Research
Its red team ran a 26B model across 15 cheap Android TV boxes. Only GLM-5.2 finished full replication in tests, but sub-30B models handle most steps.
Why it matters: If weights become the malware payload, IoT hygiene stops being a consumer problem and becomes a containment problem for every network team.
Deep Dive · FridayChips
Microsoft rebuilt its inference chip around moving fewer bytes.
Stories that change something.
Research
Why it matters: If weights become the malware payload, IoT hygiene stops being a consumer problem and becomes a containment problem for every network team.
Research
Why it matters: If guardrails regress as image models improve, any product shipping these APIs needs its own output filtering, not just the vendor's.
Policy
Why it matters: A multistate suit over minors' data puts real legal pressure on how chatbots are marketed and what they collect from young users.
Tools
Why it matters: An API that audits design docs and proofs for internal consistency is a cheap new gate to wire into CI before humans spend review time.
Tools
Why it matters: Design systems written as machine-readable specs are how teams keep coding agents from inventing their own UI, and Figma wants to own that file.
Models
Why it matters: Voice agents just got smarter and faster at the same price, but the support-call wins are SpaceXSI's own numbers from Tesla and Starlink, not yours.
Research
Why it matters: Correcting existing models instead of training giant new ones puts better subseasonal forecasts within reach of agencies with no GPU budget.
Hype we checked so you don't have to. How we judge
Epic's vibe-coded launcher beats Steam by seconds, and is missing features
What we checked: Ars Technica tested it: cold start beat Steam by a few seconds, the 0.06s tray gap was unnoticeable, and Epic calls the Preview light on features.Flock's CEO says its mics hear 'just gunshots.' Flock's own site lists more
What we checked: Flock Safety's gunshot detection page lists sideshow, crash and fireworks alerts; TechCrunch quotes a Flock spokesperson confirming them.SentinelOne's 'weeks to hours' malware win rests on one executive's estimate
What we checked: OpenAI's SentinelOne post: the 4 to 6 weeks versus 8 hours figure is a quote from SentinelOne's AI chief, with no published method or accuracy data.'Google releases MedGemma' is a 2025 launch wearing a new journal cover
What we checked: Google Research's July 9, 2025 MedGemma post announced the same 4B and 27B models and MedSigLIP; AI Weekly's Oct 14 alert covers the paper.Nothing in Chips today.
Signal only. One email, weekday mornings.