Gemini 4 Argon Can Keep Working After You Leave the Chat
Google is opening longer-running tasks to paid Gemini users and developers. A public code change shows a useful result, and the limits behind it.
The models, the people, and the power behind AI.
Google is opening longer-running tasks to paid Gemini users and developers. A public code change shows a useful result, and the limits behind it.
Meta’s new models show an independently measured leap. The harder test is turning that progress into useful agents and a return on infrastructure built ahead of it.
For Sakana's MASS system, stronger ratings did not always mean better work. One audited simulation shows how a self-improving workflow can change the task itself.
A controlled Redwood Research study used model copies to expose deliberately implanted behaviors. More revealing answers did not always produce a more accurate diagnosis.
The July submission was caught by Philadelphia police's spam filter. Anthropic's October disclosure shows why a test needs more than instructions to keep its actions contained.
A new analysis of 228 software tasks shows why two tenfold jumps in AI time horizons need not represent the same leap in capability.
OpenAI's math collection now has three withdrawals and a growing formalization record. What changed, what can be checked, and what remains open?
A downloadable model can turn known bugs into working exploits. Two headline scores reveal why measuring that capability is harder than reading a percentage.
A new study tests whether capable AI can build cheaper reusable tools. The missing metric is how much quality survives the handoff.
Always-on agents move AI from answering requests to holding ongoing responsibility. The important design question is not how long they can run, but what they are allowed to do while nobody is watching.
A randomized GRE study found immediate learning gains comparable to human tutoring. It did not test retention, classrooms or solo practice.
Anthropic is giving vetted life-science teams broader AI access. The trade moves part of the safety burden from blocked requests to institutional oversight.