AI Weekly: Autonomous Agent Workflows, Direct Weight Editing, and Developer Tooling

Artificial Intelligence, weekly news every Friday at 2:00 PM, Eastern Time, New York time. Today is Friday, November 6th. As you see a lot of updates. And, models edit models, literally, editing the weights of the models. Agents fix agents, satellites run Gemma, and producing poetry, haiku. Models run each other. So, there is a lot of collaboration is going on. And, these are a couple pictures, I have another slide, explaining them. So, but definitely transition from micromanaging, from writing code yourself to micromanaging the agent, to actually just directing the agent and let it go, let it do its stuff. So, the new way, you just provide directions, intent, and guardrails, and the agent does everything else. Okay. Leaderboard has not changed, last time it was changed was October 27th. So, this hasn't changed. Okay. Artificial analysis, intelligent index, so cost of intelligence. Uh so, look, GPT-6 Luna is very, very inexpensive. This is Claude Haiku. Uh this is Terra, GPT-6.1 Terra. GPT-6.1 Soul, which is really good model, right? Uh Quen 3-8 27B. This is a small model. This other models are usually very, very big, but this is a small model, only 27B. You can you can run it on your laptop and pretty good. Uh GPT-6, Astra, and this is Gemini 4 Carbon, which is top of the leaderboards, currently. Okay. AI news. DeepSeek, Router-Worker, Agent. So, DeepSeek has a mo, oops, what's going on? DeepSeek has a model. DeepSeek has, um, uh, what is what is it called? Oh gosh, Harness, right? DSH, DeepSeek Harness. And now, it is agent. So, it's open sourced, MIT. And, it's prototype. But, runs locally using this DeepSeek Harness, and it's very fast and good. And, monitors one GitHub repository and this Feishu, which is similar to Slack, it's the Chinese collaboration uh, platform, turning authorized request in parallel tasks. So, good. Uh, next. GPT-6, Astra can edit Quen's open weights directly. So, this is what I was talking about. So, we started with Quen model, which is a small model. You see 0.8 billion parameter model. So, it's a small model. And, Astra starts changing it, editing it. And, it end up changing it, improving it. So, the results of this model went from like about 8% to 48% on certain math tests. So, it's like a direct brain surgery, very interesting. Uh, Anthropic Public Benefit Fund. So, what Anthropic, so you remember when uh, Anthropic was formed by eight people, who left OpenAI, and they have same share of the company, and together, they control a little bit more than half of the company. They all made a personal pledge. The 80% equity. And also, they made pledge, uh not actually pledge, it's a reality, 5% of company shares goes into this fund. So, the goal share AI generated wealth and help workers facing disruption, through charitable grants, direct payments, uh interestingly that Elon Musk has a similar intent. So, he promised to give SpaceX share per American child, next year. Interesting, but this is huge amount of money. So, the 5% which they are allocating right now at 2 trillion uh company value, it's about 100 billion fund. And with all the personal, it it will like 400 billion, and this is at today's value, but it grows very fast. So, this is amazing. Uh, next. Uh, weekly videos uh, every Friday. Oh, this is just a pitch, uh for my channel, so name of the channel, Lev Selector. Uh, I provide slides which you can download. This slides or previous, like all, every week a lot of slides. Uh, and also, I really ask for feedback. If you can pause the video and ask questions or provide some feedback. Uh, next. Yeah, this this is the the images. They were generated by GPT. Uh, I find that GPT really understands me well and create nice images. So, the old way was, uh, micromanaging the agent. So, everybody is writing code using an agent. I don't know anybody who is writing code by hand. And, now is really, you can see transition where you going from, uh, micromanaging the agent, uh, making small steps to actually let agent do its thing, you just provide direction, intent, and guardrails. I want also to tell about this. So, coding with the agent sometimes feel like radar dealing with army paperwork. If you remember, MASH, it's a mobile Army Surgical Hospital, it's a famous show on American television. The last episode, when it was aired, more than 100 million people were watching it at the same time. Anyway, there is a scene there where Colonel, uh Blake and Radar talk and Radar was like his assistant, uh doing administrative work and asking him to sign some papers. So, Radar, tell me the truth, do you understand any of this? Oh, I try not to, sir. It slows up the work. So, approximately the same happens when uh, for example, I work, I tell what I want to get and uh AI starts working on it and it goes through so many tests and uh trial and error and eventually it it gets it how to do it. And frankly, I don't even understand how it did it, but I'm happy that it works now. And so, I'm kind of like, Radar, try not to understand because it slows up the work. Anyway, uh, LangChain, Self-Improving Agent. Agents. Uh, so there was, this uh, meeting. Harrison Chase is the creator of LangChain, cofounder. So, this was in London, and he was talking about, uh, improving agents, self-improving agents. And how they approach it, and how they achieve it. And agent is a complex structure, so we have a reasoning model, harness, context containing, instructions, knowledge, memory, all these pieces, and how you set the system so it improves itself. It's a very nice presentation, it's on YouTube, I recommend to listen to it. Okay. Linus Torvalds embraces AI tools. So, Linus Torvalds is a famous creator of Linux and leading the Linux kernel project, uh for many, many years. So, he has shifted from criticizing AI, in 24 to now using it, uh, using it effectively, for personal projects, for Linux kernel development and uh you see, reported benchmark gains reached about three times on a laptop and 23 on a server, not across all kernel opera, but anyway, so this, uh becoming a reality, which is good. Um, Occupy, software as a service, AI against, uh, SaaS fees. So, what happened is, this Occupy project, you see this is their website, uh occupysass.com and GitHub. So, what they're doing, they take some, famous, uh well-known, software as a service, paid services, uh and they create a completely open source variant which is doing the same thing and they creating them of course, using AI. So, for example, for Auth0 authentication and Salesforce customer management, they created ZeroAuthy and SalesCompel. It's probably not, uh quite the same, production wise, but uh, projects tend to become better with time. So, this is, yeah, Occupy SaaS. AI news. Karpathy's idea to get short clear AI output. Well, there are many projects, GitHub repositories which provide skills, or rules, how to reduce the number of tokens. So, they instruct the model how to use less tokens and they work. Now, in this particular presentation, uh, Karpathy, uh, showed yet another way how you can achieve it. So, apparently, there is a standard in Europe, in in Europe's aerospace technical writing standard, since 86. So, it's like 40 years ago, it was created, a long time ago. You know, when you're looking at any governmental or any, like bigger organization, documentation, it's usually very bureaucratic and impossible to read and understand, right? So, this standard is different because it, uh instructs how to write so it is concise and easy to understand. And what you can do, you can just take this, uh, uh, set of rules and you just give it to your agent and the agent immediately becomes better at writing in a way so that it doesn't spend as much tokens and it is clear, like its language it become like more clarity. So, very interesting. So, this is the website, ASD-STE100.org, where you can download and you can adapt for yourself if you want to do it. Impeccable, a design skill. Keeps AI-built UIs from looking generic. For Claude Code, Cursor, Codex CLI and Gemini CLI built on Anthropic's frontend-design skill. Seven reference files on typography, color, motion and more. 23 commands under /impeccable, such as audit, polish and critique. Anti-pattern rules catch template tells like purple gradients and nested cards. A CLI runs the checks with no LLM or API key; 10,632 stars in one week. Vercel CI; continuous integration. In beta, runs only the tasks affected by a code change. Spins up only the compute those checks need. At Vercel, time from git push to a passing CI fell 2-6x. Vercel pitches it as CI that can keep up with coding agents. Stop using Coding Interviews. Anthropic engineering manager, Felix Rieseberg announced that he stopped using traditional coding interviews for his candidates. He argues Claude outperforms applicable on these exercises. So results of coding interviews are poor predictors of meaningful engineering contributions. He advises junior engineers to understand business goals, identify ambitious opportunities, and build focused prototypes demonstrating impact. He now examines portfolios for taste, empathy, and principled thinking. Trajectory. This is a very interesting company, and they are raising some money, and I listened to the whole interview, I really, really like it. So, the idea is that, uh, you can, in one application, combine model training, evaluation, and deployment. So, their idea is that every company and even every individual will be able, not just use a model, but tune it, and train it, and do it, very, very easily. And, create like a perfect uh, secretary or like whatever, uh, goal of having the agent. So, example, they used it for Harvey, this is a company doing legal work, legal assistance, AI driven. A model accuracy went from 4% to 18% while costing six times less. So, this is idea where you use, uh, you can run the model locally, right? And, you can tune it, and you do it continuously. And, how you create a user interface so it's very easy to use. Claude API message threads, send only the new turn, not the whole history. A typical agent request drops from about 800 KB to 1.8 KB. The server keeps the conversation between turns. Send new messages plus the previous response ID. In public beta on the Messages API. Uses the same weights as GPT-6 Astra. Available as GPT-6 Rosalind-6 research to vetted scientists, institutes, companies and governments. Built to refuse requests for biological weapons. Rated High Capability in biology and chemistry under OpenAI's testing framework. Google runs Gemma 4 on a TPU in orbit for the first time. Project Sunchatcher prototype ran Gemma 4 on a data-center-class TPU aboard a satellite. It generated a haiku poem, sent down through Svalbard. The chip can run for about 20 minutes at a time for now. A pair of linked satellites is planned for 2027, toward solar-powered space response. Microsoft rewrote Copilot Cowork from Python to Rust while live. Global availability in 4.5 months. The team switched its runtime from Python to Rust while live. Practice 1: write clear specs before any code. Practice 2: run narrow agents with clear boundaries. Practice 3: make sure automated checks really verify each change. The team stresses clear tests, boundaries and human oversight. FlashProve: Provable LLM Inference. Open-source research system for provable inference, linked to DKIM email signing. Labs publish a fingerprint per model; outputs can be checked against it. Catches a provider quietly swapping in a cheaper model. Built to be light enough to verify on a local device. Specs and partial code on GitHub; no plan yet to deploy it at any lab. Vals: low reasoning effort keeps most of the score at a fraction of the cost. Seven frontier models on eight Vals index benchmarks. Claude Opus 5.5 at low reasoning kept 80% of its score at 16% of the cost. In 22 of 56 model-benchmark pairs, a lower setting matched the top one. Effort matters most on legal research (+19.2 points) and code migration (+15.6). A mixed Opus 5.5 and GPT-6 team did no better than either alone. Apple opens Siri AI to all app developers. New "skills" teach Siri multi-step jobs inside an app. Siri AI can see the screen and read messages. Developers can test only with internal testers; App Store submissions start in Spring. The information is available to the EU and to users under 13. Ripple-65B: Open Coding Preview. Recursive Intelligence (Palo Alto, CA) released Ripple-65B Preview. Available through Hugging Face, the Apache 2.0 software offers 65B-Bin params dense. 1 Million tokens context window. Development costs were $32 Min. Self-reported 80.5 on Terminal-Bench 2.1 narrowly reflects Reflection Beam 80.1, indicating competitive autonomous coding with fewer total parameters. Nothing's Screen-Free AI Debut. Nothing tech is based in London, UK. Announced that its first standalone AI device. The screen-free tool "couldn't have been an app," Pei says. Problem: AI gadgets have yet to demonstrate essential everyday value. Pei's proposed solution is proving usefulness at work before pursuing widespread home adoption. Importance: This marks Nothing's move into dedicated AI hardware, prioritizing simplicity rather than another display. Pei places AI's "iPhone moment" years away; rumored Nothing smart glasses remain unconfirmed. OpenAI launches Images 2.6 with GIFs and image reference. Animated GIFs, saved references, 50% faster. Rolling out globally in ChatGPT and Codex; generation is 50% faster. Makes animated GIFs from a prompt or photo; edit frames and set the speed. Keeps characters consistent across frames. Save a character, pet or product once and reuse it as a reference and gpt-image-2.6-flare (as fast as 6). and gpt-image-2.6-sunburst (higher quality). Both support image markup and SVG export. Black Forest Labs FLUX 3 Image Klein: A 5.6B open model for editing. Open weights, 5.6B parameters, built for precise image editing. Native 4K output. Also does text-to-image. Bounding-box control to say where an edit goes. Pixel-exact editing of the parts you choose; weights on Hugging Face. Axiom Formalizes Stacks in Lean. Axiom Math, based in Palo Alto, CA, announced its complete Lean. Initiated by Columbia mathematician Johan de Jong in 2005, Stacks is an open, approximately 7.5K-page reference for modern algebraic geometry. Axiom reports 16 Min code lines, calling it the largest integrated, reusable formal mathematics library ever. Pangram Image 1.0 detects AI in Image. Detects AI-generated images and video. Draws a heatmap that locates exactly where AI appears in an image. Robust to edits, compression and low resolution. It can detect modified images while browsing. API for both images and video. Together AI Compiler Speeds Inference. Neocloud - Together AI (Since 2022, San Francisco) provides cloud infrastructure and APIs for running, training, and fine-tuning, and open-source AI models. AI coding agents continuously optimize model inference by rewriting GPU code for specific models, hardware, and traffic patterns. Parallel agents optimize performance; persistent engines test speed and correctness without repeated restarts. Achievement: Together reports up to 2.6x higher token throughput than leading open inference configurations. Optimization completed within 24 hours. 252x greater experimenter throughput for testing candidate code. Faster adaptation can improve GPU utilization and reduce serving costs; reported gains vary by workload. Thinking Machines Lab - Inkling 2 Small. Thinking Machines Lab is based in San Francisco, CA - Mira Murati. Thinking Machines Lab released a new 594B total / 178 active. Accepts text and images; adjustable reasoning effort. Natively in NVFP4 on NVIDIA GB300 racks. Self-reported 80.5 on Terminal-Bench 2.1 Pro 81.6%. Coding, visual reasoning, and knowledge-work tasks. Sampling and fine-tuning are available now through Tinker, Open source weights should come soon under Apache 2.0. Cloudflare's AI Agent Trust Plan. Cloudflare explains that websites must identify the users behind AI agents, to prevent fraud, identify scraping. The new Trusted Agent Delegation Coalition will bring problems to the IETF for interoperable standards. The approach builds on Web Auth, Privacy Pass, x402 payments, fraud, identity spoofing, and scraping. It proposes four identity pillars: proof of account, and unique, balancing privacy with verification requirement. The approach builds on Web Auth, Privacy Pass, x402 payments. Customers can now distinguish personal-assistant-traffic, improving visibility and enabling targeted controls without blocking legitimate agents. Google Gemini Omni 1.2 Flash. 15-second video clips (up from 10 in Omni 1.1). 30% cheaper ($0.07 per second of 720p video). More efficient, different resolutions. Can use short motion clips to guide the movement. You can edit a full clip several turns without regenerating it. A grid view in the Gemini API as Gemini-Omni-1.2-Flash. Ranking models on Games Where they must lie to win. Dwarkesh Patel ranks models on the social deception game "Blood on the Clocktower". Cost of ranking 386 games was $20K in API credits. GPT-6 Astra leads, about 430 Elo ahead of Claude Opus 4.6. Open AI models win mostly as the good team, weighting evidence across days. Claude models tend to take reported claims at face value. A mixed Opus 5.5 and GPT-6 team did no better than either alone. Jobs and Layoffs. The Tech Layoff Tracker. In 2026, 204,867 people laid off (661 per day). In 2025, 245,953 people laid off (674 per day). In 2024, 238,461 people laid off (653 per day). Tech Layoffs by year (US only). 142.0K in 2026 (as of Nov 5, 2026). 123K in 2025. 153K in 2024. 266K in 2023. 165K in 2022. About the Speaker, Lev Selector, Ph.D. 40+ years of software engineering, data science, and building teams (hiring, training, and managing). Ph.D. in mathematical modeling and computer simulations. Interests: Generative AI, using LLMs with your data, local AI for local private data, cloud architecture, fin-tech, application security. Find and connect: Linkedin, GitHub, YouTube, Google. Thank You! Videos of every seminar are on YouTube. Slides on GitHub. Click a pptx file, then the download button on the right. Slides on Google Drive.