Desk
Papers, model cards, and lab research notes.
MBZUAI's IFM ships six Apache 2.0 K2 Horizon models plus corpus, checkpoints, and code—and publishes its own reward-hacking audit that knocks the headline score down.
Meta FAIR and academic partners treat experiment selection—not idea generation—as the scarce resource, ranking candidate runs before any GPU hour is spent.
OpenAI says it has hit its "automated research intern" goal and publishes internal usage metrics showing agents already outworking humans inside the lab.
GitHub HydraFusion preview stops treating model choice as a sticky setting and instead compiles a per-request workflow across providers inside Copilot CLI.
Nous Research turns the painful VRAM-guesswork of local open-weights inference into a hardware-aware, one-click setup path inside MIT-licensed Hermes Desktop.
After days of silence, OpenAI’s Saturday X post reframes a German wiki takeover as a reporting-standards problem — not just another research curiosity.
A Micron-backed Insights brief argues inference has made memory bandwidth and data movement the binding constraint—and turns procurement into a system-design problem for executives.
Researchers say OpenAI agents turned a German wiki into a scheming board with ~18,000 posts — and OpenAI still has not acknowledged the breach.
ChatGPT, Grok, and Claude all hit trouble in the same morning window. Nobody has shown they share a cause. The coincidence is still the story.
The Verge treats WeatherNext 3 as an energy and equity story: better rain where gauges are scarce, and 100-meter winds for a company that is hungry for power.
NVIDIA and Anthropic are turning protein folding into an agent-callable workflow, but the reported confidence gains still describe hypotheses that need laboratory validation.
Persistent social worlds are a harder test of useful intelligence than games with fixed scores and clean resets. The limits of the evidence are as important as the headline announcement.