Welcome back! Anthropic's alignment lead put a number on the end of the world after one of the company's researchers quit over it. ChatGPT's unreleased model cracked a 90-year-old math problem that carries a $1 million prize, and the mathematician who got there first says he was pushed aside. Image gen got twice as fast, and Google DeepMind listed every possible mutation in human DNA.
In today's Generative AI Newsletter:
Anthropic: What odds does its own safety lead give us?
OpenAI: What did 10,000 agents crack in 88 hours?
ChatGPT: Should designers change careers already?
Google: What scientific milestone did it hit?
Anthropic's science lead says we could all be dead in 10 years
Evan Hubinger runs alignment science at Anthropic. His job is finding the ways the company's safety work fails before a shipped model finds them first.
He just put a number on it. More than a 10% chance AI kills every human in the next decade.
He said Anthropic is trying its best, then added the part that stings. The company has no plan to solve alignment for superintelligence and isn’t on track to get one.
He was backing up Jacob Coxon, who quit Anthropic hours earlier and posted why.
Coxon spent three years on pretraining research, first at OpenAI, then Anthropic. He says neither company is acting responsibly.
He said these will soon be superhuman systems that can hack anything and acquire real power, and progress is not slowing.
Executives soften their words for the press, he says, then express the same fear privately.
On why they keep building: at OpenAI many haven't internalized the stakes, at Anthropic they understand them but believe someone less careful wins the race if they stop.
Coxon told the Wall Street Journal he's leaving the industry, and that by the end of next year things could already be out of control.
Anthropic sells itself as the careful lab. This week its alignment lead and a pretraining researcher both said the careful lab doesn't have a plan.
ChatGPT solved a $1 million math problem
An unreleased OpenAI model cracked the Navier-Stokes problem, one of seven Millennium Prize Problems and unsolved for about 90 years. Each carries a $1 million prize. This is the second one ever solved.
The equations describe how fluids move, and they run weather forecasts, aircraft design and blood flow models. The question was whether they can break down and predict infinite speed. Now we know they can.
Around 10,000 agents worked in parallel for 88 hours. Lean, the proof-checking language, took another 17 hours to verify it.
The agents sent 2.7 million messages and burned millions of dollars worth of tokens on this problem alone.
The model behind it is internal, still training, and OpenAI says it's significantly more capable than Astra, which shipped six days ago and everyone is calling it AGI.
Then it got ugly. NYU's Tristan Buckmaster and Anthropic's Levent Alpöge announced their own related proof 12 hours earlier, and Buckmaster says OpenAI heard about their method and used it. OpenAI denies seeing their work and points out it started the effort after hearing a rumor about them.
Altman called it one of the most amazing moments in OpenAI history, then said it's the strongest evidence yet that the industry needs to slow down.
ChatGPT images got a whole lot better
Images 2.5 shipped to everyone today, free tiers included, on desktop, mobile and web. People already make more than 3 billion images every week across ChatGPT and the API.
Generation takes up to half the time it did on Images 2.0.
Sketch lets you draw your idea in ChatGPT and use the doodle as the reference. Type @Sketch to open it.
Comment on a spot in an image to change only that part, the way you'd leave a note on a doc.
Edits hold up over a long conversation, so the fifth change doesn't wreck the first four.
You’ll notice a new progress bar whenever you’re generating images.
Templates cover posters, flyers, merch and product shots. Developers get two API models, Flare for speed and Sunburst for detail.
The reference-photo work is the real upgrade. Faces stay recognizable through costume changes and new backgrounds, which is where every image model so far has let you down.
OpenAI is back to leading the AI race. Makes you question if you should keep your Claude subscription.
Google now has every mutation in your DNA
Your DNA is 3 billion letters long and everyone's has mutations. Most do nothing. A few cause sickle cell or cancer, and finding the one behind a sick patient can take a lab years.
Google DeepMind worked out what every possible mutation does, all 9 billion of them, and put the answers in a searchable database. Type in a mutation and you get a damage score.
Labs used to run that analysis themselves, and now they can just look it up.
A Broad Institute team used it to solve a rare disease case.
Another researcher ran it on 54,000 people and found 19 stretches of DNA tied to body weight.
Free for academics through a browser. Drugmakers need to pay for a license, but they never run out of money so we’re sure they don’t mind.
Every entry is a prediction, and Scientific American says the whole thing is far less accurate than AlphaFold, DeepMind's Nobel-winning protein work.
Step one in working out why a disease runs in your family is now a search box.
Tool of the Day: Goblin Tools
Goblin Tools is nine tiny tools for the moments a task feels too big to start. Magic ToDo takes "clean the kitchen" and breaks it into steps, with a spiciness slider from one pepper to five that decides how small those steps get. Formalizer rewrites a message so it clicks the way you meant, and Judge reads a text back and tells you the emotion in it. This is especially useful for neurodivergent people and free forever with no ads and no paywall. Mobile apps are $2.99 and Pro is $3 a month.
Try this yourself:
Go to goblin.tools and put one dreaded task into Magic ToDo.
Push the spiciness slider to five peppers and watch it split into steps you can start.
Paste an angry email into Judge before you reply to it.
Run your draft through Formalizer to soften or sharpen the tone.
Who it's for: anyone staring at a task they can't make themselves begin.
Everything else you shouldn't miss
OpenAI won't claim the $1 million math prize: it says the goal was to report on the pace of AI progress, not to collect Clay Institute money.
Anthropic published a guide to spending less on Claude: cache hit rates, dropping old "double-check your work" instructions that now cost you tokens, and matching effort to the task.
A mathematician says the medal belongs to someone else: Buckmaster says Luis Martínez-Zoroa, who built the technique both AI teams used, deserves a Fields Medal.
Learn more about AI from the experts building it
📸 Follow us on Instagram for fast, visual AI updates in 30 seconds.
📺 Watch us on YouTube to hear insights directly from leading AI voices, builders, and innovators.
🐦 Follow us on X for breaking AI news and real-time industry updates.
🧠 Learn how to build your next AI application with practical resources and expert guidance.
🎓 Start learning with free AI courses from GenAI Academy.
💰 Invest in the GTM infrastructure of the AI economy.
