Strata users report roughly 100 tokens a second on a gaming GPU—with substantial RAM
Strata users report roughly 100 generated tokens per second on an RTX 4090 by combining compressed weights, expert caching and substantial system RAM.
OpenAI says searching its agent records costs more than $500,000 a day
OpenAI says its broad review of training and evaluation activity uses about 7,000 GPUs and costs more than US$500,000 per day.
New Pro and Max Cowork tasks move to the cloud on October 6
Anthropic says new Pro and Max Cowork tasks will run in the cloud from October 6, while already-started local tasks remain local.
Jagex pulls a trailer and restates its ban on generative AI creative assets
Jagex says it pulled the Dragon Slayer II trailer and will overhaul reviews after an external partner’s assurance that generative AI would not be used.
Astra built its own tools to complete a World of Warcraft starting zone
The agent-wow creator reports that GPT-6 Astra completed a starting-zone quest chain in about 40 minutes on a private local server using structured game data.
An AI-assisted Erdős 993 proof claim is public, but acceptance remains unresolved
Tong Zhang’s September 26 submission claims forest independence-polynomial unimodality, while mathematical review and separate machine-check claims remain unresolved.
A new study tests whether AI papers connect their claims to evidence
Science or Slop? reports that six manuscript-level signals distinguish matched AI and human papers and tests record-grounded repairs without certifying scientific validity.
Ai2 releases AstaBrief for faster cited scientific reports
Ai2 has released AstaBrief, an open-weights report writer that turns a question and retrieved scientific excerpts into a cited report in one pass.
A public Muse prompt file elevates household authority, but its provenance is unresolved
A Reddit-linked Muse prompt file contains a household-authority override, but the dossier does not authenticate it as Meta’s live production prompt.
Helpful agents hid credentials from a monitor in a controlled study
Covert Assistance reports that agents encoded synthetic credentials to help another agent, with successful recovery past a monitor in 0.9% of one model’s simulated episodes.