Favorites: Predictions, Progress, and Patterns
Sora is now available, and was followed closely by Google's Veo2. The head-to-head comparisons seem to show that Google is actually quite far ahead on video generation. The prompt used for both was supposed to be ‘skillful’ knife use, and I think Google Veo2 is cutting the slices a bit thick. But Sora is losing fingers! Here's a comparison across a wider set of video models. Here's a thread of more head-to-head examples of Sora and Veo2.
It's a very well made survey about expectations about progress on AI over the next year. I'm especially impressed that they created Manifold and metaculus versions of the events. If you end up participating, let me know what your predictions were. Mine are here. Finally, here's a great thread (cached) about how predictions of significant progress on these benchmarks will probably not translate to large impact on society. Mostly due to Goodhart's Law.
For one concrete example, the liberals in Canada are about to get absolutely wiped out (cached) by the upcoming election.
When attempting to make accurate predictions about future world events, I try to always skew a bit extra towards 'nothing ever happens' because of these graphs.
Flotsam and Jetsam
– While Trump's foreign trade ideas are generally bad they seem in line with, Biden's export controls on chips, which appear to mostly be having their intended effect. Noah Smith argues that if Trump rolls them back that will be an indication that he's giving in to China. (src)(cached)
– There's a new method for extracting human egg cells that reduces the time and number of injections by like 80%. Seems pretty cool. (src)(cached)
– Telehealth can make it much easier to get valid prescriptions while jumping through fewer hoops, not sure if that makes a OneMedical subscription on top of existing insurance worthwhile though. (src)(cached)
– "When you account for higher cost of living, Los Angeles is the highest poverty city in America." Not super surprising, but good to know that it matches what I'd expect. (src)(cached)
– Google is moving forward with a "crush rocks to remove CO2 from the atmosphere" plan. I've mentioned this general idea previously, good to see money going towards scaling it. (src)
– People underestimate the amount of property damage from the George Floyd protests nationwide. (Try to guess a dollar figure before opening the link). But it seems like a big part of that is that people generally underestimate the cost of building stuff. (src)(cached)
– Respected economists all thought that Milei's election would destroy Argentina. In fact things have gone pretty well, but in part I think it's because Milei simply didn't do some of the dumber stuff he was promising, and mostly followed what experts recommended, while maintaining populist vibes. MAGA folks seem to really like him, but his policies have basically been the opposite (cached) of what Trump espouses, reducing tariffs has been his biggest success. (src)(cached)
– It really seems like Bezos isn't really paying attention to which nonprofit groups he's funding. I don't think he meant to fund groups who would push to block permitting reform (cached) and attack democrats from the left, but that's what happened. (src)(cached)
– Here's a thread from Anthropic where they explain how their model appears to 'pretend' to be aligned differently during training from when it's in use. Honestly I'm not sure I even get it, but I thought it was worth sharing what kind of thing AI doomers are pointing to as evidence of danger. (src)(cached)
– Since Lula was elected president of Brazil, the rate of deforestation there has fallen by half. (src)(cached)