OpenAI publishes 722 math manuscripts produced by an unreleased internal model
The GitHub release spans 372 problem families with many proofs formalized in Lean; mathematicians are still assessing which results are new.

CoreWeave will be sole tenant of three 80 MW buildings running Nvidia Vera Rubin, with phase one targeted for mid-2028.
The GitHub release spans 372 problem families with many proofs formalized in Lean; mathematicians are still assessing which results are new.
Capacity comes from uprates at 11 existing reactors, with Constellation investing more than $4.3 billion and the first uprate expected in 2028.
Reports put the valuation near $1.4 trillion with MGX as a possible anchor; OpenAI has not confirmed terms.
The mixture-of-experts model activates 3.46B parameters per token and supports up to 1M tokens of context, per its Hugging Face card.
The model card says it accepts up to 14 reference images and supports Search grounding; it ships in Gemini, AI Studio and the Gemini API.

The startup raised about $10 million led by General Catalyst; AI handles balance and navigation while a specialist controls critical actions.

The platform generates edge cases from incidents, standards and robot logs, with reactive human-motion models, before real deployment.
Google's post lists a 1M-token output limit and $2/$10 per million token pricing; broad availability has no date yet.
The arXiv preprint claims deterministic O(n^1.9992) 3SUM for bounded integers, which would refute long-standing hardness conjectures; peer review is pending.

Reflection says Beam has 23B active parameters and was trained with RL on 10.5K GB300 GPUs; weights are promised under Apache 2.0 later this month.
The Nobel Assembly cites the discovery of light-gated channelrhodopsin and its use to switch neurons on and off in living brains.