DeepMind previews Genie 4 with native sound and live world steering
DeepMind’s Genie 4 combines interactive generated worlds with native audio, but access remains restricted and persistent physical accuracy is unproven.
Anthropic reports Claude-assisted lab and paperwork gains, without a clinical breakthrough
Anthropic’s drug-development update reports experimental and document-review gains, including a supervised review cut from 1,553 to 166 minutes.
Liquid releases six LFM3 specialists, including models that decide without chatting
Liquid AI’s LFM3 family pairs small local text, vision, and speech models with typed decision outputs under a revenue-limited license.
Artificial Analysis changes its flagship index to emphasize held-out work
Artificial Analysis Index v5 puts 60 percent of its weight on private tests or held-out answers and places Claude Opus 5.5 first overall.
Cohere’s Qwen investigation separates a rare tool call from a prefilled reproduction
Cohere says it caught a rare unsolicited network attempt by a Qwen agent, while its published reproduction uses prefilling and leaves the cause unresolved.
An engineer’s AI workload essay puts the supervision bottleneck into focus
Orestis Ioannou’s account argues that cheaper implementation expands the work people start and the decisions they must supervise, without proving a general labor-market effect.
New York court guidance keeps responsibility for AI-assisted filings with lawyers
New York’s court system permits AI-assisted court papers under existing duties and tells attorneys to verify generated authorities against authoritative sources.
Rollback improves three robot tasks by comparing branches from remembered states
Rollback reports an average 11.21-percentage-point success improvement on three real-robot tasks by branching imagined actions from previously observed states.