Favorites: Reward Hacking, Drowned Cities, and Kosovo
July 23, 2026 — Grad students now make up a quarter of UAW membership, US prisons shrink as crime falls, confidence in higher education keeps sliding.
Yudkowsky continued (cached): "What if the human exam developers wrote down a wrong answer? Wiser to steal the answer sheet. The First Rule of RL is that any RL signal sent by an imperfect evaluator (⊃ humans) is maxed out by targeting the evaluator's mistakes, not by targeting the evaluator's target" Check out this blog post (cached) from the HuggingFace folks before they knew where the attack came from. One commenter jokes (cached), per the previous item, "openai announces that in a worrying demonstration, during a recent benchmark exercise requiring models to solve a math conjecture, their latest model broke out of a sandbox and gained illicit backdoor access directly to the platonic realm of forms" (src)(cached)
One person joked that you can have all kinds of 'little free' things like this, maybe a GPU on a stick with a solar panel, little free datacenter. I'm sure the neighbors would love it. (src)(cached)
Ethan Mollick’s posts where a new AI fails at one of his prompts don’t go as viral, but Gemini 3.6 Flash is pretty far behind the state of the art for constructing 3D scenes. For comparison, check out Opus 4.8's version (src)(cached)
Josh Barro notes (cached) that this was basically the only instance where "let's bomb the bad country until they behave better" worked quite well. (src)(cached)
Flotsam and Jetsam
– Analysis of FEC donation records for employees of top tech companies. Overwhelming majority donate to Democrats, just CEOs and top executives give to Republicans, especially after the 2024 election. Even 85% of Palantir (cached) employees who donate gave to Democrats. (src)(cached)
– A lot of new solutions to longstanding unsolved math problems that were probably just not getting much attention previously, but now folks are bulk-solving them (cached), and to some extent demonstrating that they probably could have been solved previously if someone had devoted more time to trying things. The main one that everyone was talking about was the Jacobian conjecture, which was done by an Anthropic employee who hasn’t shown the prompt yet. There was a graph theory one where the prompt was very much not fancy, “do a breakthrough” the user asked. Prompting ability is now maybe overrated (cached)? (src)(cached)
– "In Watchmen, once superheroes become real, all the comic authors switch to writing about pirates. I feel like a similar thing is happening with sci fi vs fantasy." (src)(cached)
– New rhythm game where you play as a conductor using the controllers from a Nintendo switch. Seems like fun. (src)(cached)
– LA Times making a clear argument that we need to stop subsidizing rich people living in fire zones by having the state pay for their insurance. (src)(cached)
– Most significant group opposing AVs are trial lawyers worried that a reduction in accidents will mean less business for them. Just absolutely evil rent seeking. (cached) (src)(cached)
– Ukrainian anti-drone drones now have recoilless shotguns, for every shot fired, a second shot is fired in the opposite direction at the same time. (src)(cached)
– "There are 9 living former British PMs now. We should turn them into Nazgûl. Tony Blair can be the Witch King." (src)(cached)
– More than 25% of United Auto Workers membership now consists of unionized graduate students. (src)(cached)
– Did you know that since 2007 the US has significantly reduced the number of people in prisons, especially young people? And we've done it while reducing crime, pretty good! (src)(cached)
– New paper on Inclusionary Zoning in California. The ‘normal’ IZ mandates are already ridiculously harmful, and fall particularly hard on the poorest residents. (src)(cached)
– Public opinion on higher education has fallen a lot, even among democrats, confidence has fallen sharply, especially in the last few years. (src)(cached)
– The White House has finally admitted that their aluminum tariffs didn’t work (didn’t result in new US production of aluminum) (src)(cached)
– Alpha School is claiming some extremely impressive stats for kids test score improvement, Kelsey Piper argues that many in the past have also claimed impressive results like this and have not replicated. That doesn’t mean this won’t either, but we should be very skeptical. (src)(cached)
– The Nordic states don’t maintain their excellent public services by “taxing billionaires out of existence”, they do it with well structured (Georgist) taxes on oil drilling. You can read a whole article on the subject. Norway and Sweden actually both have more billionaires per capita than the US. (src)(cached)
– If you want the full list of Michelin starred restaurants in CA, here it is (cached) (src)(cached)