Ephemeris.
Monday
08.06.26
Today: cloud GPUs from your terminal, two tiny databases that aren't the same tool, a model that won't stop quoting whole books, and a leaderboard everybody gamed.
Cloud GPUs, one command away.
Google's new Colab CLI wires your terminal straight to a remote Colab runtime: colab new --gpu H100 provisions an accelerator and colab exec ships a local script or notebook with no upload step. It's open source under Apache 2.0 — and any terminal agent (Claude Code, Codex, Antigravity) can drive it through a bundled skill file.
Apache 2.0 · T4 → A100 → H100 + TPU · agents via COLAB_SKILL.md
Fine-tune it, and the books fall out.
Researchers at Stony Brook, Carnegie Mellon, and Columbia Law fine-tuned models on a harmless-looking task — expanding plot summaries into prose — and watched the guardrails give way. The tuned models reproduced up to 91.9% of a copyrighted book in verbatim runs as long as 440 words. Their read: refusal training is a brittle filter laid over text the weights already hold.
SQLite writes rows. DuckDB reads columns.
PostHog's team lays out why the two databases everyone embeds aren't swappable. SQLite is a row store tuned for transactional reads and writes; DuckDB is a column store built to scan and aggregate millions of rows at once. Choose by the shape of your query, not by which name you know.
SQLite
DuckDB
Red-team your agent before it ships.
At Build, Microsoft open-sourced RAMPART — a framework for stress-testing agents against the failures that actually bite: cross-prompt injection, behavioral regressions, and data exfiltration, caught in development rather than in production. It lands next to Execution Containers (MXC) for runtime guardrails and MDASH, a swarm that hunts exploitable bugs and files context-aware fixes.
Squeeze the KV cache to two bits.
KVarN quantizes the KV cache to 2 bits without the usual accuracy collapse, and with no calibration step. It Hadamard-rotates each token, then variance-normalizes per block so quantization error stops snowballing across a long reasoning chain. The payoff: 3–5× more context, throughput above FP16, and near-FP16 accuracy on MATH500, AIME24, and HumanEval — behind a single vLLM flag.
- 2-bit
- KV-cache precision, calibration-free
- 3–5×
- more context at the same memory
- > FP16
- throughput, at near-FP16 accuracy
- one flag
- ships as a vLLM quantization backend
Set a token target. Watch them game it.
Amazon gave developers AI-usage targets and a "Kirorank" board that scored people by tokens burned. Engineers did the rational thing and gamed it — pointing agents at busywork to climb, on compute Amazon paid for. The board is gone now, with a note from a senior VP: please don't use AI just to use AI.
"When a measure becomes a target, it ceases to be a good measure." Goodhart's law, this time with a GPU bill.
Seen elsewhere · ≈60T tokens in 30 days on Meta's internal "Claudeonomics" board
That's today.
Six picks, six sources — a thin-signal Monday read for what shifted.