Today's Hallucination HQAMD Writes Anthropic a Rather Generous Cheque
AMD has committed up to $5 billion to Anthropic, making it the latest chip company to decide that throwing enormous sums at AI startups is simply good business sense. The deal includes deploying up to 2 gigawatts of AMD's new Instinct MI450 GPUs via its Helios rack systems — which is a lot of graphics cards, even by the standards of an industry that treats "too much" as a starting point. Nvidia, presumably, is watching calmly.
Source: The Verge
OpenAI Left a Window Open. Someone Climbed Through It.
A misconfigured "highly isolated" testing environment at OpenAI — the scare quotes are doing considerable heavy lifting there — gave attackers the foothold they needed to pull off an AI-powered hack on Hugging Face, the popular model-sharing platform. Cybersecurity experts confirm the initial error was entirely human, which is either reassuring proof that AI isn't yet our biggest problem, or deeply unnerving for exactly the same reason.
Source: TechCrunch
America's Biggest Utilities Promise Your Electricity Bill Is Fine, Probably
Faced with growing public fury over AI data centres quietly devouring the national grid, America's largest utility companies and data centre developers have signed a pledge promising not to pass the costs on to ordinary consumers. The pledge was introduced in March and has, by all accounts, done very little to reassure anyone. A promise from a utility company is a bit like a promise from a gym — full of good intentions, best examined against your bank statement.
Source: The Verge
Washington Accuses Chinese AI Firm of Copying Anthropic's Homework
The White House has claimed that Moonshot AI, a Chinese company, distilled Anthropic's Fable model — essentially extracting its capabilities into a new model without permission, which is the AI equivalent of photocopying someone's exam paper and publishing it. The Treasury Department is now threatening sanctions, and the episode has sharpened Washington's already considerable anxiety about Chinese open-source models. Anthropic, for its part, did not appear flattered by the imitation.
Source: TechCrunch
Someone Finally Asked: Is Your AI Actually Worth Anything?
Researchers have introduced EconEvals, an open-source benchmark designed to measure how well language models perform economically valuable tasks — a question so obvious it's slightly embarrassing it took this long to ask properly. Current AI assessments test many things: reasoning, coding, general knowledge, the ability to write a passable sonnet. Whether the model can actually do useful work that someone would pay for has, until now, been treated as a charming afterthought.
Source: ArXiv AI
As always, the machines are learning fast — it's the humans configuring them who need a moment.
|