Ephemeris · Issue 100 Thursday · 30 Jul 2026 · Zürich
Ephemeris · Issue 100 01 / 05 · OpenAI
Agents · API

Two settings you already have. Three times the score.

The official ARC-AGI-3 harness kept GPT-5.6 Sol's moves and threw away its reasoning after each one, then dropped the oldest actions once history overran the window. The model remembered what it had done and forgot why. OpenAI re-ran the benchmark through the Responses API with retained reasoning and compaction switched on — the same two settings behind ChatGPT and Codex — and the public-set score went from 13.3% to 38.3% on six times fewer output tokens.

13.3% Official harness, public set
38.3% Both settings enabled
6× fewer Output tokens spent
Ephemeris · Issue 100 02 / 05 · Cloudflare
Security · Post-quantum

The last hop finally signs with lattices.

Cloudflare's edge has spoken post-quantum to browsers for two years, while the leg from the edge to your own servers still authenticated with RSA and ECDSA. Authenticated Origin Pulls and the Custom Origin Trust Store now take ML-DSA certificates, so the credential proving who is calling your origin survives a harvest-now-decrypt-later adversary. Generate with OpenSSL 3.5.0 in seed-only format, upload per zone or per hostname, teach NGINX to present or verify it, and drop trust in the classical path so nothing downgrades.

SignatureML-DSA-44 · 65 · 87 (FIPS 204)
Key agreementX25519MLKEM768
SurfacesAuthenticated Origin Pulls · Custom Origin Trust Store
PrerequisiteOpenSSL 3.5.0+ · SSL/TLS mode Full (strict)
Ephemeris · Issue 100 03 / 05 · GitHub
Supply chain · Config

RE: the ninety-two bump commits nobody asked for.

Bruno Borges went through GCToolkit and found 92 of its 578 commits were Dependabot bumps. His fix is a few lines of .github/dependabot.yml: one named group matching "*", a monthly interval instead of daily, and a separate updates entry per ecosystem. Security advisories keep their own trigger and still land the day they are published — the routine churn collapses into one predictable pull request per ecosystem, and the default three-day cooldown holds back the rest.

To
Everyone drowning in bump PRs
Re
.github/dependabot.yml
CC
Security advisories — unaffected, still immediate
updates:
  - package-ecosystem: "maven"
    directory: "/"
    schedule:
      interval: "monthly"
    groups:
      monthly-batch:
        patterns:
          - "*"
Ephemeris · Issue 100 04 / 05 · Black Forest Labs
Generative media

One set of weights for image, video, sound and a robot arm

Black Forest Labs trained FLUX 3 on images, video and audio at the same time, then extended the same backbone to predict robot actions with its partner mimic. It generates up to twenty seconds of video carrying native synchronised audio and multilingual dialogue, driven by text, a reference image, or another video. Video is in early access now, images follow in the coming weeks, and FLUX 3 Dev — the open-weight backbone — is promised later this year.

00:00–00:20 Video length per generation, audio generated with it
720P / 10S Resolution and clip length used in the evaluations
52–93% Preference rate against rivals, marked preliminary
FLUX 3 DEV Open-weight release, timing still unnamed
Ephemeris · Issue 100 05 / 05 · Sentry
Game dev · Observability

The crash report now arrives with the footage.

Sentry's Unreal Engine SDK 1.16.0 adds Session Replay, so a crash report carries video of the moments that produced it instead of a stack trace and a guess about what the player was doing.

A stack trace tells you where the game died. The seconds before it tell you what the player was doing.

A hook in Slate's rendering pipeline grabs the backbuffer at 30 fps — configurable — and hands frames to a hardware H.264 encoder running on its own thread. The right encoder is chosen at runtime from the player's GPU: NVENC, AMF or VideoToolbox.

A replay recorder keeps a rolling buffer of fragments, so what attaches to the crash is the window immediately before it rather than the whole session.

The cost is small enough to argue for. At 4K the render-thread hook measured about 0.20 ms per frame, roughly 1.2% of a 60 fps budget, because encoding and file rotation both happen off the render thread.

Enabling it takes three steps: turn on the AVCodecsCore plugin plus your vendor codec plugin, toggle Session Replay in Project Settings, then rebuild so replay support compiles in.

0.20ms / frameRender-thread hook at 4K
1.2%Of a 60 fps frame budget
30fpsDefault capture rate
1.16.0Unreal SDK version
End of issue 100 · 30 July 2026 Back to top ↑

That's today.

Issue 100. Five picks from four newsletters, seven engineering blogs, two model labs and five Telegram channels, read this morning in Zürich. The rubric, unchanged since issue 001: tools you could adopt this week, creative software, dev tooling and agentic coding, privacy and security when the signal is high, research with a practical kernel. Every item published within the last thirty days; nothing here has run before.

OpenAI Cloudflare GitHub Black Forest Labs Sentry The Batch The Zvi Import AI Jeff Bullas Anthropic Vercel PostHog Fly.io Next.js @seeallochnaya @TochkiNadAI @denissexy @rvnikita_blog @ProductsAndStartups