The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Google booted the TPUs on its Project Suncatcher satellite and got the on-board Gemma 4 model to answer prompts, including a haiku about space. It is a hardware milestone, not proof orbital data centers make sense; radiation and thermal tests come next, with twin satellites planned for 2027.
Four big labs built AI bug hunters and landed on the same rule: the agent that finds a flaw never gets to confirm it. That rule is the real story.
NVIDIA is pitching NemoClaw as the scaffolding around the agent rather than the agent itself, and the real bet is that governed open stacks can handle specialist work at a fraction of frontier cost.
A storage diagnostics tool is not glamorous, but NVIDIA's gds-diag shows a healthy pattern: deterministic checks underneath, an optional agent on top, and nothing changed without a human saying so.
Breakthrough Listen and the SETI Institute let agents steer, calibrate and fix data at the Allen Telescope Array from natural-language requests, with inference on site. NVIDIA says the pipeline made the first real-time AI detection of a pulsar. Agents rewriting processing blocks make provenance logging essential.
Built with Gemini, PlanAssist handles routine checks on householder planning applications and drafts cited reports officers edit. Pilots at Barnet, Camden and Dorset cut assessment time by up to 58%, and over 70% of reports were approved unedited, a figure that cuts both ways.
The filing warns that Anthropic’s technology could pose catastrophic or existential risks and that it may never be sustainably profitable. The candor is real and unusual. But risk bloat can bury what matters, and the poetry sits beside huge compute obligations and a short revenue window.
Microsoft Research builds synthetic enterprise “tenants” with simulated users, emails, meetings and org charts grounded in O*NET, job postings and SEC filings. It’s now Copilot’s first standardized evaluation baseline. The privacy approach is sound; whether fake offices behave like real ones is unproven.
On the Cheeky Pint podcast, Altman sympathized with mathematicians swamped by AI results but said OpenAI won’t slow down, citing about $10 million on Navier-Stokes alone. Releasing faster than the field can verify, while privately choosing what to withhold, pushes the cost onto researchers.
The Wikimedia Foundation backs neither side, arguing scraping and transformative pre-training should be fair use while query-time retrieval that replaces the source should need a license. It’s the most useful line yet drawn in the AI copyright fight, though hard to police.
OpenAI is restricting violence and weapons content, especially for users it predicts are teens, can end chats that show fixation on harming others, and promises a twice-yearly Community Safety Report. The test is whether that report includes error rates and appeals.
A small model trained alongside its own research harness just edged a frontier system on a finance benchmark for a fraction of the price, and the recipe matters more than the scoreboard.