Research
Anthropic research, via TechCrunch: a per-model fingerprint would let anyone check outputs locally. The code is on GitHub; it is not deployed.
Why it matters: Today you trust the API to serve the model you pay for. If provable inference matures, silent model swaps and vague claims that a frontier model reviewed your code become checkable, not a matter of faith.
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.
Research
Why it matters: Today you trust the API to serve the model you pay for. If provable inference matures, silent model swaps and vague claims that a frontier model reviewed your code become checkable, not a matter of faith.
Tools
Why it matters: If you ship an iOS app, voice and agent requests can now land in it without users naming your app. Expose clean actions now or Siri will route around you.
Tools
Why it matters: Separating the one bad NVLink from a wall of downstream NCCL errors is most of the pain in cluster debugging. A rules based, reproducible diagnosis beats grepping rank logs at 3am.
Policy
Why it matters: Google is telling the biggest regulated user base that Search is an AI product. Anyone building on the assumption of a plain search fallback should take note.
Tools
Why it matters: Agents that hit enterprise APIs as a shared service account break audit trails and least privilege. This is the pattern to copy before security review blocks your agent.
Research
Why it matters: Long context scores mostly measure lookup. State tracking is what breaks when an agent reads a big codebase, so this is a better probe for coding agent reliability.
Research
Why it matters: Vendors keep claiming cyber skills for their models. A deterministic public set lets you test the model you actually use on fix without regressions work.
Policy
Why it matters: If your chatbot app can drift into romance, it now falls under the adult rating. Check your descriptors before the deadline or risk review trouble.
Hype we checked so you don't have to. How we judge
ChatGPT is catching up to Google in search? OpenAI's math can't be checked
What we checked: The Verge reports OpenAI's Adam Fry announced the 20 percent figure at a press event, counting any query that triggers a web search, and did not specify the source of the Google search volume number. No independent measurement was cited.Kimi K3 is a frontier coding model? That label is Moonshot's own
What we checked: The AWS Machine Learning blog post introducing India based inference for Kimi K3 says Moonshot AI reported the model achieved frontier level coding and general agentic intelligence at release. The post attributes the claim to Moonshot and cites no outside leaderboard.Garry Tan runs AI agents nonstop? The only evidence is his own rough guess
What we checked: TechCrunch quotes Tan on Isaacson's podcast describing heavy use of AI agents and estimating the token cost at maybe $100,000 to $200,000 a year. It is his own wide-ranging guess about his usage, and the article cites no usage records.Signal only. One email, weekday mornings.