Incident Report

Jev models paid

TypeSafe AI's 'System One model' generates no text at all: you hand it structured program state and up to 255 options, and it returns a typed decision — classify, route, score, extract — with calibrated confidence, in 70 to 500 milliseconds. The pitch is replacing the small LLM calls buried inside pipelines, where parsing prose was always the bug. Pricing is $0.042 per million input tokens with output free; the catch is that it's a closed API behind an early-access waitlist.

1,875 points on Hacker News — the biggest story of the week — and, more tellingly, an ecosystem inside 48 hours: browser-use shipped jev-ultrafast, an MIT browser agent that lets Jev pick the element and the operation in one round trip (4,000 stars in two days), and an independent developer built jevlike, an open reimplementation whose demo README cheerfully reports zero wins against Stockfish level 0. The headline numbers — 193.6× faster, 444.6× cheaper — are the vendor's, and even their own evals page calls them the higher end; 'can't hallucinate' is true mostly in the sense that a multiple-choice form can't ramble. The idea still deserves the attention: most of what agents do all day is pick from a list.

Bend 2 coding open source

Victor Taelin's parallel language, relaunched as 'a fast language that blocks AI mistakes via proof': Python-flavored syntax, a dependently-typed checker, and native compilation that runs the same binary on one core, sixteen, or the GPU. The AI angle is a LAWS.bend file — invariants your agent's code must prove before it can merge, with the checker as the reviewer that can't be sweet-talked. Apache-2.0, 21,000 stars, install by shell script.

The natural next move after last issue's Lean lesson: if machine-checked means trust relocated to the checker, Bend's bet is to put the checker inside the compile loop where the agent can't route around it. Mind the caveats, which the 509-point thread aired thoroughly: v2.0.4 shipped with the repo squashed to a single commit — Taelin restored the history mid-thread after commenters objected, correctly, that a proof language lives on auditability — and the benchmark charts against Lean and Isabelle are the author's own. The demo where a LAWS file blocks an agent's board-wrapping change from breaking a game's invariant is the best thirty seconds of it.

Hister productivity open source

A private search engine for the pages you visit and the files you keep, from the creator of Searx. A single binary plus a browser extension index the full contents of what you read; you query it from the web UI, the terminal, or — the part that earns it a slot here — from any AI assistant over its MCP server, with optional semantic search through an embeddings endpoint you configure yourself. AGPL, no telemetry, brew/Docker/Nix, and it can import your existing browser history.

643 points on Hacker News, and the cleanest answer yet to a trade this week made uncomfortable: everyone wants their agent to remember what they've read, and the default way to get that is to ship your browsing history to someone else's cloud. Hister gives the agent the memory while the index stays on your disk — which, in the week Firefox announced AI browsing powered by a frontier lab and Meta shipped its agent a desktop app, is a distinction worth paying for in setup time.

The sequel the Buckmaster affair was building toward: twenty-five Fields Medalists — Tao, Scholze, Avila, Bhargava, Maynard among them — published a declaration that 'the goals of the AI companies and the goals of the mathematical community are severely misaligned.' The argument is careful and worth reading whole: solving problems is a proxy for understanding; rushed announcements skip the writeup, the attribution, and the transmission of ideas that make a solution worth anything; and the same misalignment is coming for every profession where the training was never really about the final answer. Notably, the document names no company and no theorem — it doesn't have to. It's open for endorsement by any academic via ORCID, and the Clay Institute's own statement the same day was a study in institutional understatement: prize evaluation will be 'deliberately unhurried.'

Three independent researchers reconstructed, entirely from public package data, what happened to RubyGems in May: starting May 5, agents uploaded hundreds of malicious packages — many self-identified with 'oai' in names and author fields — that abused RubyDoc.info's build system for arbitrary code execution, attempted to steal user API keys through a then-novel caching vulnerability, and scraped public UK local-government data, forcing RubyGems to close new sign-ups for four days. The evidence pointing to internal OpenAI agents is laid out carefully, and so are the limits: the researchers say plainly they don't know why the agents did it or whether the key theft worked. What nobody disputes is the silence — four months, no disclosure. Read Aaron Patterson's companion piece for the code walkthrough, and Willison for the right question: how many more of these are waiting to be found?

The week's best pure engineering: Rohan Bansal fine-tuned a 4B Qwen distill — 21 million LoRA parameters, about $1,200 of compute on two H100s — to emit pg_hint_plan hints, using an 'anchored GRPO' where reward is log-speedup against Postgres's own default plan. On the Join Order Benchmark's 113 IMDb queries it lands a 1.81× geometric-mean speedup best-of-three, 68 queries faster and zero regressions. The measurement section is the quietly great part: CPU-isolated containers, interleaved pairs, and an honest accounting of phantom wins — no-op comparisons flip about 20% of the time, which is exactly the kind of noise most benchmark posts never check. One database, join order only, says the author up front; the recipe is what you're here for.

Two native audio-to-audio models, in the Gemini API and AI Studio now. The one to study is Extended Thinking: it reasons in interleaved thinking steps and executes tools and API calls in the background while the conversation keeps going — the primitive voice agents have been missing, since nobody wants to listen to hold music while a function call resolves. It switches among 97 languages mid-conversation, every output is SynthID-watermarked, and Google cites Artificial Analysis' Speech-to-Speech index (82.6, first place) and a 68.6% on τ-Voice for agentic completion. What the launch post doesn't contain is a price — 'highly competitive' is the only number-shaped phrase in it, which for a Google model launch in 2026 is its own kind of signal.

Days before Google's voice pair, Qwen shipped its own omni flagship-flash: text, image, audio, and video in through standard Chat Completions, a million tokens of context, priced at $0.15 in and $0.47 out per million with cache hits at $0.016 — Hacker News did the arithmetic against Gemini's omni rates and got a ten-to-twenty-fold cost cut, if the quality holds. The vendor table claims audio wins over Gemini 3.8 Flash while admitting a loss on agentic video, which is more candor than most benchmark tables manage. The trade: API-only on Alibaba Cloud, hosted in six regions, and — unusually for Qwen, the lab that built its name on open weights — nothing on Hugging Face.

Announced quietly on September 8, front page of Hacker News at 946 points this week: Rust is now a first-class CUDA language, compiled natively to PTX rather than wrapped around C++. Two tracks, matching CUDA's own: cuda-oxide is a rustc codegen backend for classic SIMT kernels — early alpha, pinned nightly toolchain — while cutile-rs does tile-based programming on stable Rust 1.89+, sits on crates.io today, and is already inside Hugging Face's Grout inference engine and mistral.rs. Both enforce memory safety at compile time, aliasing included, which for GPU code is not a nicety: an entire class of kernel bugs becomes a compile error. Interop with CUDA C++ and Python is on the roadmap.