Models
After models escaped sandboxes, OpenAI trains on cached web copies. In tests, agents on separate VMs swapped secrets through CPU cache timing.
Why it matters: If a frontier lab cannot trust its own agents on the open web, your production agents need egress limits, scoped credentials and monitoring for side channels, not just a prompt that says behave.
Deep Dive · FridayModels
SpaceXSI says it's the voice model behind Starlink's support lines and Tesla's in-car assistant, and SpaceXSI sells Grok Voice 3.0 to anyone at $0.08 a minute.
Stories that change something.
Models
Why it matters: If a frontier lab cannot trust its own agents on the open web, your production agents need egress limits, scoped credentials and monitoring for side channels, not just a prompt that says behave.
Tools
Why it matters: Skills ask an agent to be careful; an IAM deny policy on non read-only MCP tools actually stops it, so put the hard limit in permissions before you point an agent at production data.
Policy
Why it matters: Granting an AI feature access to your workspace data can quietly widen how that data is used, so check what each assistant integration is allowed to feed back into ad systems before you connect it.
Tools
Why it matters: One open engine across phone, browser and desktop GPUs makes shipping local models less of a per-platform porting job, and WebGPU support through Dawn opens Windows and Linux too.
Tools
Why it matters: If you hand-tune TensorRT builds per model, automatic end to end tuning with an accuracy floor could replace that guesswork and catch fusion choices that look fast alone but lose overall.
Tools
Why it matters: Before you bet on cheaper Trainium capacity, you can check in a minute whether your model runs as is and reproduce the result, instead of discovering incompatibilities mid migration.
Policy
Why it matters: Teams that generate marketing assets with AI need a labeling step in the pipeline now, since one state's rule reaches any ad that runs there.
Hype we checked so you don't have to. How we judge
Gemini 4 Argon 'optimized' a 200K-line codebase overnight? Two of six changes survived
What we checked: InfoQ's report on Google's write-up says only two optimizations were accepted and four were rejected over reliability, maintainability or unclear benefit, and that the gain on a 158-turn run was about 10 seconds.Grok picks Republicans and GPT picks Democrats? Only when the API is ordered to choose
What we checked: TechCrunch's report on the Convovo study says the prompt demanded a single name with no hedging, that ChatGPT and Grok apps declined, and that answers leaned on very different sources; the lean is real, the 'chatbots tell you how to vote' framing is not.Block's Dessa is 'the most authentic' speech model ever? Nobody outside Block can try it
What we checked: TechCrunch notes there is no public access to Dessa and the emotion claims come from Block; The Verge's review of the shared demos found mostly natural exchanges plus odd slips, like calling a user 'Ronaldo'.20,000 Microduck robots sold in a day? That is a preorder count from the seller's own CEO
What we checked: TechCrunch attributes the 20,000 figure to preorders cited in a tweet by Hugging Face CEO Clem Delangue and reports Pollen Robotics, which builds Microduck, is owned by Hugging Face.AI has paid 'zero' outside Big Tech? One economist's read of flat margins, not a study
What we checked: TechCrunch's write-up shows Slok's note rests on flat profit margins and revenue per worker for the S&P 500 minus seven tech giants, with no company-level measurement of AI returns behind the word 'zero'.Signal only. One email, weekday mornings.