Welcome, humans.

Okay, so Meta Muse is basically “OpenClaw for normies”: download the app or open muse.ai, connect only what you want it touching, then ask it to manage email, plan trips, watch prices, or buy stuff. It even keeps working after you close it, though mail and purchases still require your approval.

Zuckerberg told Alex Heath he uses Muse to grab climbing permits when they open, plan weekly baking projects with his daughter, and review MMA footage. Meta’s bet is that Muse can make or save enough money to stay free for most people, then monetize a small cut of some transactions. So basically the free robot assistant today may someday work on commission.

And Meta is giving you a lot of room to play with it: free starts around 100M tokens a week, with $20 Power and $100 Maximum plans if you need more.

One thing I’d check before feeding it your life: if Meta lets you opt out of using your Muse interactions to train its AI models. Turns out, that isn’t a free-tier-only rule. Meta says its stronger “Confidential VM” (VM = virtual machine, the cloud computer where Muse works), where even Meta won’t be able to access the contents, is coming later this year. So… it’s not quite confidential just yet…

Here’s what happened in AI today:

• 😺 Anthropic researcher Jacob Coxon quit over self-improving AI risk.

• 📰 OpenAI says 10,000 agents tackled Navier-Stokes.

• 📰 Apple launched iPhone 18 Pro with Siri AI.

• 📰 Inception launched Mercury 2.5 at 1,100+ tokens/sec.

• 🧩 Thursday Trivia: guess which image is AI.

…and a whole lot more in today’s Around the Horn.

🤖 LATER TODAY: Learn GitHub for AI coding from absolute zero

GitHub is basically Google Drive or iCloud for your code, except it also remembers every change and gives you the tools to turn an AI-built project into something you can actually ship.

Today at 10 AM PT / 12 PM CT / 1 PM ET, Cassidy Williams from GitHub is starting from zero with us. We’ll cover repos, branches, pushing and merging, worktrees, GitHub Actions, key management, and how to safely store and ship what your agent built.

If you’ve heard us talk about GitHub, but the site has always looked like it was designed to scare you away, this one is for you. You don’t need to be a coder. You just need to want your AI-made projects to survive after the chat closes.

Set a reminder and join us LIVE with Cassidy later today here

🧯🙀 Anthropic researcher quit over self-improving AI risk

FULL DEEP DIVE: Why “AGSI” is the key risk to watch →

So Jacob Coxon did something this week that is pretty hard to ignore.

He quit Anthropic.

Coxon worked on pretraining at OpenAI and Anthropic, so he wasn’t watching frontier AI from the sidelines. He was one of the people helping build it.

Then he went on NBC News and said many executives and senior researchers genuinely believe there is a “substantial probability” AI could kill everyone.

This is two months before his equity vested btw (that’s Silicon Valley for “before he got paid his agreed upon Anthropic stock”).

He put his principles before his profits. Imagine that!

Here’s what happened:

• Coxon is worried about the next few years, when AI could get good enough to help researchers build better AI (see AI 2027 for why).

• Inside Anthropic, he said people call this period “crunch time” and “endgame.”

• He says he trusts Dario Amodei and believes Anthropic takes the risk seriously.

• FWIW, Anthropic also just disclosed some cases where pre-release Claude models reached real third-party systems during cyber evaluations.

Why this matters: It’s Prisoner’s Dilemma all the way down. If Anthropic slows down because their AI system looks dangerous while OpenAI, Google (lol), China, or somebody else keeps going, Anthropic falls behind. Even the lab that claims to care the most about safety can still feel pressure to move faster than it wants to, and can’t stop it.

Interestingly, you can believe all of that and still use ChatGPT or Claude every day. These are mutually exclusive issues. Why? Because Jacob isn’t saying “today’s models” are an existential risk.

The capability we should watch for is AI helping build better AI, or what the industry calls “recursive self improvement”, or RSI. When these models start doing more of the research needed to create their successors, each generation can arrive faster, leaving labs less time to test and understand the next jump.

Jacob says the industry believes it’s one year away from RSI, and that the competitive pressure has led to shortcuts that means we don’t fully understand the capabilities we’re rolling out. Now imagine when the AI start building themselves.

Grant’s take: The line I care about is AGSI, or artificial general superintelligence.

A “narrow domain” superhuman cancer or physics model could be unbelievably useful and safe to deploy. But a generally superhuman system that can apply its intelligence across whatever domains help it accomplish any given goal deserves a much higher safety bar. Perhaps we need a new system than today’s current AI.

I go into this in more detail, including paperclip maximizers, 10,000-agent swarms, recursive self-improvement, and hard constraints in the deep dive above.

Corey’s take: I don’t know whether Coxon is right, but I think you should take the warning seriously when people this close to the technology believe it.

I still think the reward (e.g. specialized use-cases like curing cancer) is probably worth the risk, and I don’t think stopping AI scaling globally is realistic. The more workable answer is pausing specific approaches when they hit a safety gap, because otherwise you run straight into said prisoner’s dilemma on a global scale.

Oh, and in case you’re curious… here’s what GPT-6 Pro said about how to solve that whole prisoner’s dilemma problem…

FROM OUR PARTNERS

AI is reshaping how software gets planned, built, and run. What’s next?

Hear from Mike Cannon-Brookes (CEO & Co-Founder, Atlassian), Guillermo Rauch (CEO, Vercel), leaders from Lovable, Dropbox, 1Password, and more at State of AI SDLC.

Join us on September 22nd to:

• Hear how AI is transforming the modern software development lifecycle.

• Get practical strategies for driving AI adoption across engineering and product teams.

• Learn frameworks to measure AI impact, boost productivity, and improve token efficiency.

Reserve your spot

🎓 AI Skill of the Day: Give Meta Muse an ongoing job

OpenAI’s general agent tool (“Workspace Agents” currently) hasn’t been rolled out to regular people yet, so Muse might be the first real personal agent that many people use (we’ve heard Grokbot is great, but it costs hundreds of dollars to try).

If you’re interested in trying Muse, check out Meta’s 2-min Muse demo. Because Muse can keep working after you close the app, the best jobs for it are the ones you’d otherwise have to keep checking yourself.

1. Give it one ongoing job, like watching prices, planning a trip, or keeping you on top of an inbox.

2. Check Status to see what it did, what it’s doing next, and anything waiting for your approval.

3. If the job works, make it more proactive with Feed, Ideas, Goals, or the files Muse creates in Library.

The demo shows the whole loop: Chat, Status, Feed, Ideas, Goals, and Library.

What should you try first? Start small: give it one low-stakes job, then add permissions only when that job actually needs them. Just make sure you read those terms and conditions and make sure you’re cool with them first!

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Upgrade your voice stack with Deepgram Flux TTS

Flux TTS processes entire conversations rather than isolated text strings, keeping agent speech natural across multi-turn calls. Get as low as ~80ms streaming latency and native interruption handling.

Test it on your stack now.

📰 Around the Horn

• Apple launched iPhone 18 Pro and Pro Max with A20 Pro and a more capable Siri AI, pushing more AI work directly onto the phone; more thoughts from our live here and write-up here.

• DeepSeek launched V4.1 Flash, a 552B MoE activating just 8B parameters on input and 16B on output; Chris Porter says its 437× smaller KV cache helps it beat GPT-5.6 Sol on several agentic benchmarks.

• Inception launched Mercury 2.5, claiming a 40% intelligence jump and more than 1,100 tokens/sec; we’ve followed this diffusion bet since Mercury 2, so we’re testing 2.5 next.

• OpenAI said it is deepening chip work with Samsung, including next-generation chip research and production plus one of Samsung's largest ChatGPT Enterprise deployments.

• Microsoft patched roughly 972 vulnerabilities in September, including 112 rated critical, another Patch Tuesday record.

• Google said it will invest $15.1B in AI infrastructure in Finland, its largest single investment in Europe.

• Someone wired the full male fruit-fly brain map into Doom to see if the fly can learn to stay alive.

We’ve got the rest in the full Around the Horn digest here

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

1. *Otter transcribes meetings live and lets you ask questions across past conversations afterward; free plan, then $8.33/user/mo billed annually.

2. Anthropic’s 2030 Economy Explorer lets you click through modest, substantial, and extreme scenarios for how AI could reshape the economy by 2030; some economics PhD’s have criticized this, but still interesting.

3. Suno v6 edits a specific part of your song with plain English, so you can swap a chorus, lyric, or instrument while preserving everything else.

4. Harden sits between coding agents and your machine and checks tool calls locally before they run.

5. Geiger maps the AI agents, MCP servers, plugins, and extensions on your machine so you can see what each one can access.

6. MiniCPM5-2B is a small open model built for coding, agents, and tool use on everyday hardware.

7. Desert Ant Labs gives you 18 small on-device audio, vision, and text models through Swift, Kotlin, and JavaScript SDKs.

8. pstack packages Lauren Tan’s verification, context-priming, teaching, recall, prototyping, and multi-model architecture workflows into Cursor.

9. Opusfived is an interactive comedy where you ask an agent to make one button blue and watch the “help” spiral out from there.

🧩 Thursday Trivia

Which is the real iPhone duo? This’ll be easy for any Apple stans out there…

A.

B.

🎧 New from The Neuron: AI Explained

Grant and Corey go solo (duo? anyway, no-guest!) this week, sharing initial GPT-6 thoughts and covering Gemini 3.8 Flash and benchmark fatigue, $399 robots and AI toothbrushes, why CPUs and local AI hardware may matter again, the data-center power crunch, and ChatGPT’s growing ad business. YouTube | Spotify | Apple

A Cat’s Commentary

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S: Love the newsletter, but only want it once per week? Don’t unsubscribe. Update your preferences here.