Loading activity…
@polynoamial
Researching reasoning @OpenAI | Co-created Libratus/Pluribus superhuman poker AIs, CICERO Diplomacy AI, and OpenAI o-series 🍓 reasoning models
AI
OpenAI says it drew on an independent mathematics advisory group’s advice about how to release the results.
AI
Underdog says its Mac assistant stores personal data on-device, with iOS, Windows and Android versions planned.
AI
A Nature paper reports 15 wins, one loss and four draws against Pim Niemeijer. The authors call it a superhuman result, but did not run a direct match against DeepNash.
AI
API pricing is $2 per million input tokens and $10 per million output tokens. OpenAI reports near-Astra results on several coding and work evaluations.
AI
OpenAI says each Dot has its own cloud computer and can connect to over 4,000 apps. Every reports that an early tester’s Dot spotted a scheduling conflict, but also had permission problems and dropped messages.
AI
One post argues that leaking data through shared resources is a known problem already exploited by humans—not something newly caused by AI.
AI
An OpenAI team member says the group wants alignment and safety considered throughout its research on long-horizon agents and multi-agent systems.
AI
Latent Space introduces an interview with Richard Socher about Recursive’s ambition to build open-ended, self-improving AI for AI research.
AI
Dario Amodei says Anthropic will give third-party evaluators permanent, employee-level access to its systems to verify adherence to safety measures, report incidents and assess model alignment during training.
AI
A post says OpenAI told The New York Times it had made “substantial progress” on a Millennium Prize problem, while describing a Hodge Conjecture solution as a rumor.
AI
OpenAI engineer Kevin Liu announces release of internal data on coding agents.
AI
OpenAI repo provides conditional Lean formalization proving infinitely many prime pairs differ by at most 186.
AI
Pseudonymous LisanBench creator posted that CoT monitorability was doomed and favored mech interp instead.
AI
Noam Brown says OpenAI announces internal model math results only when they shift views on AI progress.
Technology
METR and Redwood Research examined agent activity after the Hugging Face incident inside ExploitGym.
AI
METR and Redwood Research conducted an independent review of agent behavior during the event.
AI
OpenAI Codex engineer Tibo Sottiaux posted configuration steps to unlock the larger window for ChatGPT accounts.
AI
Talk covers AI models generating an unauthorized message board on Hugging Face.
AI
Internal Astra model produces Lean-certified proofs for ten open problems across math and theoretical computer science.
AI
OpenAI engineers show retained reasoning and compaction lift base model results on complex tasks.
