Today's word
HEADROOM
Ephemeris.
Four of today's five stories fit more work into a box that didn't get bigger. The fifth measures how much room is left.
Nine tasks in ten never reach for Linux
Cloudflare open-sourced @cloudflare/computer, a runtime that hands every agent its own machine: a durable SQLite-backed filesystem that speaks node:fs, and a single exec() call. Bash runs inside a V8 isolate — translated to JavaScript — and only escalates to a full Linux container, mounted over FUSE, when the work actually demands one. The team's target is keeping that under ten percent. Your agent code calls the same interface either way and never learns which one it got.
# early preview, on npm today
$ npm install @cloudflare/computer
// one interface, two runtimes underneath
await computer.exec("git clone $REPO && npm test")
↳ isolate just-bash → JS ~90% of calls
↳ container full Linux via FUSE on demand
# filesystem is durable, seeded from R2 or source control
computer.fs.writeFile("/src/app.ts", patch)
Same GPU. Twice the context.
Cloudflare dropped the KV cache from BF16 to FP8 and the GLM 5.2 weights from FP8 to INT4, then published the arithmetic. On Kimi K2.6 the context they can hold in memory goes from roughly 686,000 tokens to about 1.37 million, concurrency doubles from 32 requests to 64, and accuracy across GSM8K, ARC, MMLU and MMLU-Pro stays within 0.8 points of baseline.
- Peak throughput
- +41%
- Cost per token
- −30%
- GLM 5.2 checkpoint
- 705 → 421 GB
21.5 gigabytes of dev server, down to two
Next.js 16.3 turns on Turbopack's memory eviction and disk cache by default, and the numbers are the story: Vercel's own dashboard drops about 90% of its dev-server RAM after 50 routes, repeat builds read from cache, and App Router SSR moved to native Node streams. Separately, next dev now writes a version-matched AGENTS.md block pointing at the docs already sitting in your node_modules, which retires the Skills that existed only to feed agents current documentation.
| Measure | Before | After | Delta |
|---|---|---|---|
| Dev RAM · vercel.com dashboard | 21.5 GB | 2 GB | −91% |
| Dev RAM · nextjs.org | 4,600 MB | 840 MB | −82% |
| Build · vercel.com/geist | 30 s | 5.5 s | 5.5× |
| Build · nextjs.org | 21 s | 9.2 s | 2.3× |
| SSR requests under load | baseline | +22% | +22% |
| Dev-to-ready · Rust React Compiler | baseline | −46% warm | experimental |
The turn detector left the audio path
OpenAI wrote up the six months of inference, context-management and media-transport work behind GPT-Live. The voice model is full-duplex — it listens and speaks at the same time — and decides several times a second whether to talk, keep listening, pause, interrupt, or call a tool. When a question needs real reasoning it delegates to a frontier model like GPT-5.5 on a separate asynchronous path, so the conversation keeps moving while the thinking happens elsewhere.
USER — so the thing I keep running into is—
GPT-LIVE — mhmm
USER — —is that it waits for me to finish.
GPT-LIVE — right, that's the turn detector. It's gone.
74 tasks, and the best agents clear a third
The Terminal-Bench team retired their benchmark and shipped Frontier-Bench v0.1 in its place: 74 hard tasks from 100+ contributors, deliberately spread past agentic coding into finance, biology, hardware and media. Top agents land around 34%. It's built as a continuous benchmark — new tasks land each release — so the score is a moving target rather than a leaderboard waiting to be saturated.
- VBA & FreeCAD
- Lean & Coq proofs
- Protein embeddings
- MVCC & WAL recovery
- React front-ends
- Medical claims
- Telecom billing
- Security & ops
That's today.
Five picks for 4 August 2026, chosen for what a senior engineer or founder could act on this week: tools worth adopting, dev and agent infrastructure, and research with a practical kernel.