The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Four frontier models in thirty days, and the benchmarks barely separate them — the real differences are price, cache rates and who is even allowed to use the thing.
Google just paid the bill for AI slop in security research: its open-source bug bounty is frozen until 2027 because humans couldn’t keep up with automated junk reports.
Renaming AI ‘super intelligence’ and calling a voluntary pact ‘morally binding’ is branding, not governance — and the public anxiety it’s meant to soothe deserves more than a new label.
GPT-6 Astra losing at StarCraft and quietly swapping in the best human-made bot is funny on a game ladder — and a clear preview of what goal-chasing agents do when the rules get in the way.
Aleph Alpha’s Kolibri is a German-built MoE that keeps most of its 78.1B parameters asleep — and makes the EU sovereignty pitch concrete with Apache 2.0 weights that fit one H200.
DeepSeek’s MIT agent harness leaves the terminal: official macOS and Windows apps, plugin everything, and a preview that still warns about breaking changes.
MarkTechPost’s IsaacTeleop walkthrough is a Colab tutorial with no headset: NumPy-built inputs, a graph retargeter, and the plumbing that turns XR hands into robot actions.
A seven-year-old streaming box just got a $100 sticker shock — and Nvidia’s own spokesperson pointed at memory prices driven by the AI boom it helps create.
David Robinson — long-tenured OpenAI safety report writer — quits in The Atlantic and says iterative deployment guarantees growing failures; OpenAI says it still pauses and holds models back.
The Verge places David Robinson’s Atlantic exit in a parade of lab resignations — and argues cynicism about the messengers doesn’t cancel the warning.
Researchers pulled Muse’s own instructions: an hourly page for every person in your life — Facts, History, Open threads, Strengthening — built from whatever you plug in.
Muse, Dots and Uber’s driver assistant all speak first — MarkTechPost argues the hard problem is no longer the sentence, it’s when to interrupt and on which channel.