2025-07-12
mesoarchean
Jasmine Sun interviews Daniel Kokotajlo. I attended one of their wargames, but if anything, it’s made me even less confident in the accuracy of their AI 2027 predictions1. At Manifest, there were a couple conversations about which of the attendees was the best predictor, and generally the consensus was that accuracy tends to highly domain dependant, with one’s calibration being highly dependent on the domain one is forecasting in. Related, here’s an interesting interview by Alethios with Nuño Sempere of Samotsvety: there’s an interesting discussion about different types of forecasters, and it’s relevant to note that the AI 2027 team seems to operate in the quant-style modelling category, which suffers particularly when data is scarce.
Mechanize (Epoch AI) has an interesting post on how solving difficulties around reinforcement learning environments are the current bottleneck around further scaling. I wonder if we’ll have something like Bayesian methods for reinforcement learning rewards, but this might be a case where creating the metamodel for assigning meta-parameters is just as complicated and high-dimensional as actually determining what the correct weights should be in the first place.
Packy speculative tech linkthread. Particularly interesting is a graphene oxide “taster”, which has implications for my theory that food will be elevated in status as an art form in the era of AI because cooking and tasting will be difficult for AI to handle autonomously.
ACX not-a-book review on the amyloid hypothesis for Alzheimer’s disease.
Alex Sorondo vignettes about writers and money. I probably go farther than most in separating the artist from their art, but one area where I do find it harder to do is when someone is writing advocacy for policies which seem counter to their own lifestyles. One plausible explanation is that there’s limited ways to write from wealth which are popular these days, which is presumably why many still act as if they are starving artists even after making it. But it might actually just be force of habit, in that you develop your style in your youth and the experiential scaffold feels missing or even false when you try to adapt it to your newfound circumstances.
Freddie de Boer with notes on the top magazine writers today.
I’m also pretty skeptical about their AI-mindreader idea, because humans are actually pretty good mind-readers already, and even if you don’t buy Moravec's Paradox, you can solve this by selecting for political leaders for sociopathy.