In partnership with
IN THIS ISSUE
Welcome back to The AI Field.
The people building AI said the quiet part out loud. A researcher quit Anthropic warning that the labs are racing toward self-improving AI, and Anthropic's own alignment lead agreed with him. The same day, OpenAI announced that 10,000 of its agents had cracked a 90-year-old math problem in 88 hours. And the two biggest names in AI audio just bet their next products on licensed music.
- An Anthropic researcher quit and the lab's alignment lead put extinction odds above 10%
- OpenAI says its agents solved the Navier-Stokes Millennium Prize Problem
- Suno retrains on licensed music as Universal signs with ElevenLabs
Read time: 5 minutes
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
RECENT UPDATES
01 · ANTHROPIC
Jacob Coxon, a 27-year-old researcher who worked on GPT-4o at OpenAI before joining Anthropic this year, quit with a thread on X that has passed 100 million views. His charge: neither lab is acting responsibly. "They are racing straight to self-improving superintelligence and gambling with our lives."
He wrote that the people building AI "earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt." His fix is a coordinated slowdown, possibly "a temporary ban on improving model capabilities."
Then Evan Hubinger, Anthropic's Alignment Science lead, replied that "we really do earnestly believe AI could kill all humans!" He put his own odds above 10% within the next decade and said Anthropic does "not yet have a plan to solve alignment for superintelligence." He later clarified that today's models are low risk. Self-improvement is the danger.
The AI Field: Resignation threads aren't new. An alignment lead agreeing with one in public, with odds attached, is. The case for safety-first labs has always been that careful people need to be at the frontier. This week one of them left, and another admitted the plan for superintelligence doesn't exist yet.
02 · OPENAI
OpenAI says an unreleased internal system, more capable than GPT-6 Astra, solved the Navier-Stokes existence and smoothness problem, one of the seven $1 million Millennium Prize Problems. The proof shows that a smooth fluid can develop a singularity in finite time, a question mathematicians have chased for about 90 years.
About 10,000 agents worked on it at once. They launched on September 1 and hit the result 88 hours later, using roughly 130 billion output tokens. GPT-6 Astra then formalized the proof in Lean in another 17 hours, so the logic is machine-checked.
Two asterisks. The proof uses a smooth external force, which the official problem allows, but the unforced version most mathematicians care about is still open. And NYU's Tristan Buckmaster, who announced a related result with Anthropic's Levent Alpöge the night before, questions whether OpenAI built on their work. OpenAI says it didn't use their prompts or proofs, and it won't claim the prize.
The AI Field: Forget the prize for a second. A lab just pointed 10,000 agents at an open research problem and came back with a machine-checked proof in less than a week. Put that next to Coxon's thread and the nerves make sense. Automated research isn't a slide in a pitch deck anymore.
03 · AI MUSIC
Suno launched v6 on Wednesday, built with licensed music from Warner Music Group, BMG and Believe, and says it will retire every previous model. v6 and v6-wild are for paid plans. v6-mini is free for everyone.
A day later, Universal Music Group signed a multi-year deal with ElevenLabs, the $11 billion voice AI company, to build a licensed platform where fans can remix, mash up and reinterpret songs from artists who opt in. It's ElevenLabs' first major-label deal. No launch date or payout details yet.
The lawsuits aren't done. Universal and Sony are still suing Suno over how its earlier models were trained.
The AI Field: The labels never wanted to kill AI music. They wanted to get paid, and this week the two biggest bets in AI audio showed they'll pay. If you make music with AI, check what your tool was trained on. Licensed models are where commercial use is heading.
THE AI TOOLBOX
TRENDING RIGHT NOW
Six tools worth ten minutes each. Click the name to open it.
👨💻 base44: Describe the app you want and it builds the frontend, database, login and hosting in one go, then publishes it on your own domain. Free plan to start.
🎵 Suno v6-mini: The free version of Suno's new model family, built with licensed music from Warner, BMG and Believe.
🖼️ ChatGPT Images 2.5: Up to 50% faster than Images 2.0, and the new Sketch feature turns a rough doodle into a finished image.
📊 Econ Scenario Explorer: Anthropic's free tool lets you plug in your own AI predictions and see what they mean for US jobs, wages and GDP by 2030.
⚡ Mercury 2.5: Inception's diffusion model writes at over 1,100 tokens per second and launched with 100 million free API tokens.
🤖 Muse: Meta's new personal agent sends emails, books travel and fills out forms from the app or WhatsApp. Only in the US for now, where it's already No. 2 on the App Store.
QUICK HITS
OpenAI adds Paul Christiano to its nonprofit board. The alignment researcher, who did foundational work on RLHF, also joins the Safety and Security Committee. His line: "alignment remains a difficult technical problem."
OpenAI paused new sign-ups for its $200 Pro plan. Demand for Astra is "really unprecedented," according to OpenAI's Thibault Sottiaux. Plus, Go and the API aren't affected.
Anthropic says Alibaba-linked accounts ran 151 million Claude exchanges. The campaign ran from May to July across 3,500+ accounts to train Qwen models, peaking near 3 million exchanges a day. Moonshot AI and DeepSeek were named too.
Nvidia expects about 70% revenue growth next year. Jensen Huang told a Goldman Sachs conference that Nvidia powers "every model" at the major labs and is "tracking every single gigawatt of land, power, shell around the world."
MORE FROM THE FIELD
WHAT TO WATCH · WHO GETS THE CREDIT
OpenAI says it won't claim the $1 million, and the Clay Institute's rules wouldn't allow it yet anyway. A solution has to be published in a qualifying outlet, sit for at least two years, and win general acceptance before the prize is even on the table.
The real fight is over credit. Charles Fefferman, who wrote the official problem, called Diego Córdoba and Luis Martínez-Zoroa "the heroes of the story" for the technique both teams built on. When 10,000 agents do the grinding, someone still has to decide whose idea it was.
YOUR TURN
Your feedback shapes what we cover next. Pick the answer that fits best.
What did you think of today’s edition?
Until next time!
Olle
Founder of The AI Field