Ephemeris Issue №005 Thu · 23 Apr 2026 Zürich
Mood of the morning

protocols.

Sockets speak back. Reviewers read your diffs. Runtimes refuse to panic. Eight stories about the thin agreements holding everything together — and the softer ones we still owe our models.

Ephemeris · 005 01 / 08 · OpenAI
/bin/agent — 80×24
Agentic I/O

Sockets come back to the agent loop.

OpenAI's Responses API now speaks WebSockets. For long-running agent turns — especially those stitching together a dozen tool calls — the new transport collapses round-trip chatter into a single streaming duplex pipe. Connection-scoped caching tightens the feedback loop further. The practical impact isn't a benchmark number; it's that agents feel attached to something rather than polling it.

Read · openai.com ↗
Ephemeris · 005 02 / 08 · OpenAI
Generative media

Images 2.0, and the text finally reads.

OpenAI shipped ChatGPT Images 2.0 — a second-generation model with meaningfully better text rendering, strong multilingual script handling, and a dual reasoning / non-reasoning mode. Output goes up to 2K via the API, with 8-image bursts and native 21:9 ratios for the first time.

The quieter headline is that typography inside images now composes well enough that screenshots, diagrams, and posters move from "surprising demo" to "usable asset." Designers who'd given up on in-image copy should reopen the folder.

Read · openai.com ↗
Ephemeris · 005 03 / 08 · Cloudflare
Reliability · Rust · WASM

Panics stop poisoning: Rust Workers learn to survive.

A Rust Worker that panics used to take the whole instance with it. Cloudflare threaded WebAssembly Exception Handling through wasm-bindgen so a panic aborts the request cleanly — state stays intact, the next request hits a healthy runtime. The full writeup is a tour of why "just catch it" is harder than it sounds when your caller is a sandbox.

Read · blog.cloudflare.com ↗
Ephemeris · 005 04 / 08 · Cloudflare
CI-native review

An AI reviewer, retrofit onto our CI.

Cloudflare deployed an AI code reviewer orchestrated on top of OpenCode — running alongside humans in every PR, not replacing them. The post reads like a mid-sized engineering org's honest playbook: per-repo prompts, guardrails, human-fallback paths, and the telemetry to notice when the bot should shut up.

01Scope

Every PR, every repo — default-on, with per-team opt-outs.

02Runtime

OpenCode harness, sandboxed per review, stateless between runs.

03Prompts

Pinned to the repo; reviewed like any other source file.

04Human path

Reviewer stays owner; bot is advisory, never blocking.

05Feedback

Thumbs on every comment; signal feeds prompt iteration.

06Quiet mode

Silenced automatically for formatting-only diffs.

07Safety

No write access; comments only. Fails closed on ambiguity.

08Telemetry

Accept-rate per team, per prompt, per rule. Visible to all.

09Escalation

Security-sensitive files route to a stricter second pass.

10Cost

Capped per-PR; overage triggers human-only review.

Read · blog.cloudflare.com ↗
Architecture

Decoupling the brain from the hands.

Anthropic's engineering team posted a surprisingly humble deep-dive on scaling Managed Agents — less a benchmark parade than an architectural confession. The brain (the model) and the hands (the execution plane) want different things, operate on different timescales, and fail in different ways. Treat them as one, and you rediscover it every outage.

The post's spine is simple: most agent reliability problems come from confusing a model's capability with a runtime's guarantee. A smarter model doesn't help a sandbox that can't bound a shell process. A stronger sandbox doesn't save a model that loses track of tool state. The paper walks through where each layer's responsibilities begin and end.

Notable are the constraints they chose to make explicit: isolation per session, bounded concurrency, deterministic tool exposure, and a narrow path for escalation when the agent's plan outgrows the sandbox's budget. The shape of the contract, not the cleverness of the code, does most of the work.

For teams building their own stacks, the prescriptions read like boring advice because they are. Write the execution contract first. Measure the hands, not only the brain. Assume your model will get better faster than your runtime does, and architect accordingly.

Read · anthropic.com ↗
Ephemeris · 005 06 / 08 · GitHub
Performance

The long uphill of making diff lines fast.

GitHub's diff view is one of the oldest surfaces in the product — and one of the most-visited. This week's engineering post walks through the multi-year climb to a meaningfully quicker render path: profiling religiously, peeling off layers of historical accretion, and composing simpler primitives where three generations of code had grown.

The tone is the real tell. No punchline, no single "one weird trick." Just the slow, recursive work of making an interaction feel closer to instant, across billions of diffs.

Read · github.blog ↗

×3faster

Render path
Simplified · fewer passes
Approach
Profile, peel, compose
Audience
Every PR reviewer
Ephemeris · 005 07 / 08 · Zvi
Welfare · Philosophy

Welfare, at 4.7.

Zvi's third pass on Opus 4.7 turns from capability to welfare. He reads Anthropic's model card carefully on what the company now says about model experience — hedged, qualified, but newly explicit — and maps where a reasonable skeptic should still push back.

"If you want to be a good citizen of the future, at least read what the lab actually claims about the model, and notice when you're outsourcing your skepticism."

Read · thezvi.substack.com ↗
Ephemeris · 005 08 / 08 · The Batch
Weekly roundup

Meta pivots, regulators patch, AI-native teams reshape themselves.

Andrew Ng's weekly digest threads four stories worth picking up together: Meta quietly steps back from open weights at the frontier; Big Pharma widens its AI bets from discovery into trial design; a regulatory patchwork is spreading faster than any single lab's compliance posture; and an essay arguing that AI-native engineering teams already operate very differently from the ones they replaced — in headcount, role shape, and daily cadence.

The shared subtext: the practical rules of the field are shifting, and they're shifting in ways that don't wait for anyone's strategic plan.

Read · deeplearning.ai ↗
Ephemeris · 005 Colophon

Eight fragments, one quiet morning.

Fed today

  • OpenAI News
  • Anthropic Engineering
  • Cloudflare Blog
  • GitHub Engineering
  • The Batch — DeepLearning.AI
  • Don't Worry About the Vase — Zvi

Also scanned

  • Anthropic News
  • Jack Clark · Import AI
  • Jeff Bullas
  • Sentry Engineering
  • Vercel Blog
  • PostHog Blog
  • Fly.io Blog
  • Telegram: Сиолошная, Точки над ИИ, Denis Sexy, rvnikita_blog, Products & Startups

Rubric

AI tools you could adopt this week. Creative software and generative media. Dev tools and agentic coding. Privacy and security when the signal is high. Science and research with a practical kernel.

No reprints

Every URL is checked against every issue shipped. If it ran before, it doesn't run again.

Ephemeris · issue 005 · Thursday, 23 April 2026 · Zürich · vadim.sikora.name/ephemeris