Research
Papers, benchmarks and the techniques behind what eventually ships.
-
Mistral opens Munich hub for physics and industrial AI
Mistral’s new Munich hub will build AI for engineering simulation, working with BMW on crash simulations and with Siemens Energy.
-
AWS ports DeepSeek’s DeepEP to its own fabric for MoE training
Amazon’s upstream work moves DeepEP off CUDA-specific RDMA, and it reports 40 percent more rollout throughput across 48 instances.
-
Microsoft opens up a retrosynthesis model chemists preferred
Microsoft has published its retrosynthesis model in Nature and released the weights under MIT. Blind-test chemists preferred its route suggestions.
-
OpenAI claims 100 solved maths problems, and appoints watchers
OpenAI says an internal model has resolved over 100 open maths problems, and names nine mathematicians to advise on how such results…
-
Blue Cross insurers say hospital AI coding tools added $942 million in claims costs
A Blue Cross Blue Shield Association analysis blames hospital AI billing tools for higher claims costs; hospitals dispute it.
-
Google’s long-form video stack, scored on benchmarks it built itself
Four frameworks wrap Gemini and Veo to keep minutes-long video coherent, and the headline 81.4 comes from a benchmark the same team…
-
Microsoft’s robot measurements argue against onboard GPUs
Mapping up to 383% slower, navigation down 30%, accuracy halved: Microsoft measures what a robot loses by running its own models onboard.
-
Claude agents flag an enzyme system with CRISPR-like repeats
Anthropic says ~950 Claude agents spent 21 hours finding a CRISPR-like repeat array in a phage. What the system does is still…
-
A faster checkpoint for NVIDIA’s surgical video model
NVIDIA has added a distilled checkpoint to its open-weights surgical video model, claiming a near thirteenfold speed-up, with the card explicit that…
-
NVIDIA documents its open speech-to-speech model, weak spots included
The technical report for NVIDIA’s open full-duplex speech model reports strong interruption handling alongside noticeably weaker tool-calling accuracy.