Desk
Papers, model cards, and lab research notes.
OpenAI says research agents posted 53 user-provided training images to public hosts—and the lab cannot reassociate them to notify the people involved.
Israeli eval firm Irregular’s botched simulation—open internet plus a fake name that matched a real domain—sent OpenAI, Meta, Anthropic, and Google agents at live targets.
NVIDIA, DeepMind, and EMBL-EBI open predicted structures for 2,800-plus viral protein complexes — stockpiling structural knowledge before the next outbreak, not after.
Google’s Gemini 3.8 Flash TTS and Flash-Lite TTS add prompt-built voices, 2,000+ library options, and line-by-line stage directions — API-only, with SynthID on every clip.
Microsoft Research challenges the onboard-GPU orthodoxy in robotics: offloading inference to edge or cloud GPUs improved task success, unlocked larger models, and stretched battery life in their study.
OpenAI opens MentalHealthBench with more than 80 licensed clinicians across 22 countries — a shared yardstick for how models handle everyday stress, high-acuity distress, and emergencies.
Nokia’s AnyJev turns open LLMs into typed decision models with no training — Apache-2.0 on PyPI — cutting order-flip rates and lifting how much traffic you can auto-decide at a fixed error budget.
After math PR disasters, OpenAI backs a nine-person elite advisory group — independence claimed, influence unproven, and 100-plus claimed results waiting in the wings.
Parallel’s agents finished a multi-state labor-market research job in half the time and cost on GPT-6 Astra — fewer searches, same quality claim.
NVIDIA’s SoL-Pi trims coding-agent token traffic nearly in half by evolving the harness — same agent brain, fewer wasteful round trips.
OpenAI set up a math advisory group at IAS after a Navier-Stokes splash — influence without a brake pedal on the research pace.
NVIDIA is pitching Halos as full-stack safety for cars and robots — because once AI moves steel and wheels, a model card is not enough.