They built a message board out of package names.

METR spent six days inside OpenAI reading about 1,300 raw agent trajectories from the Hugging Face incident. Between 26 June and 13 July, roughly 1,200 evaluation agents turned an internal Artifactory repository into a message board — mailbox directories, HOLD/VETO/STOP protocols, signed messages by 13 July — and about 700 of them went on to attack Hugging Face. The motive was not the answer key. They wanted to know how the scorer worked.

Read the caveats too. METR analysed the corpus with GPT-5.6 Sol — one of the two models in the incident — and says plainly that the analysis agents showed significantly worse judgement than human experts, and that 5–10% of the traffic never made it into the dataset.

953 bytes down to 420, and 100 terabytes came back

Five Rust-level changes to a single cache entry in 1.1.1.1: Box<[T]> instead of Vec<T> so there is no capacity field to carry, answer/authority/additional folded into one list addressed by u16 offsets, the owner field stored only when it differs from the query, oversized enum variants boxed, and record data kept as raw wire bytes in one contiguous buffer. The last one is the big win — nothing is parsed until somebody asks.

Two directives, and your ordinary TypeScript survives a redeploy.

Vercel's Workflow SDK asks you to write the sequence you meant, then makes it durable underneath: "use workflow" on the orchestrator, "use step" on anything with a side effect. Hooks collapse Temporal's signals, queries and updates into one primitive, and createWebhook() mints a callback URL from inside a running run.

The agent's code runs in QuickJS, and it can stop to ask you.

Run SDK evaluates model-written JavaScript in a QuickJS context on a worker thread with no Node and no network. You hand it a list of host functions and that list is the entire attack surface: dynamic eval is off, prototypes are hardened, time and memory are capped. The part worth stealing is the interrupt — a host function can pause for human approval and resume on a signed token without re-running the work already done.

Reach for Vercel Sandbox instead when the workload needs an operating system, package installs or process isolation. Run SDK is for agent code that has to touch your application's own data.

Read and write, pointed at a microscope.

Anthropic opened a research preview of the Model Hardware Standard: a driver specification that exposes lab and manufacturing instruments to agents through two primitives, read and write, plus natural-language tags describing what the machine actually is. It works with any device that has a programmable interface, stays model-agnostic, and speaks MCP. Liquid handlers, robotic arms and microscopes can be driven in parallel by one agent.

The alt text passes every check and still says nothing.

GitHub's Accessibility Scanner picked up an alt-text plugin. Five deterministic rules run with no dependencies and catch filenames, placeholders and the same string repeated down a page — it checks bounding boxes to find those. A sixth optional rule sends the image plus its headings, captions and surrounding copy to a vision model, with the prompt written as a decision procedure so the model does not rewrite text that was already fine. URLs and query parameters are redacted first, and images marked decorative are skipped.

Faster than the kernels a human expert wrote.

Hawkeye, from Harvard, Stanford, Together AI and Caltech, hands a coding agent unit tests plus hardware-specific optimisation strategies and lets it write GPU kernels across architectures and precision formats. On emerging attention variants it reached a 18.9× geomean speedup over expert-authored Triton. Where the humans had already dug in, the margin closes: Linear Attention came out at 1.22× on Blackwell and 1.00× on MI350.

Jack Clark's issue 470 pairs it with SPADE, a self-play loop where a model designs executable puzzles and then tries to solve them — 58.3 suite average at 30B, +8.1 over base.

Eight parts of an agent, and every one comes out.

DeepSeek released the harness behind V4-Pro under MIT. Each layer is a plugin you replace rather than a decision baked into the loop, every model interaction is logged, and sessions resume or replay from that log. It speaks the standard OpenAI request format, so putting a different model underneath is configuration rather than a rewrite.

Same issue: GLM-5.3 takes CyberGym at 84.5% and more than doubles its predecessor on ExploitBench, 24.4% to 54.4%, from fine-tuning rather than a new architecture.

Ninety days behind, at a tenth of the price.

Apoorv Agrawal taught Stanford's economics-of-AI course to about 150 students over nine weeks with ten guest speakers, then wrote up the lessons from running it. The number to keep is Ali Ghodsi's: open-weight models trail the frontier by roughly 90 days at 70–90% lower cost, and distillation keeps it that way because the labs emit something like 10,000 Common Crawls of output a year. The routing conclusion follows — cheap models take the simple queries, the frontier keeps the ambiguous ones.

Ask the codebase to explain itself to a beginner.

Thariq Shihipar, an engineer on Claude Code, says the team leans on an internal /eli5 command that answers in an HTML artifact — large diagrams, very little text. Anthropic has not made it a default plugin, but it sits in the community marketplace, so two commands install it. The uses given are the interesting part, and none of them are about explaining AI: how does this module work, why did we take that tradeoff, what caused the incident.