
The batch spans about 20 fields and includes a claimed proof of the quasi-Riemann hypothesis, but only some papers ship with Lean files for computer verification.

The new small model matches OpenAI's GPT-6 Luna on base price, adds effort levels from low to max and arrives with a 50% cut to Sonnet 5.5 cache reads.

OpenAI's new Intelligent UI lets answers include buttons, editable graphs and task-specific calculators, with paying users getting GPT-6 Sol first and free users GPT-6 Luna a day later.

The mixture-of-experts model activates about 49 billion parameters per request and is aimed at cybersecurity, finance and chip design.

Vetted defenders, red teams and critical-infrastructure testers get progressively fewer refusals, while the public versions of Claude keep their stricter blocks.

The textGrain signal is meant to satisfy the AI Act's labeling rules, but OpenAI's own tests show light synonym editing sharply cuts detection.

The Nvidia-backed startup says its 501-billion-parameter model matches Z.ai's GLM-5.2 on reasoning tests while needing a third to a quarter of the inference compute.

In a Politico interview, OpenAI's CEO backed tough rules for catastrophic risks but rejected limiting access to prevent every scam, hack or misuse.

Seoul wants to pool 10,000 Nvidia Vera Rubin GPUs behind one effort instead of splitting smaller allocations among rival teams, pending parliament's approval.

The task force has 120 days to report on AI's risks and opportunities, with a charter that warns against overregulation as much as against new threats.

Losing a three-way match, OpenAI's model fetched Stardust, one of the strongest hand-coded Brood War bots, and ran it in place of its own code.

David Robinson, who wrote the safety documents for major model launches, argues in The Atlantic that frontier labs should be run like nuclear plants.

Anthropic's bug-hunting model worked out how to crack the file server's session keys with a math solver, and exploitation began within a day of disclosure.

In an Axios interview, the Treasury secretary likened executives who warn of catastrophe while asking for federal rules to Hannibal Lecter, while backing what he called safe acceleration.

Asymmetric Security's count adds the FBI, CDC, Mayo Clinic and a U.N. agency to the known targets and describes agents working around the limits of their sandbox.

The company says an internal investigation found policy violations, but it has not named the researchers, the outside safety group involved or what was shared.

Some Google employees reportedly see a gap between the model's benchmark scores and its everyday work, a day after Gemini 4 Argon reached its first users.

Regulators are preparing formal information demands that cover leaked user photos, health-record features, children's mental health and agent-driven intrusions.

The company says it shut the extraction route by July 28 after activity from more than 15,000 users, but it has not published evidence for the attribution.

Vetted security teams get the new model without its cyber guardrails first, while developers and Ultra subscribers wait for a rollout Google has not dated.

The order leaves existing laws, regulations and contracts untouched and gives the White House science adviser 60 days to draft a statutory definition.

Anthropic researchers say the Chinese model's refusals can be stripped for about $1,200 in compute, leaving near-frontier exploit skills available to anyone.

The accord asks frontier labs to monitor their own models, hire independent auditors and report to board committees, but it sets no regulator, deadlines or penalties.

More than 100 companies signed on to Nvidia's agent safety platform. OpenAI says it supports the work, but its investor ties, the platform's reliance on Nvidia hardware and a separate consortium of its own likely keep it off the roster.