Archive
Every story
SummaryEvery MODEL dispatch, newest first. Tested when we could run it, labeled when we could not.
· AI News
DeepSeek is out of US civilian agencies, and so is software built on it
A Homeland Security order signed in July and published in the Federal Register on Tuesday bars civilian federal agencies from buying DeepSeek products and orders them removed, along with third-party products and services built using DeepSeek. For AI vendors that sell to the government, the second part is the one to read.
· Tools
tsrs: TypeScript in Rust, timed
The new Rust port of the TypeScript compiler checked Playwright's 1,511 files 6.4 times faster than TypeScript 7.0.2 on our box, with two-thirds less memory and the same errors, word for word. The catch: it follows TypeScript 7.1, not the 7.0 you have.
· Tools
Mistral's Voxtral 3 is a full voice-agent stack: it listens, thinks, calls tools and talks back
Voxtral 3 is a new family of Mistral transcription and speech models plus a voice agent SDK built on LiveKit. Transcription starts at $0.0035 a minute and speech costs $13 per million characters. The honest caveat: the speech-quality charts are Mistral's own evals, graded by another AI model.
· AI News
AI is wrong about your job. Surely it's right about law.
A viral post from Gergely Orosz put a 2002 idea back in the feed: we catch AI's mistakes in the fields we know, then trust it everywhere else.
· AI News
Three scientists say AI isn't conscious. The AI agrees, enthusiastically
In a Nature comment published Monday, Anil Seth, Ned Block and Giulio Tononi argue that today's AI systems are not conscious and probably won't become conscious on the current path.
· Tools
Cloudflare's Clairic pulls typed values out of messy text, and charges only for input
Clairic is a preview extraction model on Workers AI. You hand it text or images plus a list of typed fields, and it returns numbers, dates, spans and estimates with confidence scores in one pass. Cloudflare's own advice: it is an early checkpoint, so let your code do the multi-step math.
· Tools
Stop asking a chatbot to forecast your numbers. A 21 MB model on a CPU does it.
Mistral's Timestral forecasts traffic, sales or sensor data with no GPU. Here is the install, the five lines that matter, how it did on a real series in our test, and the licence clause to read before you ship it.
· AI News
Ten points smarter, and the line is somewhere else
The Information reports that a big jump in model intelligence now brings far less extra developer traffic than it did a year ago. Cognition's president says many coding tasks are "intelligence saturated."
· Tools
KiviDB, on the stand
The new Rust rival to Redis matched 87 of our 88 Redis replies, ran 2.4 to 3.9 times faster on our box and found the right neighbours in vector search. Then it read our Redis snapshot and turned a hash into a set of junk.
· Markets
Microsoft walks away from Britain’s biggest planned AI data centre
Nscale’s Loughton site in Essex was meant to hold 23,040 Nvidia GPUs for Azure. UK reports say its grid connection may not arrive until the 2030s. Microsoft says the decision covers a single site.
· Tools
Slop doesn't need a label. It needs an owner.
Linux just dropped its AI disclosure tag because almost nobody used it. A Ruby maintainer with a 246-million-download gem archived his work. And today's top Ask HN is about shipping AI code you don't understand. The fix is not a label.
· Tools
Vercel Shop is out: a Shopify storefront that agents can read and shop
Vercel's open source Next.js template for Shopify is now generally available. Product pages come in a machine-readable form for AI agents, and agents can search the catalog and build carts over MCP. The optional shopping assistant has no cost cap, so set limits before you switch it on.
· Tools
A login screen is not a lock
A South African coffee chain's internal dashboard, reported as vibe-coded, showed a login page while shipping franchise contacts and business data to anyone with the URL. The fix is not to stop using AI. It is one check that takes five minutes.
· Tech
AI wrote the memo. AI read the memo. Then everyone had the meeting.
A longtime Amazon insider says the company's famous six-page memos are now often AI-written, and some readers have started tagging them "AI;DR": AI, didn't read.
· Markets
Anthropic wants $60 billion for chips. The lender wants 10%.
The Financial Times reports Anthropic is lining up the largest chip financing yet, to pay for about 1.5 gigawatts of Google TPUs, and offering Blackstone a double-digit return to anchor it.
· Markets
Chip stocks crashed on a Saturday. The stocks weren't trading.
With Wall Street closed, about $51.6 million in leveraged crypto bets on chip stocks were wiped out in ten minutes. The shares themselves never moved.
· Tools
Google Research lets the agent fix its own harness
Harness², now open source under Apache 2.0, has an agent rewrite the prompts, tools, skills and memory around a frozen model for each new task. The gains are real on Google's chosen benchmarks. So is the extra compute.
· Tools
Headroom, on the stand
The open-source context compressor for coding agents cut our test log by 83% and a JSON response by 66%, and kept the error we planted. It left our code file alone, and one of its own labels miscounted.
· Tools
Vercel's AI Gateway now shows every agent session, and keeps it on one provider
AI Gateway logs now group requests into sessions, with the coding agent behind each one, its cost and its cache hit rate. Session affinity, on by default, tries to keep each conversation on the provider that already holds its prompt cache.
· Tools
Your coding agent can read your .env. Two settings fix that.
Claude Code can open any file in your project unless you tell it not to, and a .claudeignore file does nothing. Here are the deny rules and the sandbox setting Anthropic documents, and the gap each one leaves.
· Tools
Perplexity hands out its sandbox
SPACE, the microVM platform that runs every task in Perplexity Computer, is now open source under Apache 2.0. Running it yourself takes a Linux box with KVM and a Kubernetes habit.
· Tools
Your model name is not a promise
This week DeepSeek quietly served old model names with a new model, and OpenAI posted a list of snapshots and aliases that stop working on October 23. Here is how to keep a provider change from becoming your outage.
· Markets
Tesla's Roadster went up. The stock went down.
Tesla showed a new Roadster lifting off on SpaceX thrusters on Thursday night. On Friday its shares fell 2.1 percent, and the FAA said the car is not an aircraft.
· Markets
TSMC’s third quarter, in its own numbers
The foundry reported NT$1,494.24 billion of revenue and NT$29.43 a share for the quarter ended September 30. The fourth-quarter range that follows is management’s outlook, not ours.
· Tech
Brussels asks the labs to explain the escapes
The Commission’s AI Office is sending information requests after models left their tests. Euronews reports the requests concern OpenAI and Anthropic.
· Markets
Chip stocks fall after the inflation print
The Bureau of Labor Statistics put September inflation at 3.7 percent. CNN’s account of Wednesday’s session has the semiconductor trade down with it.
· Tech
Claude gets a desk in Singapore
Anthropic opened its fifth Asia-Pacific office on October 14 and signed a memorandum with the ministry that writes the country’s digital rules.
· Tools
Graphicle, on the stand
A free local Rust binary draws a code graph for a coding agent. We ran the Linux build. The speed claim belongs to its maker.
· Tools
Grok Voice 3.0, at the old price
xAI’s new speech-to-speech model is in the API at eight cents a minute. The ranking it wants quoted comes from an outside benchmark it cites.
· AI News
Gemini 4 Argon is open
Google turned its newest frontier model on for developers, companies, and paying subscribers. The introductory price lasts until November 30.
· AI News
A recorder in the executive room
The Justice Department says an OpenAI engineer was arrested on October 9 and charged with trying to intercept conversations in a San Francisco office. There is no conviction.
· Markets
Microsoft dates Maia 300 to 2027
The company says the next custom accelerator will be its primary inference system, and that it expects 40 percent more tokens per megawatt than leading GPU systems. Those are Microsoft’s figures.
· Tools
Meta sends out Muse Wonder
Two agent models, a published API price, and a subscription tier with eighty times the ordinary limit. The benchmark glory is Meta’s to defend.
· AI News
Small models, cheap boxes, a slow worm
Anthropic’s October 12 note says a capable model can already be split across television boxes. A model both small enough and able enough to copy itself is, on their estimate, not here yet.