Tools
The open-source release means moving a job off GPUs mostly takes swapping cuda for tpu and nccl for tpu_dist. Eager mode is first class, unlike PyTorch/XLA.
Why it matters: TPUs stop being a JAX only option. Teams with PyTorch code can now price TPU capacity against GPUs without a rewrite, though Google's own speed figures still need outside benchmarks.
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.
Tools
Why it matters: TPUs stop being a JAX only option. Teams with PyTorch code can now price TPU capacity against GPUs without a rewrite, though Google's own speed figures still need outside benchmarks.
Tools
Why it matters: If you lean on Codex for coding, heavy ChatGPT use now eats the same budget, and OpenAI warns some users will hit resets more often. Plan your limits as one pool.
Tools
Why it matters: This is the reference design for MCP at scale: registry as control plane, proxy for runtime, per tool auth and PII redaction. Copy the pattern before your tool list explodes.
Tools
Why it matters: Silent prompt changes alter what your coding agent will do. If a legit project trips the new rule, Anthropic says to report it, but expect more policy tuning inside the tool itself.
Tools
Why it matters: Anyone shipping model code has the same gap between fast unit tests and costly GPU runs. Shrunk real checkpoints plus recorded outputs is a pattern you can adopt in your own CI.
Policy
Why it matters: If your product pulls grounding data from a third party search API, the provenance of that index is now a legal question, not just a quality one.
Tools
Why it matters: Agent subscriptions are still being priced in real time. Anything you automate on a flat plan can stall when the vendor quietly retunes limits, so keep a fallback.
Hype we checked so you don't have to. How we judge
Anthropic is on the verge of self building AI? Its filing says no current model qualifies
What we checked: TechCrunch quotes the revised S-1/A: internal estimates about future models suggest Anthropic may imminently meet its Responsible Scaling Policy definition of dramatically accelerated progress, but it has not designated any current model as exceeding that threshold and plans to keep humans in control of model development.An AI that predicts human behavior with 98% accuracy? Not per decision, it turns out
What we checked: TechCrunch reports CEO Amit Gal-Or clarified the 98% figure measures whether trading flows fall within a predicted range, not whether each trade was predicted, cites an F1 of 0.8 to 0.85, and notes the $62.5M seed came from the founders and Infinity Group, founded by the CEO's father.OpenAI's data centers will protect the grid? So far it is one pilot at an unbuilt site
What we checked: The Verge reports OpenAI will pilot the program at an upcoming data center in Effingham County, Georgia, promising to free up to 1,000 megawatts during extreme weather or emergencies, framed against spreading local opposition to data centers. It is a commitment for a future site, not an operating result.Agents cut incident triage by 75%? That figure comes from Circle K's own metrics
What we checked: The AWS Machine Learning blog post, co written with Circle K's Marek Klatowicz, states mean triage dropped to under 15 minutes, a 67 to 75 percent reduction based on the team's internal incident response metrics. No independent data or methodology is published.Signal only. One email, weekday mornings.