Seven engineers built the new source control system in five months, and it uses about 200 times less CPU and memory than the old Git mirrors. Read it if agents now write most of your team's commits.
Datalab says it scores 41.3 on OmniParseBench text, against 19.3 for Tesseract 5, on 32 CPU cores. Try pip install indra-ocr before you pay per page for document parsing.
Timescale says its PRISM index beats pgvector's HNSW at equal recall on 100 million vectors with a quarter of the memory. It is alpha and needs Postgres 18, so test it before you trust it.
You upload robot data, it fine-tunes a pi model and serves it from a hosted gateway, and one customer built a deburring model in four days. Access is by waitlist.
The reinforcement learning policy trained with no human demos on parts as small as 4 mm. It taught itself tricks like setting a peg down to regrasp it.
In Composo's physics sim, GPT-6.1 Sol scored 190 of 220 points on the breakfast level, and both models beat a careful scripted player. A fun test of planning when space runs out.
That is under a third of Claude Opus 5.5's $8,708 and under half of GPT-6 Astra's $5,324. His point: count what a finished task costs, not the price per token.
His hook rejects any commit that skips his diagnostics library and tells the agent what to use instead, and he says that made it comply every time. Copy the pattern for your own house rules.
He argues types are guardrails for agents and proposes a goal of unsafe code without unsafety, backed by proof tools. Meta wrote twice as much Rust in 2026 as in the nine years before.
Seeded randomness, save and load from day one, and screenshot diffs give Claude Code a fast way to check its own work. The same tricks help any app an agent edits.
Jev has a lot to answer for. Ever since AI startup TypeSafe unveiled the decision model in mid-September, the industry The post “Take work away from… · The New Stack
Most people who try to run a model at home end up with a $3,000 graphics card that sits idle twenty hours a day and a chatbot that forgets the… · Codez (@0xCodez)
Prosecutors say a recorder hidden behind a TV at OpenAI's San Francisco office was found on Sept. 30, and OpenAI says its own models helped trace it. It sees no sign that weights or customer data were taken.
Its Global Financial Stability Report names an AI correction as one of three big risks, with the 'magnificent seven' now over 35% of the S&P 500. Asian markets are the most exposed.
Bloomberg reports an internal memo ordering an 'AI first' plan, with all art made through AI workflows. Epic says its developers still lead the creative work.
Lambda Web Functions deploy an Express or Next.js server with one command and bill CPU time, so waiting on a model's streamed tokens costs little. Node.js 24 on Arm only, in five regions.
Its 2024 rule treats code from LLMs as tainted unless the core team approves it, and September donations beat a $50,000 goal. Debian voted the other way in August.
Tampered Nano X units read the seed phrase and sent it out over a mobile network, from more than 300 wallets. Ledger says not to use any bought from dealer Cryptobilis in the last three months.
Aftonbladet blamed an AI agent for erasing several hundred days of editing work, and TV4 says a failed backup system did it. Either way, test your restores.
The BriefGoogle opened Gemini 4 Argon, its new top AI model, to paying users on Tuesday: developers on the paid Gemini API, businesses, and Google AI Pro and Ultra subscribers in the Gemini app. It costs $2 per million input tokens and $10 per million output tokens (a token is about three quarters of a word) through November 30, then doubles to $4 and $20 on December 1. Artificial Analysis put a full test run at $2,407, against $8,708 for Claude Opus 5.5, so the next seven weeks are the cheapest time to try it on your own work.
Level UpBefore November 30, run one real task from your own work through gemini-4-argon on the paid Gemini API. If it can wait, use batch mode at $1 in and $5 out. Write down what the finished task cost, not the price per token, and compare it with the model you pay for now. →Gemini API pricing: Argon's intro, batch and December 1 rates