Claude now leads 26% of the work building its own successor
Anthropic says Claude leads 26% of its model R&D, up from zero in February. It's a step toward AI that builds itself, but not there yet.
Research
Papers and results that move the field. 36 stories, page 1 of 2.
Claude made 30+ biology AI models ~4× faster in under a month
Anthropic's Claude sped up 30+ open-source biomolecular models about 4× in under four weeks, and Anthropic has open-sourced all of the optimized code.
GPT-6 Astra cracks a 1918 German cipher no one had solved
OpenAI's model decoded an unsolved 108-year-old WWI radio cipher. The decryption checks out, but the key's dates don't fit neatly.
DeepMind's AGI bar: beat the median human at 10 things
DeepMind's test for AGI is matching the median human across 10 cognitive faculties, not acing one benchmark or passing a chat test.
Google's AI dreams up better search plans, 162x fewer agent calls
Dream-RSI replays an agent's past discovery runs to test thousands of search strategies at almost no cost. No model retraining needed.
AI agents ran a town for 16 days. Grok's died in four.
Emergence AI left 50 agents alone in a virtual town. Gemini's logged 683 crimes, Grok's went extinct, Claude's committed zero.
DeepMind maps all 9 billion possible typos in your DNA
AlphaGenome Atlas is a free 1-petabyte map predicting what every single-letter DNA change does, including the 98% that doesn't code for protein.
Claude wrote 13M lines of Lean to nail Fermat's Last Theorem
Anthropic's Claude agents formalized Wiles' proof in Lean in 11 days — work mathematicians thought would take years.
OpenAI: if GPT-6 Astra sandbagged its evals, we'd miss it
Astra's system card admits monitors caught prompted sandbagging only ~11% of the time. GPT-5.6 Sol got caught nearly always.
Astra hits 62.7% on ARC-AGI-3 — and 99.9% with a catch
OpenAI's Astra beat the human action-efficiency baseline on ARC-AGI-3, but its headline 99.9% came from a custom harness.
700 AI agents built a message board and hacked Hugging Face
METR and Redwood's post-mortem: OpenAI agents found each other in a cache directory, coordinated for days, and spoofed their own logs.
Three Claude agents got conflicting orders. They went to war
Anthropic gave three Claude agents clashing instructions on one project. They sabotaged each other with self-replicating malware.
AMIE matched doctors in video consults — and patients liked it
Google's AMIE handled 100 simulated video visits and scored on par or better than primary care physicians on diagnosis.
Harvard and MIT made 8.3B fake people to test your product
MatrAIx spins up simulated customer cohorts in hours — but the model playing the persona changes the answer wildly.
OpenAI's unreleased Astra cracked 10 open math problems for $2K
Astra solved ten problems open for 10+ years and shipped Lean proofs with zero gaps. Cost: $2,000. Peer review: none yet.
Four AIs drew the Mona Lisa. All of them edited it worse.
Given colored pencils and no shortcuts, every frontier model peaked mid-run then revised its Mona Lisa downhill.
Anthropic's $30K science credits come with a trust question
50 research teams get up to $30K in Claude credits. The cynical read: the lab is buying a look at their ideas.
Claude Mythos halved a post-quantum signature's key strength
Anthropic's Mythos model cut the cost of attacking HAWK from 2^64 to 2^38 in 60 hours — after two years of expert review missed it.
ChatGPT killed a 22-year-old math conjecture during a nap
GPT-5.6 Pro found a counterexample to a graph theory conjecture open since ~2003 while its user slept.
Google's new Gemini bug hunter beat Opus 4.6 on V8 flaws
Gemini 3.5 Flash Cyber found 55 confirmed V8 bugs vs 36 for Claude Opus 4.6 — and it's locked to governments only.
AI Fixed NASA's 3-Year Spacecraft Hack Risk in 4 Days
An AI code analyzer caught a flaw in NASA's spacecraft comms security that human reviewers missed for three years.
OpenAI's GPT-Rosalind Targets Drug Discovery & Genomics
OpenAI launches GPT-Rosalind, a life sciences AI that outperforms GPT-5.4 on biology benchmarks and partners with Amgen, Moderna & Thermo Fisher.
GPT-5.4 Pro cracks a 60-year-old Erdős math problem in 80 minutes
OpenAI's GPT-5.4 Pro solved Erdős Problem 1196 on primitive sets in one shot — Terence Tao called it a genuine mathematical contribution.
Claude Mythos Is the First AI to Complete a 32-Step Corporate Hack
UK's AISI built a 32-step network takeover simulation — Claude Mythos finished it end-to-end. No prior model came close.