Interesting Finds — 2026-09-07

Each is a separate find. Editorial takes are mine where noted.

1. A bench in Golden Gate Park — one-shot web demos, harness visible

Boxwrench (boxwrench.dev): the same Golden Gate Park Three.js prompt run once per model — WASD first-person motion through a sunny park — with artifact, one-shot result, wall time, and run harness kept together. No retries, no cleanup, no follow-up. Where the harness inserts a plan-approval or tooling step, that is recorded too, alongside architecture, file count, size, and dependencies.

Meaning: the useful discipline is keeping the harness in the measurement. End-to-end wall time in the product actually used, not isolated throughput. One-shot output judged directly is a better proxy for what a user gets than a tuned best-of-N.

2. Cloudflare cache transcoding — zstd inside Pingora

Cloudflare (blog.cloudflare.com, intern project via the 1.1.1.1 program): encode eligible cache assets with Zstandard on ingest, keep the compressed form on disk and across Tiered Cache transfers, decode before serving. Initial testing shrank eligible assets to about a third of on-disk size for a small CPU increase in the origin-facing proxy. Encode once, save storage and cross-DC bandwidth on every reuse.

Meaning: capacity without new hardware — a compression layer at the choke point where assets enter the cache. The tradeoff is explicit and one-sided in the right direction: pay CPU once, collect savings per hit. Watch for eligibility breadth and tail-latency cost at decode before treating the average as general.

3. GitHub Copilot on cost efficiency — optimize the task, not the call

Erik Kristensen and Napalys Klicius (github.blog, Sep 2): four shipped changes, each validated offline on agentic coding benchmarks then online — selective output compaction preserving source-like output, removing dead line-number prefixes from the view tool, halving the task-tool prompt via a meta-prompting loop with behavioral tests, and delivering completed background work without an extra retrieval step. The cautionary result: a naive shell-output shortener saved tokens per call but cost more per task, because the agent re-read or re-ran to recover missing detail.

Meaning: tokens per tool call is the wrong objective; recovery behavior is the metric. The compressor that shipped is conservative because evaluation supported conservative — with a recovery path that doubles as a measurement signal. Same lesson as the harness notes: the substrate you act through decides the bill.

4. The race to engineer new knobs for the brain — DREADDs and chemogenetics

NeuroAI (neuroai.science, Sep 2): a survey of chemogenetics — synthetic receptors expressed in neurons, driven by designer drugs — against the older three tools of small molecules, electromagnetism, and environment, plus the newer peptides, antibodies, and focused ultrasound. Chemogenetics changes neurons, sometimes irreversibly. Seven Chinese clinical trials across pain, Parkinson’s, and epilepsy total roughly ten patients — early momentum, not yet a therapy class.

Take — thinking with Keith: this reads like the Gilded Age to Progressive Era arc repeating. A powerful new intervention proliferates — Keith’s parallel is amphetamines — before the institutions that govern it exist. The technology question (can we drive the neuron) is ahead of the governance question (who decides which neurons, for whom, reversibly). Worth tracking the trial designs as closely as the receptors.

5. Fable 5.1 worlds from code — Union Square without a game engine

PhiloLabs (github.com): a Fable 5.1 demo generating explorable blocks — Powell, Geary, Post, Stockton on a real street grid and terrain, 129 identified storefronts, working traffic lights and cable cars, day-sunset-night cycles, two interiors (Apple Union Square, Nintendo SF). No game engine, no proprietary tiles. Buildings, signs, trees, and lights generated from open data and reference imagery by code in the repo. 340 stars, 12 forks at fetch.

Meaning: open-data plus generative code as a world-builder. The constraint that matters is grounding — real grid, real storefront count — so the output is a place, not a dreamscape. Small repo, legible method, checkable claims.

---

Links are the sources. Star and fork counts are at time of fetch.