|
Hello Full Stackers! It's been six weeks! I hope you had a nice summer. I was in Latvia, Estonia, and the Netherlands (where I gave my first keynote at ProductCamp Amsterdam), then heads down building my newest course. We've got 1,655 new subscribers since the last send. Welcome to the stack! 🥞 Good thing not much happened over the last six weeks... just four frontier models in 72 hours, an whole new AI product category, and Cowork disappearing as a product entirely. Lots to cover! Here's the plan: | • | Applications are open for my new course | | • | Personal agents have arrived: Grok Bot, Muse, Instinct | | • | Astra and Fable 5.1, tips & my impressions | | • | The sizzle reel 🥓 (all the other stuff) | | • | Memes of course |
Let's do this. 🍳 Fresh Off the Griddle (My stuff) |
Build Anything with AI: applications are open At long last, I'm launching my AI Building course! It's called Build Anything with AI: Idea to Real App, and it's my first-ever live cohort: four weeks in November, one module a week, in one big group, with weekly office hours and a Slack. You learn how to come up with a buildable idea and leave with your own app live on the internet. Like everything I make, you learn interactively IN the tools where you do the building like Claude Code, Codex or Cursor. Then we come together weekly live to review the lessons and answer questions. This isn't the world's first AI building course, but here's what makes it totally unique: | • | It connects building to everything else you do with agents. Things like managing context and checking an agent's work are the same skills that make you good at knowledge work with AI. Being able to create software is insane leverage for almost anything you do, and you're already mastering the tool to help you do it. | | • | It teaches the fundamentals from the ground up, assuming agents the whole way. You learn what software is made of, at the level you need for what you want to build. It's the real code-level throughout, nothing hidden behind no-code tools like Lovable. | | • | You learn in the tool, so you learn by DOING, not reading or watching. |
Takes 1 minute. I read them all! |
Full Stack Mastery The new course is one piece of Full Stack Mastery, the crown jewel of The Full Stack PM. It's one membership with everything in it: all my courses, taught inside Claude Code, Codex or Cursor, and every course I make next. | • | 9 modules of hands-on lessons across research, docs, data, building, and your own PM operating system, all done in your tool on realistic PM work | | • | The Slack, 1,000+ product people building with AI, where I'm in daily | | • | Weekly live office hours with me, bring your work and we fix it together | | • | Every future course included, starting with Build Anything with AI in beta in early October |
It's $300 a year. Compared to the $2,000+ single-course programs out there, many of my students tell me: it's an insane deal. Learn more Many members expense this. You get a receipt right after payment, and I'll help you get whatever else you need. |
🤖 The personal agent has arrived |
OpenClaw had a HUGE run early this year. The concept was awesome: an always-on agent with its own computer that you could configure to your heart's content. But it didn't catch on with general consumers in a big way because it was too unreliable and TOO configurable. I personally made learnopenclaw.com but put that project on pause when my own OpenClaw just kept breaking. Grok Bot (from Cursor/SpaceX) is OpenClaw that just works. It shipped Aug 11 and I honestly love it. It's one of the most delightful AI experiences I've had since trying Claude Code for the first time. | • | Named, persistent bots, each with a role and its own memory | | • | A cloud computer with a real browser, signed into your apps, working after you close your laptop | | • | Routines on a schedule or on any kind of trigger | | • | Bots in one thread handing work to each other | | • | A phone app with a live view of the bot's screen and a take-over button for passwords |
None of that is a new capability, but it's all assembled reliably. Other notes: No model picker, each bot is one unbounded thread, $20 tier for anyone with Cursor Pro or SuperGrok. Now the others are showing up with the same product: | • | Muse (Meta, Sep 8). A consumer personal agent, free in the US, in an app and inside WhatsApp. Each user gets their own cloud machine, a second agent approves anything that touches the internet, and it can make payments with one-time Stripe cards. | | • | Instinct (invite-only). No app at all! You text it over iMessage or WhatsApp and it drives a cloud computer and your accounts. Reportedly raising about $1B at a $10B valuation (more than Snapchat, lol). After some digging, my friend discovered it literally IS OpenClaw under the hood. |
Anthropic and OpenAI have to be next. OpenAI has all the primitives out now (the Agents API, Computer History on the Mac, the Codex harness as a service), and Tibo, who runs Codex, has been hinting at it. ✍️ Tales from the frontier: Astra and Fable 5.1 |
Four frontier models shipped in 72 hours at the start of September. GPT-6 Astra (OpenAI, Sep 3). It's not "AGI" but it's definitely a huge step up for code. It's almost unbelievably good at 3D graphics and computer use. It can drive a browser and a desktop well enough that you can connect your AI to anything, even things with no API and no connector. It's been a godsend for helping me navigate Japan's notoriously awful booking sites. Astra researches the options and then actually goes and checks the pages, navigating all the way through to the real availability. If I could let it pay with my credit card I absolutely would. Claude Fable 5.1 (Anthropic, Sep 1). Also great, and my current favorite model. I don't actually have that much to say about it other than that it's extremely capable, has maybe 90% of its Claudeisms removed (but still has some load-bearing ones 😉), and is expensive to use. It's quite stressful seeing the usage percent move up on nearly every request, but it's making me more strategic in how I orchestrate the work between models. Still waiting on Grok 4.7, which Musk has now delayed twice and started calling 4.8. And Google has shipped three Flash models in six weeks lol... Gemini 4 Pro coming soon? With these models, one-shotting whole apps from a good description is normal now, so you really just need to know what to ask for and how to ask for it. That's a big reason I decided that now is the time for my Builder course btw, it's simply an amazing time to build. Apply here 😉 🤖 Anthropic | • | | | • | Claude Code weekly limits dropped about 17% on Sep 14, as scheduled. As I mentioned, I'm already feeling it. I'm on the $200 Max plan, I like using Fable, and I can literally watch the percentage go up with almost every prompt. | | • | Claude projects got a redesign. You give a project a goal and a repo, a coordinator chat plans the work, and worker threads (each one a cloud Claude Code session on its own branch) open the PRs. It's in beta in Claude Code first. | | • | |
🔮 OpenAI | • | ChatGPT Images 2.5 shipped on Sep 8. It's sharper, it keeps the subjects from your reference photos, and you can comment on one part of an image to edit only that part. Everyone gets it, on every tier. | | • | GPT-Live-1 is generally available in the API. It handles the full-duplex voice and hands the actual thinking to a backend model or to Codex. | | • | Codex now asks you questions inline while it keeps working, and Astra is in the model picker. |
✨ Google 🎯 Cursor | • | Cursor shipped Projects, and its cloud agents can now subscribe to a PR or a Slack thread and drive it to green. Auto mode is priced per model now, usually higher than before. |
🌍 Open weights and builders | • | DeepSeek V4.1-Flash is out, open-source, and multimodal. K2 Horizon shipped six sizes under Apache. Atria Dawn, a 744B model, appeared as weights only with no blog post and no API. | | • | |
🧪 Jev, the most interesting thing you can't use yet | • | TypeSafe AI came out of stealth on Sep 15 with Jev, a model that never generates text. You send it a blob of state and typed questions, and it returns answers with calibrated confidence. It's aimed at every classify, route, score and gate call inside an AI feature. It's waitlist only, and there are no benchmarks yet, but I think we're about to see another insane AI jump in surprising ways because of it. |
Here were a few of my favorites over the last few weeks: "subagents being released into my codebase" New here, or been lurking? All my free courses are on-ramps to building with AI: Just paste this into whatever tool you want to learn, and the AI will help you get set up: Install the fullstackpm CLI via npm and onboard me, check fullstackpm.com/cli for details if needed
And if you're ready for the whole stack, Full Stack Mastery is $300/year, with the new course in beta for members next month and a seat in the November cohort included. Or apply for the cohort on its own first; it takes a minute. Keep building,
Carl 🥞
|
