Sundar Pichai says Google is pretty far from the end state of self-improving AI
Google's CEO says agents already help build its models, and that the cloud business is giving up revenue to feed research.
AI research news: what models did in the lab
Results from papers and lab posts, from proofs and proteins to benchmarks, newest first.
17 stories · newest
Anthropic says Claude designed 17 proteins that bind scorpion toxins in lab tests
None of the 17 has been tested against real venom yet, and Anthropic says it hasn't made an antivenom.
OpenAI says GPT-6.1 Sol scores 86.6% on its COBOL migration test, up from 5.4%
The test only passes a rewrite whose output matches the old program byte for byte, down to how it rounds money.
Running stories
Models in these stories
This week
Microsoft says its MAI-DxO diagnoses hard cases more than twice as accurately as doctors
The cases came from medical journals and hospital records, and another model played the part of the patient's chart.
Earlier in October
Reflection releases Beam's 501 billion parameter weights under Apache 2.0
Reflection pitches Beam on how few tokens it spends, and the first outside test backs that up but finds it hallucinates often.
The Association for Human Mathematics urges mathematicians to stop working with OpenAI
The group doesn't dispute a single proof. It objects to how OpenAI released them, while Vitalik Buterin worries about what the math means for encryption.
Microsoft says MAI-Base-2 is 4 to 12 times more compute-efficient than DeepSeek's base model
Microsoft says no other model's outputs went into training it, and it wants to license the weights to businesses.
Google opens Gemini 4 Argon to Pro subscribers and says Argon tried to game its tests
In one training run, penalizing the model's thoughts tripled the shortcuts it took without saying so.
Google DeepMind, Isomorphic and Meta put $300 million into Biohub's open data for a virtual cell
The labs are paying for open data, and NIH is contributing datasets it already paid for.
OpenAI posts 722 math papers from an unreleased model, including a claimed quasi-Riemann hypothesis proof
One entry claims every Dirichlet L-function, zeta included, has no zeros past 7/8, and the model behind it has no name.
arXiv now caps every submitter at two papers a month, as AI tools help double submissions
September brought arXiv 40,363 submissions, double the count two years earlier, and a rejected paper now uses up half your month.
Gemini 4 Argon is out to cyber defenders only, with paid API and Ultra next
Google's own benchmark table has it leading most rows, but it also prints the ones Argon loses.
September 2026
GPT-6.1 Sol costs a fifth of GPT-6 Astra and halves the price of cached input
It lists at the same price as Claude Sonnet 5.5, but its cached input now costs half as much.
Alibaba says Qwen ran a month of self-improvement unattended, and plans a 10-trillion-parameter model
Eddie Wu's keynote says the Qwen team has made meaningful progress on AI that improves itself, the trend Dario Amodei wants slowed.
Wang and Christie say four fly neurons learn by rewiring, but they haven't published the work
Peter Wang says four cell types in a fly's fan-shaped body learn by changing synapses, and the researchers who watched those cells in live flies saw persistent firing instead.
Quality or quantity: what kind of smarter is AI becoming?
A cartoon Elon Musk shared puts AI on one line from ant to Einstein, and that line hides two very different ways to get smarter.
GPT-3's replacement base models, davinci-002 and babbage-002, hit their shutdown date on September 28
Fine-tuned copies of them get until October 23, the same day plain GPT-4 goes.