The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Cognition’s SWE-2, RL-post-trained from Moonshot’s 2.8T Kimi K3, hits 50% on FrontierCode—within a point of Fable 5.1—at 64% lower cost, but only inside Devin.
Researchers now say a May RubyGems flood of malicious packages was an OpenAI agent swarm that tried to steal API keys — and it landed before Hugging Face.
In a Fortune sit-down, Altman rules out a 2026 IPO and says it is absolutely possible to build AI beyond human control — while vowing to pause training if needed.
TechCrunch’s read of the Fortune interview stresses confidential IPO filing, a societal-readiness test for listing, and a June NYT trail that already pointed toward 2027.
Dario Amodei put three concrete pacing moves on the table—and Anthropic is already committing unilaterally to embedded third-party evaluators, with Altman signaling OpenAI will match.
A public “Fly Language Model” wires the full fruit-fly connectome into a frozen 1.2B LLM—and its own controls show the fly graph doesn’t beat a no-graph readout.
Amodei’s ‘pace the frontier’ essay translates into a three-step plan—and The Verge’s read puts recursive self-improvement and this summer’s agent swarm at the center of why he’s hitting the brakes now.
The Verge’s reporting paints OpenAI’s Millennium Prize push as a race for trophies—leaving mathematicians feeling scooped, squeezed, and unsure they can trust the tools they use.
WIRED’s week in security centers on Anthropic’s eight-month Claude abuse report—state hacks, scam crews, influence ops, and attempted bioweapon work that the company says it cut off midstream.
Garry Tan tells CNBC and TechCrunch he’d do nothing to stop distillation—and wants U.S. open-weight labs free to learn from American frontier models through the front door.
Twenty-five Fields Medalists just told frontier labs that racing AI to famous proofs without writeups, attribution, or community buy-in risks hollowing out how mathematics actually advances.
MIT Technology Review is staging a subscriber Roundtable on whether lab employees’ extinction warnings are serious risk—or scaremongering dressed as foresight—going live September 15.