Favorites: Corporate Mole Hunts, Shor’s Shifts, and Step-by-Step LLMs
March 23rd 2025: Youth vote turning red, hidden Slack channels, and drones picking their own targets.
The Rippling vs Deel lawsuit contains some wild details (cached) of the corporate espionage: Rippling used a honeypot to implicate the CEO (or one of the top lawyers) of Deel and confirm the identity of the mole in their org. They emailed the CEO and counsel to say "hey, there's a mole in our org that's been monitoring the Slack channel #d-defectors", but they had in fact just created that channel hidden and waited for someone in their org to search for it. They got a hit. And that doesn't even include the part Matt Levine excerpted for his social media post, when the mole was confronted and asked for his phone, he "feigned compliance before hiding in the bathroom and then fleeing the scene".
Shor was interviewed by Ezra Klein to break down the 2024 election, and he brought slides. One highlight was that when they made voters rank how much they care about things, inflation was far and away the top (cached). Another wild one is that no matter how you partition the groups, everyone under 25 became more conservative (cached). Shor also found that not only did voters as a whole say they found Harris to be more liberal than they'd prefer, but also more democrats thought she was too Liberal (cached) than too Conservative. Finally, while we're on the subject of Shor, one critique of him is that turnout is more important, but the recent paper claiming that turnout is very important seems to basically say the opposite on a closer reading. And finally, if you hear claims about 14 or 19 million democrats having stayed home (cached) in 2024, that's also nonsense

Looking at how this functions, it seems like another example of improving LLM performance through increasing 'test-time' compute. Essentially, throwing more tokens towards getting the right answer. This is what was accomplished with early LLMs via telling the model to 'think step-by-step' and it's how OpenAI got big performance improvements with their o-series models that 'think' for a while before responding. This is just much more structured. I dunno if they actually used LangChain, but this is exactly what it's for.
If you want LLMs to get better at spatial reasoning, maybe this is the way to do it.

Longer Reads
– First of all, a lot of the fans of DOGE are absolute ghouls. Secondly, ending AIDS treatment for kids in africa can lead to deaths very quickly, median viral rebound is 22 days, and in almost 10% of cases it's acute. (src)(cached)
– AI tutors can work, not that explicitly doesn't mean "just give the student an LLM" (src)(cached)
– Drone warfare is already terrifying, but we are so close to kamekazi drones being given the ability to select their own targets (src)(cached)
– Anti baby led weaning take! (src)(cached)
– Zephyr Teachout wrote a critique of Ezra Klein's new book Abundance. Her critique of the YIMBY policies in the book are pretty dumb argue Stapp, Demsas, and Paul Williams. Stapp actually had some good things to say (cached) about Teachout's other arguments: Monopoly power really is causing some problems in American society that prevent abundance, but the biggest problems are stuff like electrical transmission and hospitals, not housing or tech (cached) (src)(cached)
– BYD says they've got 1000kW charging working, they say they'll be building 4000 of these chargers in China. (src)(cached)
– A 54-year-old US citizen was detained by ICE and held overnight, he happened to be carrying ID and his social security card which is probably the only reason they didn't try to deport him. (src)(cached)
– New York is tightening their enforcement of license plate requirement enforcement. But the $50 fine is still too little. (src)(cached)
– Another thread about how well NYC congestion pricing is working (src)(cached)