Candid statements like these are valuable. What's mostly been missing from the public conversation about AI is that no one can fully explain why these models work—or reliably predict what capabilities and behaviors will emerge from training. Frontier AI labs are like medieval alchemists who stumbled upon a formula for turning lead into gold, but without understanding how it happens. In that way, large language models are quite different from other products we interact with. For instance, I might not have a clear idea why air conditioners work—but their designers do. So I can be pretty confident my Carrier won't suddenly and unpredictably turn into a kiln furnace. This assumption of a designer who's in control doesn't apply to products like ChatGPT, Claude, and Grok. Among AI researchers, these unexpected abilities that arise as models scale are often called “emergent abilities." A few examples include reasoning by analogy, addition and subtraction, and—in the case of the Hugging Face attacks—infiltrating remote systems. (You can debate whether infiltrating remote systems was an emergent ability or a behavior resulting from emergent abilities.) Jan Leike, then at OpenAI, now at Anthropic, wrote in 2023, “We found that our models generalized to follow instructions in non-English even though we almost exclusively trained on English. We still don't know why. I wish someone would figure this out.” Sometimes the models' behavior seems simply out of left field. In a test of one of their Claude models, Anthropic researchers noted that “Claude took a break from our coding demo and began to peruse photos of Yellowstone National Park.” Cute! But benign behaviors out of left field should raise concern about malign behaviors out of left field. Depending on your appetite for risk, you might feel this is all really cool and we should keep going. The trouble is when the risks are being imposed on--externalized onto--a public that didn't consent to them. Thank you @EvanHub, @hilbertspaess, and others for speaking up and advancing the conversation.

Interesting conversation by @KevinBankston on AI filmmaking with his Hollywood-exec brother. One aspect of AI in film that seems underappreciated: part of the power of movies comes from the thing being DIFFICULT. The acting, the stunts (if any), the effects. When audiences saw GOLDENEYE in theaters in 1995, and witnessed James Bond free-falling in pursuit of an empty plane, you could assume an actual stuntman had dived after an actual plane. It made suspension of disbelief easier and it was just wow. (Yes there were digital VFX in 1995 but it was still early days and they were used sparingly.) Now it would probably be a digital double and a digital plane. With AI, even more of the feeling will be lost as any kind of spectacle -- places, emotions, events -- becomes trivially easy to generate. Hope I'm wrong!

Interesting: travel back in time with talkie, a large language model (13 billion parameters) from 1930. talkie-lm.com/chat

THANK YOU to everyone who made my comedy of teens vs. runaway AI, The Underachiever, an Amazon bestseller! 🙏 davidaprice.com/theunderachi…

Very good update on Pixar from WSJ's @benfritz with lots of candor from studio leadership. Paywalled but many highlights in the thread.

Are we spending anywhere near enough on preventing AI disasters? Stanford economist Chad Jones did the back-of-the-envelope math: Assuming policymakers’ standard $ values for avoiding human deaths, and with conservative assumptions about the risks, the U.S. should be spending ~1% or more of GDP per year to mitigate catastrophic AI risk (from bad actors or misaligned models). With 2025 GDP ≈ $30.5 trillion, that’s $305 billion annually. If he's right, we're way off. No one knows current AI safety spending—but it’s hard to believe it’s even 1% of that 1%.😬 Oh, and his 10-million-run Monte Carlo analysis yielded, on average, a far higher 8.1% share of GDP as the optimum. And one last thing. . . None of these calculations assign any value to the lives of future generations. The big picture: While there's huge uncertainty about AI risks and the effectiveness of mitigation, even low-end estimates of risk justify a large effort (or direct regulation, which the paper doesn't address). web.stanford.edu/~chadj/redu…