Ephemeris.
Eighteen feeds read, ten stories kept. Both frontier labs shipped a model this week — and underneath that, everyone else spent it handing agents real work: a bug backlog, a brand guide, a machine of their own.
Three weeks, 1,462 issues, and the agent could not close one.
The Next.js backlog peaked at 3,109 open reports. Rather than closing by inactivity — which had already thrown out real bugs — the team built closability, a research agent that reproduces each issue in a fresh sandbox against its reported version, latest stable, and canary, then hunts for evidence against its own conclusion. It is read-only outside that sandbox: it cannot comment, close, push, or deploy. Maintainers read the evidence and made every call.
| Already fixed | 543 · 37% |
| Duplicate | 278 · 19% |
| Expected behaviour | 237 · 16% |
| No longer reproducible | 89 · 6% |
| Unsupported or obsolete | 66 · 5% |
| Other | 249 · 17% |
Publish the brand as one file the agent can fetch.
Vercel's design.md sits at a public URL so any coding agent — inside their codebase or not — can read how a Vercel page is supposed to be built. It ships with a public stylesheet that constrains type and spacing, and an eval harness of seven fixed scenarios. Across 200+ runs, pages built with it carried 39 mechanical failures against 91 without. The file, the stylesheet and the agent template are all published.
- 01The guidance file: how to frame a reader's task, structure evidence, choose a page composition.
- 02A public stylesheet of classes and tokens, so the agent picks from a set instead of inventing values.
- 03Seven fixed scenarios — renewal proposals, perf reports, benchmarks — scored by humans and by critique.
- 04The honest caveat: every output still had at least one failure serious enough to block shipping.
A minute of 1440p, and the camera goes where you point it.
Atlas is a multimodal autoregressive diffusion transformer pretrained from scratch on text, images, video and 3D together. It generates novel views with pixel-perfect camera control, reconstructs real scenes from sparse photographs, and emits both 2D frames and explicit 3D — point clouds and Gaussian splats — rather than only pixels. World Labs claims it beats recent video models on camera control and the best open reconstruction models on their own ground. Early access is by request.
$40 million, and the model stays in your building.
Thomson Reuters built a domain model family on Qwen and skipped fine-tuning as the whole answer.
200 billion tokens, curated down from a 19-trillion-token candidate pool.
The recipe is continual learning — mid-training plus fine-tuning — on a corpus that is one-third proprietary documents, one-third synthetic task data, one-third general material. Three months and roughly $40M.
The result beat GPT-5.4 and Claude Sonnet 5 on legal, tax and finance tasks, and led by about 15 points on factuality over open web content. Customers can run it on-premises.
The strategic read matters more than the benchmark: rather than licensing its archive to a lab, Thomson Reuters spent the money to keep the archive and the weights. A 35B model, Thomson-Small, is out for academic use.
The same issue covers Z.ai naming Ox Alpha as GLM-5.3-Flash — 320B total, 18B active, MIT-licensed weights, and $0.09 per task against $0.68–$2.03 for comparable proprietary models.
Shorter tool output made the bill go up.
GitHub's Copilot team went hunting for token savings and found the obvious lever backwards: trim a tool response too hard and the agent makes another call to recover what you cut, so the task runs longer and costs more. They started measuring cost per completed task instead of per response, then shipped four changes that stack to roughly 14% off the bill with task quality held.
The same model, behind two different doors.
Fable 5.1 and Mythos 5.1 are one model at two safeguard levels: Fable generally available, Mythos with fewer restrictions for vetted cybersecurity and life-sciences professionals in the US. On Terminal-Bench 4.0 that gap is visible — 55.8% against 60.9%. The number to plan around is the cache read price: at $0.25 per million tokens, Anthropic puts typical workloads about 25% cheaper and heavily agentic ones up to 45%.
Terminal-Bench-Science 0.1 — 52.6% (was 24.7%)
Humanity's Last Exam — 65.0% with tools
CursorBench 3.2.0 — 73.4%
Bio false positives down 85%; cyber interventions down ~60%
The first model rated Critical on cyber, shipped with the guards bolted on.
GPT-6 Astra is the first model to reach the Critical cybersecurity tier of OpenAI's Preparedness Framework, and it shipped alongside a safety overview and a separate note on the safeguards that made release possible. The capability story is computer use, coding and scientific research; the argument to watch is architectural. Astra reasons with recurrent depth, so a chunk of its thinking never surfaces as readable text — which is exactly the surface safety researchers have been monitoring.
Point the agent at a URL and it gets a whole computer.
Fly's argument is that a stateless sandbox is the wrong shape for an agent: what it wants is a real machine with a real filesystem on a real network, kept across turns. Sprites are disposable cloud computers, and the MCP server hands an agent the API to create them, run commands, manage files and take checkpoints. There is machine-readable documentation at a fixed path so the agent can learn the surface itself, and a five-per-session cap so a loop cannot spawn a fleet.
› tools: create · exec · fs · checkpoint
› docs for the model: /.sprite/llm.txt
› guard: 5 sprites per session, mcp- name prefix
› idle cost: effectively nothing
Four nanoseconds a byte in, a third of the disk out.
Cloudflare re-compresses cacheable text — HTML, JSON, CSS, JavaScript — with Zstandard inside Pingora, and the arithmetic is the whole story. Encoding is paid once on cache fill, decoding on every serve, which is why the decode side had to be the fast one. Pre-compressed media is skipped because there is nothing to win, and objects under 4 KiB are skipped because the CPU is not worth the bytes.
The estimate was 83% low, and one metric saw it.
An assistant had written a file-size estimator that multiplied source size by resolution ratio instead of bitrate by duration. Nothing threw, nothing logged, and every trace looked healthy — the number was simply wrong. One distribution metric recording actual bytes over estimated bytes, tagged by output format, made it obvious: MP4 to WebM averaged 1.57.
Record the KPI itself, not a log line about it. Attributes — format, quality, resolution — let you slice for a failure you had not predicted.
That's the week.
Issue 119 · Saturday 5 September 2026 · Zürich. Ten picks, ten sources, everything published between 31 August and 4 September.