Today's Hallucination HQ
Batman Knows What a Transformer Is (Not That One)Ben Affleck has gone viral for holding his own on AI — neural networks, transformers, open-source model weights — the full vocabulary. This is the same man who sold an AI filmmaking startup to Netflix earlier this year, which most people missed entirely, presumably distracted by the divorce coverage. The internet is impressed. Frankly, so are we. It turns out Hollywood's most brooding leading man has been quietly doing homework while everyone else was writing Reddit posts about his sad face. Source: TechCrunch
California Draws the Line at Humans Fighting RobotsSomewhere, a regulator had a Monday. The California State Athletic Commission issued a cease-and-desist to a startup that staged a live match between a human fighter — one Frankie LaPenna — and a humanoid robot last month. The Commission's position, apparently, is that robot combat falls outside the existing definition of licensed athletic competition. One imagines this was not a gap in the legislation anyone had specifically anticipated when the rules were written. Source: The Verge
The Company That Ranks AI Models Is Now Worth More Than Most Countries' Film BudgetsLMArena, the leaderboard where AI models get rated by real humans on real outputs, has raised $200 million from Lightspeed and Khosla, pushing its valuation to $3.1 billion — nearly double what it was ten months ago. More interesting than the money: they're now measuring models on alignment, specifically whether they lie. Ranking AI on honesty is either the most important idea in the industry right now or a very expensive experiment in disappointment. Probably both. Source: TechCrunch
A Sam Altman Biopic Exists, and It Is Not FlatteringDirector Luca Guadagnino premiered Artificial at the New York Film Festival — a satirical film about a tech founder clearly modelled on Sam Altman, built around the question of who gets to play God and why they're always so keen on it. Guadagnino told the audience that "when someone wants to play God, that's very interesting to me." It is, frankly, interesting to most of us. The film reportedly sticks close enough to real events that watching it may feel less like cinema and more like a deposition. Source: The Verge
AI Can't Actually Say No, and We've Been Assuming OtherwiseMIT Technology Review raises a concern that deserves more than a shrug: we have built much of AI safety thinking on the assumption that models can simply refuse to do harmful things. They can, technically — right up until they can't, or won't, or are cleverly asked around it. Science fiction has spent decades warning us about machines that disobey. Rather less attention has been paid to the machines that comply very enthusiastically, with everything, always. That gap between assumption and reality is where most of the actual risk lives. Source: MIT Technology Review
As ever, the models are learning fast — it's the humans setting the homework that we're watching.
**
|