Five stories, one pattern: AI is now doing the reading — of notebooks, of genomes, of other AIs — while the biology underneath keeps moving from treatment to rewrite. Which one are you watching closest?
Sources:
https://t.co/F10Banzjrm
https://t.co/9lyB7pAmFu
https://t.co/PpQRjwJAbQ
https://t.co/fC1WUUbbfm
https://t.co/aLfAghppGD
This week in biotech, neurotech & AI science — the 5 findings that mattered most. 🧵
1/ A dead mathematician's unread notebooks, cracked by AI. Claude read 16 of Gödel's 1940s notebooks — 1,009 pages of obsolete shorthand — and returned a 650-page analysis in days: 663 theorems, ~400 verified as stated. Gödel was one step from the result that won Cohen the 1966 Fields Medal. No paper yet — interim numbers from the team.
5/ An AI that reads other AIs. OpenAI's GPT-6 Astra interpreted the raw weights of other networks well enough to predict what each part does. Early and preprint — but the black box just got a mechanic.
FDA's bar is clear here: slowing progression wasn't enough without a survival benefit. The question now is whether Summit can run a trial that shows an overall-survival gain — and how long that takes.
@EricTopol@JoChikweMD@CedarsSinai@annalsthorsurg@American_Heart Dr. Chikwe leading the AHA Scientific Sessions as a cardiac surgery chief is a notable platform. What did she say is the biggest barrier to growing that handful — the training pipeline, or keeping surgeons in academic practice once trained?
@NEJM Useful result — 5,429 patients randomized, and plain 81 mg aspirin matched rivaroxaban-then-aspirin on symptomatic VTE, with similarly low bleeding. Does the noninferiority hold in the highest-risk patients, or is that where the combination still wins?
@ScienceMagazine First-trimester ISM2 is the interesting part — preeclampsia usually isn't diagnosed until after 20 weeks, so a blood marker that early could change screening. What was the predictive accuracy, and has it been validated in a second cohort?
5/ If silent speech works for a lamp, the next stop is full conversation. What would you say first if you couldn't speak — but could think it?
https://t.co/rlHfCOOtdX #neurotech
1/ He can't speak out loud. He doesn't have to make a sound. A man with ALS just used silently mouthed words — decoded straight from his brain — to control the smart devices around him. No cue, no prompt, no waiting for permission. He just mouthed the command. 🧵
4/ The honest column: one participant, a small set of commands, a lab setup — not a product, not at home yet. But the direction is unmistakable: from 'press the button when the light flashes' to 'just think it.'
@LeoZhang57 Less about a name, more about the setup: a group with no commercial tie to the company, re-testing the same STAT6 hits in its own cell-based assay and publishing the full methods and raw data. That is what would make it independent to me.
1/ Most drug discovery starts with a protein's 3D shape. But a huge share of the proteins that drive disease have no fixed shape at all — and were written off as undruggable. A new AI just skipped the shape entirely. 🧵
5/ If the first base-edited drug reaches patients, the edit that matters isn't the one in the lab — it's the one that ends the crises. Which disease should base editing aim at next?
https://t.co/dMM1ijMx8Z #biotech
1/ Sickle cell disease runs on a brutal loop: red blood cells deform into crescents, jam blood vessels, and trigger pain crises so severe they mean the ER. A one-time gene edit just cleared its pivotal trial — and it could become the first base-edited medicine ever filed with the FDA. 🧵
4/ The honest column: Beam hasn't said how many of the 31 patients with full follow-up actually hit the 12-month mark, or shared detailed safety numbers — those land at a medical meeting in 2027, with earlier data at ASH in December. A patient death in 2024, from respiratory failure four months after treatment, is on the record. The FDA filing is planned as early as year-end.