Welcome to Jumble, your go-to source for AI news updates. This week, DeepMind scored all 9 billion single-letter typos in human DNA. Meanwhile, OpenAI claims a million-dollar math proof, and the mathematician it beat says the company played dirty. Let’s dive in ⬇️
In today’s newsletter:
🔬 DeepMind scores 9 billion DNA edits
⚔️ A Millennium Prize claim turns ugly
🏛️ CA becomes the first state to register AI auditors
💳 Card networks agree on how to ID shopping bots
🎯 Weekly Challenge: look up a famous gene variant yourself
🩻 DeepMind Just Scored Every Possible Typo in Human DNA
On September 8, Google DeepMind released AlphaGenome Atlas, a free database predicting what each of the 9 billion possible single-letter changes in human DNA does. Doing that in a lab would take many human lifetimes; the dataset weighs a petabyte.
Would you look up your own genome in a tool like this?
📊 One Score to Rank Them All
Every variant gets one AlphaGenome Variant Impact score, where a 30 means the top one in 1,000. In beta tests it put the true causal variant in a patient's top 50 candidates 29.5% of the time, versus 12.5% for existing methods.
🩺 Already Finding Real Answers
Broad Institute researchers used it to catch a DNM1 variant behind a child's epilepsy, later confirmed at the bench. The catch: it is free for non-commercial use only, weak on enhancers, and not validated for the clinic.
⚖️ Did OpenAI Play Dirty to Win a Million-Dollar Math Problem?
On September 8, OpenAI said an internal model beyond GPT-6 Astra proved the 3D Navier-Stokes equations can blow up, resolving a Millennium Prize Problem. Roughly 10,000 agents took 88 hours and 130 billion output tokens, and the proof checks out in Lean.
🕵️ The Mathematicians Who Got There First
NYU's Tristan Buckmaster and Anthropic's Levent Alpöge had quietly worked the same route since mid-August, with their Euler blowup verified in Lean on August 22. Buckmaster says OpenAI called after hearing rumors, and never answered clearly whether its models trained on his Codex sessions.
🔥 "I Don't Have to Be Nice"
Buckmaster says OpenAI's Sébastien Bubeck proposed dropping Alpöge from the paper, asked "Why would you ruin your career?" when he threatened to go public, then said: "If you don't want me to be nice, then I don't have to be nice."
OpenAI denies using their prompts or proofs, cannot rule out that de-identified usage data helped, and will not claim the $1 million prize.
Weekly Scoop 🍦
📜 California starts licensing the people who audit AI
🛒 Visa, Mastercard, and Ant align their agent ID protocols
🤖 OpenAI considers slowing AI model development
💻 Cognition post-trains Kimi K3 into its cheapest coder
🎓 Underemployed Chinese lawyers grade AI answers for $15
🗂️ Cursor's new coordinator agent manages months-long builds
🧪 Weekly Challenge: Look Up a Famous Gene Variant Yourself
Challenge: AlphaGenome Atlas is free to browse, so this week you become the genomics lab. Pick a famous DNA variant, pull its score, and have ChatGPT, Gemini, or Claude translate it.
📍 Step 1: Pick a famous variant Try rs4988235, the single letter behind adult milk digestion, or rs1805007, one of the red hair variants.
🔎 Step 2: Search it in the Atlas Open alphagenome.google/atlas. If it wants coordinates instead of an rs number, ask your AI for the GRCh38 position first.
🧠 Step 3: Translate the score Paste the score and breakdown into ChatGPT, Gemini, or Claude: "Explain this like I have zero biology background. Is it hitting splicing, expression, or the protein?"
🧾 Step 4: Check it against the research Ask what published studies say about the variant and where the prediction disagrees. That gap is the whole story of this tool.
Does a Lean-verified proof settle who solved Navier-Stokes, or does an 88-hour sprint after a rumor make it hollow? And what would you build with AlphaGenome Atlas now that 9 billion scores are free? See you next time!
Stay informed, stay curious, and stay ahead with Jumble!
Zoe from Jumble
