Today's mood
Plumbing.
Today's issue lives below the floorboards: boot firmware, query partitions, observability pipes, and the silicon now small enough to sit on your desk.
Four hours to reboot a server. Now it's minutes.
After a firmware update, Cloudflare's core units took up to four hours to come back. The slow path was UEFI and iPXE sitting through timeouts that never needed to fire. Delete the waits and reboots drop to minutes — a reminder that if you run bare metal, the boot path itself deserves a profiler, not just the OS on top of it.
An agent read the query plan and found a three-year-old bug
PostHog pointed an autoresearch agent at ClickHouse and it surfaced a defect that had hidden since 2023: timezone-wrapped timestamp comparisons that blocked partition pruning, so every affected query scanned far more data than it needed. Remove the wrapper and granule scans fall 62%. The signal worth keeping: agents are now good enough to audit the slow paths humans stopped looking at.
You shipped the LLM feature. Now watch it.
A day-one observability kit for small teams who just put their first AI feature in front of users — no platform team required. Three setups, in the order they'll save you:
- Trace every model call. Inputs, outputs, latency, the prompt version — so a bad answer is a row you can open, not a story a user tells you.
- Track cost with alerts. Per-feature token spend with a threshold, so a runaway loop pages you before it pages finance.
- Capture the LLM-specific failures. Tool-call errors and token-limit blowups don't throw like normal exceptions; instrument them on purpose.
The AI economy you can't see is growing 2,600% a year
Jack Clark's latest issue rounds up the parts of the field that never make the homepage — and they're where the curve is steepest.
Invisible economy
A machine-to-machine AI economy — agents transacting with agents — compounding at roughly 2,600% a year, mostly off the charts the rest of us read.
Protein-folding scaling
Fresh scaling laws for protein-folding models: bigger and more data still buys accuracy, with a practical map of where the returns bend.
Pricing the risk
A sober case that automated alignment research is harder than the optimists claim, plus how economists might price extinction risk into policy.
Codex and the frontier models now run inside AWS
OpenAI's frontier models and the Codex coding agent are generally available on AWS. For teams already committed to AWS, that's a real change of options: run GPT-5.5 and Codex against your own accounts, data boundaries, and spend commitments instead of routing everything through a separate vendor. Worth a fresh look if procurement or data residency had kept OpenAI off the table.
A 120-billion-parameter model, on the desk
NVIDIA's RTX Spark pairs a Blackwell GPU with a 20-core Grace CPU and up to 128GB of unified memory — enough to run a 120B-parameter model locally at a million-token context. Built with Microsoft for secure on-device agents, it ships this fall in laptops and mini-desktops from the usual makers. Real local inference stops being a server-room privilege.
- 128 GB
- unified memory
- 1 M
- token context
- 1 PFLOP
- AI performance
That's today.
Six picks for Tuesday, 02 June 2026: a four-hour reboot cut to minutes, an agent that found a bug humans missed for three years, a day-one playbook for watching your LLM feature, the corners of the AI economy that don't trend, OpenAI's models landing inside AWS, and a desktop that runs a 120B model on its own.
Sources today: Cloudflare · PostHog · Import AI (Jack Clark) · OpenAI · NVIDIA, via Сиолошная (@seeallochnaya). Rubric: tools you could adopt this week, dev tooling and agentic coding, research with a working kernel, and hardware that shifts what's possible.