Ephemeris.
Issue 038 · A morning brief from Zürich · Six files
“There must be people outside those incentives.”
Chris Olah · Anthropic · 25.V.2026
Two hundred and seventy gigabytes became five.
PostHog moved local feature-flag evaluation from Django to Rust. The new service runs on five pods instead of thirty, costs twenty-four times less CPU, and answered the median request in four milliseconds. The rollout happened in three weighted steps over two days.
What broke at fifty million actions a day.
Cursor's first work-stealing scheduler topped out at 90% uptime. Rebuilding on Temporal pushed reliability past 99% across seven million workflows. The bigger lesson, Josh Ma writes, is that environment setup is the silent quality lever no one watches.
- TO
- Anyone who has a CI step they'd rather not write
- FROM
- Roshan Sadanani · Cursor
- RE
- @cursor/sdk · v1 · 29.IV.2026
Build your own coding agent over coffee.
The Cursor SDK is a TypeScript package that exposes Agent.create() and agent.send() with configurable models, sandboxed execution, and persistent state. Teams are already wiring it into CI pipelines, internal review bots, and customer-facing flows.
// summarise a PR diff and post the review back import { Agent } from "@cursor/sdk"; const a = await Agent.create({ model: "composer-2.5", cwd: process.env.GITHUB_WORKSPACE, }); const reply = await a.send( "Read the diff, flag risky changes, post a CR comment." ); console.log(reply.text);
Skip seventy percent of attention. Lose nothing.
BLASST drops attention blocks whose softmax score falls below a single scalar threshold. No retraining, no architecture change, no special hardware — just a kernel-level shortcut that earned the MLSys '26 best-paper award.
The technique uses online softmax statistics to identify attention scores that contribute nothing to the output, then skips the corresponding block — both the softmax and the matrix multiply.
At 71.9% block sparsity the authors report a 1.52× speedup on prefill; at 73.2% sparsity, 1.48× on decode. Benchmark accuracy on long-context tasks degrades by less than one percent.
Threshold calibration is automatic and reveals an inverse relationship to context length: the longer the prompt, the more attention you can skip. The kernel is a drop-in for existing inference stacks.
Anthropic answered the Pope.
Pope Leo XIV's encyclical Magnifica humanitas arrived on 15 May. Ten days later Chris Olah, Anthropic's co-founder, conceded that AI labs work inside incentives that can pull against doing the right thing — and named three places the Church should weigh in.
- I. That AI's benefits reach the people whose labour it is most likely to displace — starting with the global poor.
- II. That moral imagination keeps pace with capability — what does human flourishing look like when work itself is reshaped?
- III. That somebody investigates the mysterious internal states of these models — Olah's word — and what they may amount to.
Four questions Anthropic now has a building for.
The Anthropic Institute will publish its own research and data, run a four-month fellowship, and hand its findings to outsiders so they can argue back. The agenda splits four ways.
That's today.
A morning brief for engineers and founders. Six files, four sources, one rule: useful by tomorrow.