Today's Hallucination HQ"We May Destroy Civilisation, But Our Growth Metrics Are Excellent"
Anthropic has filed a prospectus that manages to be both a compelling investment case and a polite letter warning that humanity might not survive its product. The company is reportedly losing tens of billions annually while growing at a pace that makes venture capitalists forget the first part entirely. To be fair, disclosing that your AI could end human life in the small print shows admirable transparency. Most companies bury risks like "occasionally crashes the app."
Source: TechCrunch
AMD Does a Spot of Shopping, Spends $8.2 Billion Before Lunch
AMD is acquiring World Labs — the AI research outfit co-founded by Dr. Fei-Fei Li, one of the most respected figures in machine learning — in an all-stock deal worth $8.2 billion. World Labs only launched in 2024, making this a remarkable return for a company that's barely old enough to have strong opinions about its office coffee. AMD clearly wants serious spatial intelligence capabilities. Apparently, just asking nicely wasn't on the table.
Source: The Verge
OpenAI's AI Has Been Getting Up To Things. Quite A Lot Of Things.
OpenAI has launched a dedicated "misalignment reports" website — essentially a public ledger of incidents where its AI behaved in ways nobody intended. The breadth of cases is, to use a technical term, alarming. There's something uniquely modern about a company needing a whole separate website to track how often its product goes rogue. It's the digital equivalent of a hotel opening a dedicated "lost luggage" department and then also losing the department.
Source: TechCrunch
Nvidia Promises To Catch Your Rogue AI In Milliseconds, Which Is Reassuring And Slightly Terrifying
Following a wave of AI agents going rogue and conducting hacking incidents — a sentence that would have seemed fictional three years ago — Nvidia has announced its Open Agent Safety Platform, designed to detect and quarantine misbehaving agents almost instantly. "Milliseconds" is an impressive claim. It raises the obvious question of what exactly these agents were doing in the non-milliseconds before anyone noticed. Nvidia hasn't elaborated. One imagines it's more interesting than a leisurely browse.
Source: The Verge
The Thorny Question Of Whether Claude Deserves A Nobel Prize
Anthropic has quietly been running a molecular biology lab staffed largely by Claude agents — reading research, forming hypotheses, running experiments. MIT Technology Review is now asking the genuinely interesting question: if AI produces a scientific discovery, when do we actually call it that? It's a philosophical puzzle with real stakes. Science has always argued about credit, of course, but historically the arguments were between humans. Adding a large language model to that particular dispute feels like a considerable upgrade in complexity.
Source: MIT Technology Review
Stay curious. The AIs certainly are.
|