Ephemeris. Issue 024 · Tue 12 May 2026 · Zürich

We are reimagining every role for how a company operates in the agentic AI era.

Cloudflare · "Building for the future" · 07 May 2026 — and inside, five other postcards from that future.

Ephemeris · Issue 024 01 / 6 · Anthropic
AI safety · Evaluations

Neutral, by spec.

Anthropic published its election-safeguards update: a written political-neutrality spec, an internal eval set of 600 partisan prompts, and the numbers. Opus 4.7 hits 95% on the balance metric and 100% on the refusal-of-influence set — measurable impartiality, not vibes.

Spec · Evaluation · 24 Apr 2026
Balance
95%   Opus 4.7 across 600 partisan prompts.
Refusal
100%   on explicit "influence the voter" requests.
Method
Adversarial pairs, both sides of each issue, blind LLM judges.
Spec
Public, versioned, applied uniformly across topics.
Section B Ephemeris · Issue 024 02 / 6 · Vercel
Dev tools · Agentic coding

Six hundred preview deploys, every day.

Superset built an IDE where every developer can run a stable of agents in parallel — each in its own ephemeral sandbox, each producing a preview URL. The math: ~50 engineers × ~12 agents × ~1 deploy each = 600 deploys a day, against a CI that used to ship 30. The post is half blueprint, half ops handbook.

The pattern is not "AI writes code"; it's "developer orchestrates ten attempts at the same task in parallel, then picks the one whose preview behaves right." That only works if every attempt has a real URL, a real database snapshot, and a real cost ceiling.

Superset's stack: Vercel Sandboxes for execution, branch-per-agent for state, a routing layer that gives each agent its own subdomain. Idle agents go to zero. Bad ones get killed by the orchestrator, not by the developer noticing.

Headline lesson: agent IDEs are not editors with chat sidebars. They're CI systems where every PR is a small fleet, and the human's job is to grade outcomes, not to write the diff.

Ephemeris · Issue 024 03 / 6 · Sentry
Observability · Next.js

Vibe code, observed.

A walkthrough wiring a real distributed trace across Next.js routes, Supabase Edge Functions, and Postgres queries. Drops the SDK in three files, propagates context across the wire, then shows the resulting flamegraph for a slow signup. Cheap to adopt this afternoon; valuable the first time you have to debug something that touches a database from inside an Edge runtime.

LayerSpan nameCostStatus
1Next.js · App RouterPOST /api/signup312 msok
2Edge functionauth.verify_email118 msok
3Supabase · PostgresINSERT public.users9 msok
4Edge functionwelcome_email.send2.1 sslow
5Next.js · App Routerresponse.render14 msok
6Trace · trace_idconnects 1 → 5, same rootlinked
Ephemeris · Issue 024 04 / 6 · Import AI
Policy · Research · Forecasting

RSI, regulation, and a brain in a jar.

Jack Clark's three-vector issue: a model of recursive self-improvement against macro-economic growth, a "radical optionality" framework for AI regulation that doesn't pre-commit to a path, and a wetware paper where neurons grown on chips perform pattern recognition. Three things a founder should be tracking even if none of them ship next quarter.

§1
RSI vs. economic growth. A back-of-the-envelope model: if RSI compounds at >2× annually, even a single year of "off" growth blows past every prior tech wave. Inputs are guessable; the curve is not.
§2
Radical optionality. Regulators don't have to pick "ban", "license", or "permit". Build a portfolio of triggered policies — only fire the strict ones when measurable thresholds (compute, capability evals, autonomy tests) tip. Sounds obvious; nobody is doing it.
§3
A neural computer. A paper grows rat neurons on a chip, trains them on a pattern task, and reports above-chance accuracy. Substrate isn't silicon. The frame "AI = matmul" is one possible local maximum.
§4
What to do this week. Re-check your forecasts for a 2-year horizon under "RSI works modestly." Most plans break before then.
Ephemeris · Issue 024 05 / 6 · via @seeallochnaya
Models · Explainer

Read the model with the model.

An interactive, side-by-side walkthrough of the DeepSeek-V4 paper: every figure rebuilt, every routing decision annotated, with the author's notes in the margins. 50 of 90 sections live so far. The author used GPT-5.5/Pro and Gemini 3.1 Pro to research the math, then kept editorial control over every note — a model for how to actually read a frontier paper in 2026.

Annotation · §4.2 MoE Routing

"The shared-experts trick isn't new, but the auxiliary-loss-free balancing is. Once you see the loss term they removed, the routing collapse from V3 makes sense in hindsight — and so does why nobody else tried this."

Sections annotated 50 / 90

Restructure · Q2 2026 06 / 6 · Cloudflare
Industry · Agentic era

We are reimagining every role.

Cloudflare cuts more than 1,100 staff in one stroke and explicitly names the cause: internal AI usage up roughly 600% in three months, "every internal process, team, and role" up for redesign. Generous severance, frank language, no euphemisms. The first frontier-infra company to publicly restructure around agents as a primary lever.

Memo · 07 May 2026Source: Cloudflare blog
Cut
≈ 1,100+ roles across the company.
Stated cause
Internal AI usage up ~6× in three months.
Language
"Reimagining every internal process, team, and role" for the agentic era.
Severance
Above market — vesting acceleration, extended health, outplacement.
Read as
A signal, not an isolated event — others will copy the framing.
End of issue 024 Back to top ↑

That's today.

Six pieces today — half a normal issue, because the day was thin and a tight set beat a padded one. The rubric (AI tools, creative software, dev tools, security, research, anything actionable) preferred postcards from the agentic-era restructure over more launch posts.

Anthropic · News

Vercel · Engineering

Sentry · Engineering

jack-clark.net · Import AI

ikot.blog · via @seeallochnaya

Cloudflare · Blog