin has a strong management team, a foundational technology platform, and has already extensively adopted Anthropic.

  Unlike traditional private equity funds, Sequence does not aim to buy as many companies as possible, sell them after a short period, or primarily create value through financial structuring. The company plans to pursue no more than one transaction per year and takes a much longer time horizon for technological transformation. Lee sees traditional acquisition firms as being at a disadvantage particularly when it comes to recruiting world-class engineers, giving them influence over investment decisions, and dealing with the typical three-year pressure to sell, although he acknowledges their progress in AI transformation.

  Lee identifies exceptional management teams, deep industry expertise, an already evident openness to technology, and markets in which established companies possess structural advantages as the key criteria for new investments. His overarching lesson from investing and building companies is that exceptional people in large markets matter more than the supposedly best idea, because execution and persistence ultimately determine success.

  OpenAI, Anthropic, xAI, ChatGPT, and the Sequence platforms Atlas, Lattice, and Artifacts were explicitly discussed; format: deep dive.

NVIDIA Developer (1 new video)

• Vertically Integrated, Horizontally Open | AI Factory Insider Ep 5
  September 22, 2026, 21:43:02

  The video explains the concept of “vertically integrated, horizontally open”: NVIDIA optimizes the entire AI stack for maximum performance while opening every layer to hardware providers, middleware, orchestration, model, and library partners, as well as companies with their own data and workloads. This makes it possible to reuse common foundations such as data management, training, fine-tuning, and inference across many industries, while adapting data, models, and domain expertise to the specific use case. Examples mentioned include automated data processing in healthcare, AI-powered drug discovery, recommendation systems in retail, simulations in manufacturing, image reconstruction in medical devices, as well as fraud detection and portfolio optimization in finance. CUDA is described as the central link between applications and GPU hardware and as a multilayer platform ranging from frameworks to hardware controllers. CUDA libraries provide reusable, optimized building blocks, so developers do not have to implement everything themselves. At the same time, it is emphasized that “open” here means freely available, supporting open standards, and enabling others to build their own businesses on top of it—not necessarily open source or vendor neutrality. Close, collaborative development of hardware and software is intended to enable new capabilities and allow some improvements to benefit older GPU generations as well. In the context of agentic AI, CUDA libraries are to become available as “skills” for agents, allowing them to select suitable tools themselves and perform tasks such as data processing or performance optimization. According to the discussion, this shifts the role of developers from manual implementation toward guiding agents, applying deep domain knowledge, and designing systems holistically; AI and humans are intended to complement each other.

  NVIDIA, CUDA, CUDA-X libraries, and agentic AI were explicitly discussed; format: deep dive.

Odysseas

No new videos during this period.

OpenAI (2 new videos)

• One Prompt. A Feature Built and Tested. | Ramp × GPT-6 Astra
  21.9.2026, 19:22:03

  At Ramp, Astra is used as a kind of AI colleague who is given broad goals instead of detailed instructions and then works independently on tasks. As an example, the team cites a request to implement routing strategies based on API key granularity: Astra found the relevant place in the application, made the changes, worked on them for around twelve minutes, and then validated them for about fifteen minutes. In the process, it independently identified the correct API key and the appropriate settings, and even checked the changes in the mobile view, although this had not been explicitly requested. Particular emphasis is placed on its concise responses, screenshots of the changes, and ability to understand vague requirements. In the future, Astra is also expected to help automatically retrieve receipts for customers, with tests planned to determine how well it can be adapted to many different merchant websites.

  Astra and Ramp are explicitly discussed; format: demo.

• “A Staff Engineer Collaborator” | Notion’s First Look at GPT-6 Astra
  21.9.2026, 19:17:35

  Astra is described as a kind of “Staff Engineer Collaborator”: Rather than assigning small, clearly defined tasks, users should give the model an overarching goal and specify the desired state. Quinn reports that, compared with earlier models, Astra communicates much better, explains its results more clearly, and establishes relevant connections more quickly. As its first major task, Astra was asked to re-examine an existing check for token reuse and identify overlooked issues; according to Quinn, it actually found points that had previously been missed and described them precisely and without unnecessary elaboration. As a result, the developer spends less time deciphering the model’s follow-up questions and results and can continue working with it more quickly. For team use, Quinn therefore recommends specifying a central goal or metric rather than an individual task.

  Astra and Soul were explicitly discussed; format: opinion/reflection.

Pinecone

No new videos during this period.

Productive Dude (1 new video)

• ChatGPT Sites + Higgsfield: Build Beautiful ANIMATED Websites! (ChatGPT Sites Tutorial)
  22.9.2026, 20:05:21

  Using ChatGPT and the Higgsfield plugin, consistent reference images are first created for preparing a burger: from the burger on the grill, through flipping it and melting the cheese, to adding the toppings, packaging it, and adding the fries. Higgsfield combines these images into a coherent video sequence; Kling 3.0 Pro is used to generate the clips with start and end frames because, according to ChatGPT, Veo 3.1 does not support end frames. The generated clips are reviewed and, when errors occur—such as an incomplete flip or unnatural transitions—reworked using additional images and in-between frames. The final video is almost two minutes long, though it still contains minor texture issues on the table. The video is then used in ChatGPT Sites for a parallax website, where scrolling controls the video sequence and displays text in the empty areas of the images. After selecting font and color variations, the page is revised several times, including by making certain motion sections faster or slower, having text move into view instead of appearing abruptly, and removing unnecessary navigation elements. Finally, more creative text animations, larger descriptions, color correction, light reflections, and vignettes are added, resulting in an immersive burger website with heat effects, melting cheese, ingredient badges, and scroll-dependent animations. Explicitly discussed: ChatGPT, ChatGPT 6 Astra, ChatGPT Sites, Higgsfield, Kling 3.0 Pro, Veo 3.1, Claude, and HeyGen; format: tutorial.

Sebastien Dubois

No new videos during this period.

Simone Rizzo (1 new video)

• Jev è la nuova era dell’AI: l’ho rifatto in 1 giorno e te lo regalo
  26.9.2026, 10:23:26

  Jeev is described as a closed and paid “System-One” model: it is not a conventional chatbot, but processes a state and multiple questions in parallel, returning typed answers such as Boolean values, classes, or scores. The author therefore developed a local open-source alternative called Rizzo Flow, which, according to the demo, makes several decisions in about 175 milliseconds, also runs on the CPU, and can process up to one million input tokens. Technically, he explains three possible approaches — encoders, decoders, and diffusion models — and argues that Rizzo Flow is based on a decoder: autoregressive token generation stops after the first token, after which only the probabilities of relevant answer options are evaluated. For multiple questions, the KV cache of the unchanged state is reused, so only the respective questions need to be processed again and answers can be generated in parallel. Instead of Qwen, Rizzo Flow uses the Spark X 2.5 model in variants with 4 and 1.7 billion parameters, was fine-tuned with LoRA and QLoRA, and calibrated using distilled Jeev data; users can also provide their own datasets for specialized tasks. In the demo, the model classifies résumés, support tickets, and yes-or-no questions, assigns categories or scores, and, when it lacks knowledge, can also choose an “I don’t know” option; the API is intended to match Jeev’s, allowing existing code to continue to be used unchanged. The author cites browser automation, context cleanup in agents, ad analysis, lead qualification, and the classification of tickets, emails, and documents as potential use cases, particularly in combination with a larger agent model such as Claude. Finally, he addresses objections concerning the novelty, the use of encoder rather than decoder models, and the legality of distillation; his own assessment is that Jeev is not architecturally anything special, but is highly useful as a fast and practical classifier for business workflows.

  Jeev, Rizzo Flow, Semif, Laya, Open Jeev, Spark X 2.5, Qwen, Gemma, DeepSeek, Mistral, Claude, and Llama were explicitly discussed; format: Deep-Dive.

Supabase (1 new video)

• The future of AI…
  23.9.2026, 13:00:03

  Jev is described as a complement to generative language models, not a replacement for them. While a conventional LLM can be used like a hammer for many different tasks and generates responses token by token, Jev works like a screwdriver: It processes data in a single pass, selects from predefined answers, and additionally provides a confidence probability, such as “Spam 0.97.” This makes it particularly suitable for classification tasks such as spam detection, fraud checks, content moderation, or routing support tickets. Classifiers have been useful in the past, but they were costly to develop because they required labeled data, training, hosting, monitoring, and a new model for each new task. Jev is intended to simplify this process through a single API request, without requiring custom training, hosting, or a machine-learning team; what previously took an entire quarter could thereby become an afternoon’s task. According to the video, the biggest difference lies in the cost, since no output tokens are generated, making Jev 400 times cheaper. The central thesis is that developers have often used generative models for simple classification tasks and can now choose from several tools that are better suited to the job for the first time.

  Jev, LLMs, ChatGPT, and OpenAI are explicitly discussed; format: deep dive.

Tech With Tim (1 new video)

• LM Studio Full Tutorial – The BEST Way to Run Local Models
  26.9.2026, 11:50:21

  LM Studio allows you to download and run local models for free on your own computer and make them available to other applications via a local inference API. When choosing a model, the amount of available VRAM or Unified Memory is crucial; LM Studio filters models to match the hardware and recommends suitable variants. It explains that quantizations such as Q4, Q6, or Q8 affect memory requirements and accuracy, with lower quantization generally requiring less memory. After downloading, a model is loaded, and settings such as context length, cache quantization, and other options can be adjusted; multiple models can be loaded in parallel as long as sufficient memory is available. The integrated chat interface supports system prompts, parameter changes, structured output, plugins, file processing, and MCP servers. The tutorial then shows how to access the local server through an OpenAI-compatible endpoint from VS Code and the Kline extension; after an initially oversized context configuration, the model is successfully used with a reduced context. Finally, it demonstrates how a local model generates a Python script that sends a request to the LM Studio API itself.

  LM Studio, local models, Gemma, OpenAI-compatible endpoints, VS Code, Kline, and Claude Code were explicitly covered; format: tutorial.

TheAIGRID (1 new video)

• Google Just Took A Massive Step Towards AGI
  September 26, 2026, 00:00:07

  With the DeepMind Institute, Google DeepMind has created a permanent institution dedicated to building, measuring, regulating, and deploying AGI. Participants include Shane Le, James Manika, and Demis Hassabis; researchers from Google, DeepMind, and academia are expected to collaborate and publish their findings.

  The ten key points:

  1. Institution für AGI: The founding of the institute is neither a new Gemini model nor proof that AGI has already been achieved, but it is a clear institutional signal.
  2. AGI rückt laut DeepMind näher: The founders define AGI as a system with the cognitive abilities of the human brain and expect today’s gaps to be closed soon. They cite possible consequences such as communities of AI agents, institutional changes, scientific advances, and cyber, biological, and control risks.
  3. Zeithorizont 2028: Shane Le estimates the probability of minimal AGI by 2028 at 50 percent, while emphasizing that today’s chatbots have not yet reached this threshold.
  4. Messung durch ein kognitives Profil: A framework evaluates ten capabilities using new tasks that were not seen during training. This is intended to show where a model is strong, unreliable, or fundamentally deficient.
  5. Selbstverbessernde Agenten: Safety research should also examine systems that help design, program, test, or train better AI. This could accelerate improvement cycles while making human control more difficult.
  6. Überwachung von Denkprozessen: Chain-of-Thought data can provide clues about deception, awareness of testing, or misunderstandings. Future models, however, may work increasingly with numerical representations that are unreadable to humans, so their monitorability should be specifically assessed.
  7. Tests vor der Veröffentlichung: Hassabis proposes an independent body involving experts and government representatives, along with large-scale tests and evaluations whose content is not known in advance. Companies could initially share powerful models voluntarily for up to 30 days before launch; in the event of serious risks, a coordinated slowdown could also be considered.
  8. Auswirkungen auf Arbeit und Wirtschaft: One study compares scenarios ranging from limited task automation and widespread job displacement to an economy that continues to grow while demand for human cognitive work declines. Policy measures should be tied to measurable indicators such as wages, duration of unemployment, and labor’s share of income.
  9. Einfluss auf die Definition des Problems: Although the disclaimer says the contributions are merely prompts for discussion and do not represent an official Google position, leaders associated with DeepMind and Google could influence how governments and other labs assess AGI, benchmarks, and safety measures.
  10. Entscheidende Signale für die Zukunft: It will be important to see whether DeepMind regularly publishes cognitive profiles for Gemini, keeps thought processes monitorable despite more capable and opaque methods, introduces independent tests for agentic systems, and translates the institute’s proposals into concrete Google decisions.

  The founding of the institute therefore does not prove that AGI has been achieved; rather, it shows that DeepMind considers the arrival of human-like systems close enough to institutionally prepare not only for their development, but also for their measurement, release, safety, and societal consequences. Google DeepMind, Gemini, and AGI were discussed; format: deep dive.

Theo – t3․gg (1 new video)

• Getting the most out of Opus 5.5
  9/25/2026, 00:11:51

  Opus 5.5 is supposed to work particularly well when given a complete task with a clearly defined end state: for example, implementing a function, getting the tests to pass, and then delivering a screenshot or a PR. Instead of writing “think carefully,” users should let the model decide for itself how much reasoning it needs; the “Max” mode in particular is criticized because, according to the experiences shown, it consumes significantly more tokens and time while delivering hardly better results. For long tasks, the speaker recommends adjusting the process while it is running instead of stopping and restarting the thread, and explicitly specifying when the model should continue and when it should ask questions. In one example, Opus moves a task to another computer, sets up the repository, environment variables, and access there, and receives clear permissions along with the instruction to ask questions when uncertain and not interfere with the local computer through browser or computer control. Large tasks can be divided into subtasks or Sub-Agents, and their use should be explicitly allowed when necessary.

  For design tasks, users should not only describe the desired styles but also name specific undesirable patterns; screenshots and marked-up image excerpts are said to be particularly helpful for this. After long runs, one should first check whether the model is waiting for a decision or approval, and then read its summary, changes, open issues, and aspects that could not be verified. The speaker also uses models from other families for code reviews and considers OpenAI models as well as other mentioned models valuable reviewers, because they sometimes find different errors than Opus. Overall, the central thesis is that Opus 5.5 needs more room to maneuver, clear validation criteria, and trust, while the human sets the boundaries for risky or destructive actions; model switching in response to safety flags and avoiding requests to disclose internal reasoning traces are also discussed. Anthropic, Opus 5.5, Claude Code, Claude Apps, Browserbase, T3 Code, OpenAI, Fable 5.1, Grock 47, GB6 Soul, and GPT6 Astra were explicitly discussed; format: Deep-Dive.

Tim Carambat

No new videos during this period.

Tina Huang

No new videos during this period.

Unsupervised Learning

No new videos during this period.

Weights & Biases

No new videos during this period.

WorldofAI (1 new video)

• HUGE OpenAI DevDay LEAK! “o” AI Agent, Sonnet 5.5 BEATS GPT-6, MiniMax M3.1 OUT & More! AI NEWS
  Sep 27, 2026, 06:17:31

  Ahead of OpenAI DevDay on September 29, leaks about a product called “O” or “AON” are being discussed: a potentially always-on assistant for long-running tasks that could work for hours, days, or weeks, execute code, conduct research, and possibly coordinate multiple agents. There are also indications of three API speed tiers—Standard, Fast, and Ultra Fast—as well as a possible expansion of particularly fast access. Anthropic is reportedly planning to release Sonnet 5.5 in the coming weeks; testing reports and the speaker’s personal assessment portray the model as very fast, efficient, and in some respects more capable than “GPT-6 Soul,” although these claims and the presumed pricing have not been confirmed. MiniMax M3.1 Flash was initially leaked through a removed GitHub entry and subsequently officially released as a fast model for everyday development tasks; examples show solid results with graphics and SVGs, even though it reportedly does not match GPT-6’s level. It is also mentioned that LongCat 2.5 Preview is supposed to be available free of charge for two weeks through Open Code, with a one-million-token context window, multimodality, and no data retention. Another example presented is Opus 5.5, which reportedly created an extensive Minecraft clone almost entirely autonomously, including tens of thousands of lines of code and tests, hundreds of files, and generated images and sounds. A sponsorship segment promotes the Firecrawl Developer Index, which searches developer documentation, issues, and pull requests and can connect to Cursor, Codex, and MCP clients, among others.

  OpenAI, Anthropic, GPT models, O/AON, Astra, Sonnet 5.5, Opus 5.5, MiniMax M3.1 Flash, LongCat 2.5, Firecrawl, Cursor, Codex, and MCP were explicitly discussed; format: roundup.

Zubair Trabzada | AI Workshop (6 new videos)

• I Combined Claude Opus 5.5 With Jev – Here’s What Happened
  26.9.2026, 19:17:23

  The video compares Claude Opus 5.5 on its own with a combination of Jev and Opus 5.5 when handling customer inquiries. Jev takes care of quick classification decisions, such as whether a payment issue is urgent, requires follow-up, or should be assigned priority one; in one example, this took 293 milliseconds compared with 1.85 seconds using Opus alone, without calling Opus at all. For twelve different messages, the combination processed the test in 19.7 seconds with four Opus calls, while Opus alone required 33.9 seconds and twelve calls. As soon as an inquiry requires an actual draft response, such as a pricing request for a small website, Jev forwards it to Opus; this keeps the results similar while making the preliminary decision faster. The video then explains how to build such a system for customer service or other business processes using Claude Code, MCP connections, and Zapier, with Jev integrated through a corresponding connection. It also shows how to connect a Jev account using an API key, as well as an example with the assistant Jarvis, which first has Jev classify inquiries and then uses Opus 5.5 for more complex tasks. In the later section, Opus 3.5 is explicitly recommended for Claude Code; a free prompt collection and a paid version of Jarvis are also mentioned.

  Claude Opus 5.5, Opus 3.5, Jev, Claude Code, MCP, Zapier, Type Safe Jev, Zendesk, and Jarvis were explicitly discussed; format: tutorial.

• My AI Called Me During a Meeting (Real Recording)
  24.9.2026, 20:07:46

  Jarvis is an AI assistant with access to files, folders, and various tools. In the demo, it recognizes an urgent business matter, calls the user during a meeting, and reports that Dana Whitfield has signed a $3,400 contract; it then checks the calendar, adds a kickoff appointment, and sends a confirmation along with a voice message via Telegram. Jarvis can also retrieve information from the personal “AI second brain,” see and take over the screen, and carry out actions in an advertising campaign until the user stops it. Through the camera, it can observe the user and, for example, provide feedback on their clothing. A reflection module called “Jev” is also demonstrated. It quickly classifies inquiries and decides whether they relate to files or require an action or tool. The creator presents Claude Opus 5.5 as the preferred model for Jarvis, but also mentions switching to other models and offers a free prompt package for recreating it, as well as a ready-made version in a paid community.

  Claude Opus 5.5, Claude Fable 5, Sonnet, Jev, Telegram, Google Calendar, Gmail, GitHub, Notion, Google Sheets, Google Drive, Zapier, and Canva were explicitly discussed; format: demo.

• Claude Opus 5.5 Is a Game Changer for Small Business
  23.9.2026, 23:33:29

  Claude Opus 5.5 is presented as particularly well suited to recurring tasks in small businesses, such as following up on invoices, contacting leads, and answering frequently received emails. The video demonstrates a combination of Opus 5.5, Claude Code or the Claude desktop app, and the “Claude for small businesses” plugin, which contains prebuilt workflows for areas including inbox management, SEO, invoicing, and lead management. Installation takes place through the plugin section, where users search for “small business” and activate the plugin from Anthropic. For additional services, Zapier is set up as a connector, allowing applications such as Gmail, Google Sheets, and Slack to be connected and granted the necessary permissions. The concrete demo creates a workflow that searches Gmail for messages from real people, skips newsletters and receipts, saves replies in the user’s personal writing style as drafts, logs the inquiries in a Google Sheet, and then sends a summary to Slack. In the demonstrated run, five emails were found, three of them relevant; drafts, a CRM-like spreadsheet log, and a Slack summary were created for these messages, with the emails still requiring review before being sent. Other workflows, including weekly briefings, marketing, lead generation, job postings, payment reminders, and inventory planning, are also mentioned; the speaker emphasizes that the system can infer the user’s way of working from previous interactions.

  Claude Opus 5.5, Claude, Claude Code, Claude for small businesses, Anthropic, Zapier, Gmail, Google Sheets, Slack, Google Calendar, Zoho CRM, and the assistant “Jarvis” were explicitly discussed; format: tutorial.

• I Upgraded JARVIS With Jev (INSANE Results!)
  22.9.2026, 19:51:01

  Jev is demonstrated as a preliminary decision layer for the Jarvis AI assistant. Before every response, Jev evaluates an inquiry based, among other things, on which area it belongs to, whether it is complete, whether Jarvis is addressed directly, whether an action involving greater risks is required, and whether the information is available in the “Vault.” For a request involving social media notes, Jev assigns the task to the Vault, after which Jarvis retrieves the relevant files and folders. For email-related tasks, Jev instead recognizes that an action is required, classifies the request as a mail/calendar task, and asks for confirmation; in the example, Jarvis also points out that identical messages have already been sent. A vague request to create an invoice for John Doe and send it via Telegram likewise results in a follow-up question because the recipient and invoice details are missing and the request probably cannot be answered from the Vault. According to the video, the benefit is that this fast routing means Jarvis has to check fewer functions and tools itself, allowing it to work more efficiently, quickly, and cost-effectively. A free PDF guide for implementing the integration is also announced, while a ready-made version of Jarvis is offered in a paid community. Jev, Jarvis, and the Type-Safe API were discussed; the video is a demo.

• I Let GPT-6 Astra Run My Meta Ads – Here’s What Happened
  22.9.2026, 00:36:30

  GPT-6 Astra is intended to create Meta ads largely automatically: after being given a product idea, it researches active competitor ads through the Meta Ad Library, analyzes their hooks and messaging, and uses these insights to create new ad creatives. In the example, a prompt for the product “PacNest,” a compression packing cube, is entered into Codex; the system then automatically searches for relevant ads and uses Hicksfield to create several photorealistic static ads. According to the display, the process took 9 minutes and 29 seconds, while the researched pages and ads used remain traceable in Codex. The video also shows how Hicksfield is connected through a plugin and authorized once. In addition to static images, video and UGC ads can also be generated, for example for the product “Nook Beam Mini”; Cance 2.5, Nana Banana Pro, and GPT image 2.5 are mentioned in connection with Hicksfield. Codex can also incorporate existing campaign data, previously successful ads, and the user’s own notes, as well as create ad campaigns in the browser; the budget and the decision between cheaper static ads and more expensive videos remain with the user. A PDF containing prompts and research-based concepts is offered as free assistance.

  GPT-6 Astra, Codex, Hicksfield, the Meta Ad Library, Cance 2.5, Nana Banana Pro, GPT image 2.5, Jarvis, and Jev are explicitly discussed; format: tutorial.

• How to Use JEV With Claude Code and GPT-6 Astra
  21.9.2026, 17:56:18

  JEV is explained as a decision model rather than a language model: it does not write text, but evaluates predefined situations and provides probabilities, yes/no decisions, multiple-choice selections, or ratings. According to the video, this makes it particularly fast and inexpensive; in the example shown, it responds to traffic-control situations in some cases in under 100 milliseconds. A city simulation built with GPT-6 Astra uses JEV to control traffic at intersections and respond immediately to changes such as blocked roads. JEV is not intended to build apps itself, but to serve as a decision layer within applications created with Claude Code or GPT-6 Astra. To use it, the video describes creating an account with TypeSafe.ai, obtaining an API key, and then integrating the key into the application; Vercel and OpenRouter are mentioned as alternatives. As another application, the video shows an AI assistant, or “AI second brain,” in which JEV quickly decides which notes or files match a request. The central message is that many apps supposedly built with JEV were actually created with other models, while JEV should be used solely for fast decision-making.

  JEV, Claude Code, GPT-6 Astra, ChatGPT, Claude, Opus 5, TypeSafe.ai, Vercel, and OpenRouter are discussed; format: tutorial.

---

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