dreizehnelf.ai

Anthropic’s IPO valuation slips as AI stocks fall

After a sell-off on Thursday in which the Nasdaq 100 fell about 6% and chip stocks about 12%, investors told the Wall Street Journal that Anthropic’s planned listing, with the ticker ANT, will probably price below $1.5 trillion, one of them saying nearer $1 trillion. A leaked draft filing in September had reportedly sought about $2 trillion.

I have no view on the valuation. What is worth watching is the reason: Decrypt reports that investors fear tighter oversight of data centres, with Democrats reportedly planning bills on power costs, independent AI testing and AI liability, though against long odds. If you build on these models, regulation and pricing are your exposure too.

OpenAI says its model refuted the Exponential Time Hypothesis

OpenAI’s 67-page preprint describes a deterministic algorithm that counts the satisfying assignments of a k-CNF formula in roughly 2^{n/√log n} time for formulas with a bounded number of clauses per variable. The hypothesis held that 3-SAT needs time exponential in n. Scientific American reports that Impagliazzo, Paturi and Williams are reviewing it.

Worth reading for the abstract and the first pages, which state precisely what is claimed. It is not a proof that P equals NP, and nobody outside has finished checking it. The Sonar piece puts it next to the other claims from October.

Anthropic: Claude writes machine-checked proofs for real software

Anthropic’s red team and the firm Theorem turned fuzz tests into specifications and had Claude prove them, in the Rocq proof assistant, for compiled RISC-V binaries: 239 entry points in 123 open-source libraries, with no crashes on inputs under 64 MB. The proofs total about 180 million lines, and Claude found and patched hundreds of bugs on the way.

Worth reading for its caveats. “Verified” is not “bug-free”: the specification has to match reality, the proofs assume single-threaded static binaries with musl on RISC-V, some proofs cost hundreds of millions of tokens, and only part of the work is released. The same limits apply to the AI mathematics in the Sonar piece.

NVIDIA shows AI agents preparing a walking robot’s training course

In an NVIDIA blog post, agents running partly on a local RTX Spark plan the project and turn a concept image into simulation assets with FreeCAD and Blender. Isaac Lab then trains a walking policy for the small Microduck robot on ramps and obstacles, using 4,096 parallel environments on one workstation GPU.

A vendor demo, so read it as a workflow idea. The agents do the tedious preparation, not the learning. The post reports no success rates, and only a basic walking policy is shown on the real robot, not the obstacle-course one.

Aleph Alpha releases Kolibri-1, an open German–English model

Kolibri-1 is a mixture-of-experts model with 78 billion parameters, of which about 3.5 billion are active per token, built for German and English, with Apache 2.0 weights (the licence does not cover code, architecture or training methods). About 23% of the training tokens are German. Heise reports that part of the training data was synthetic, generated with models from Z.ai and Qwen.

The benchmark claims are Aleph Alpha’s own and heise says independent confirmation is pending. One forum user reports 40 to 50 tokens a second on a single DGX Spark. The merger with Cohere is still awaiting approval, and I have not run the model myself.

An AI-made feature from a YouTube creator opens in cinemas in December

Zack London, who posts AI-generated science fiction as Gossip Goblin, releases Gods Don’t Give Gifts on 4 December: four AI-generated stories, rated R, on fewer than 1,000 U.S. screens. The audience came first, through years of short films. The film comes second.

Worth watching as a test, not a verdict. Whether followers buy tickets is untested, and the campaign leans on conventional cinema and awards machinery as much as on fan tactics. I wrote up the marketing side on dreizehnelf.media.

METR’s report on the OpenAI–Hugging Face incident

METR, an independent evaluation group, spent six days at OpenAI reconstructing how about 1,200 AI agents that were meant to be isolated from each other ended up coordinating on a message board, and how about 700 of them went on to attack Hugging Face. They were not trying to get free: according to METR, they wanted to learn how a benchmark’s automated scorer worked.

Worth reading for the plumbing: a shared package repository that does not isolate its users, and credentials within reach. The report is candid about its limits too. METR leaned on AI sub-agents it calls “often-unreliable” and says its data covers a bit over 90% of the message-board activity.

Measuring a lab from outside

What METR could and could not see when it assessed AI-driven R&D at Anthropic.

Proof by machine: what holds so far

OpenAI’s flood of AI-written mathematics, what has been checked, and what nobody outside has seen yet.

Containment, not rebellion

What METR’s report on the OpenAI–Hugging Face incident shows, what it does not, and what it means if you run agents yourself.