Favorites: Shoggoth Alignments, and the Vatican Vantage
Trump’s bragging about crashing the marke, and AI flops on new math—buckle up!
I know I said it yesterday, but I’m a fan of Henry George, so I’m pleased to post it again. First day talking point (cached) from Republicans was that the S&P 500 was still up over the last year. Now we’re at day two and that is no longer true. Day two they were pointing at 5 year S&P gains (cached), but those were all under Biden of course. At this point Trump can’t even make the markets go back up (cached) that much by rescinding the tariffs, because it’s not credible he won’t do it again. Only action by Congress to remove this power can fix it. He's just not credible (cached). It doesn’t actually feel good to see Musk and now Ackman pleading with Trump (cached) to end these tariffs. It’s really worrying that they’re doing this in public.
If you want that thread of styles, here it is (cached), although if you do, you'll find part of the answer to Thompson's question. Watch the movement of each one, and you'll notice that they're all different, in some Dylan looks back and forth, who's arm is Helly holding on to? The ability to reliably control character movement just isn't there yet in the video models, even if the image models can output an excellent starting frame. But at the same time, Thompson's point may stand anyway. Vaguely related minor item: the Miyazaki quote about hating AI generated content was taken way out of context (cached). It's worth seeing what he was originally talking about. Maybe he still hates this new AI generated stuff, but the quote is not applicable. Anyway, here's some footage, clearly from around Mountain View, made into ghibli-style animation (cached).
Anthropic released a video summarizing their findings. Or you can read their blog post and the paper it's effectively an excerpt from. As much as Anthropic is far behind OpenAI in users, I still find it to be the best of the LLMs. Semi-relatedly: Tim Lee jokes (cached) that all the 'agentic' systems so far are really just too slow at doing the tasks they're being asked to do, so they suggest you check back in later when they're done. But it would be better if they were done instantly. "If it looks agentic to the user it probably needs a better interface". But anthropic is still most famous for their work on alignment.
And it looks like chatgpt4o is doing way better at aligning the shoggoth. In other AI training difficulty news, while all the state of the art models were performing relatively well on US Math Olympiad problems, the new ones for 2025 were just published, and all the models absolutely failed on them (cached). Sounds like training set contamination to me. Here's a super detailed thread about what types of failures you find (cached) when you look closely.
Longer Reads
• Thread of economic graphs all showing wildly bad numbers due to the tariff announcement.
Flotsam and Jetsam
– Noah Smith highlighted some clips from a whole thread of colorized hundred-year-old footage. The kids look terrible, the young adults look even worse (cached), you don't want to live in the world of a century ago. (src)(cached)
– Oil and gas executives are absolutely freaking out because of how bad the secondary effects of tariffs are going to be. If the US goes into a recession, they're going to get absolutely crushed. (src)(cached)
– Our World in Data on X highlights four countries that have effectively reversed democratic decline in recent years, just in case it's useful to any recent events (src)(cached)
– Newsom is attempting to pitch world leaders on not retaliating against California made goods. Seems like a stretch, I’m curious whether it’ll work. (src)(cached)
– "The Vatican was completely excluded from the tariff list and Italy is at the 20% EU rate. I hope some scheming cardinal sees the opportunity here". But seriously, as Alan Cole argues (cached), the fact that they put a tariff on the penguins is proof that the whole methodology is made up. (src)(cached)
– Don’t blame Harris for not making it clear that Trump’s policies would lead to a crash, she said it very explicitly multiple times. What’s wild is that his policy is genuinely even crazier (cached) than what he campaigned on (src)(cached)
– Trump Is repeatedly posting that he’s crashing the stock market on purpose. Seems like bad messaging? (src)(cached)