2025-06-03
endless summer
Dwarkesh predictions on AI timelines, with a similar intuition to my own that some fast equivalent of RL finetuning is necessary for strong-AGI to develop. I don’t share his belief that this is due to lack of online learning ability, which appears already at least partially doable, but is limited by a lack of a proper developer UX. Rather, if my current belief about what RL is actually doing is correct, then the issue is actually that there is a tradeoff between generality and task-specific performance for any particular model size: this indicates that the demand for compute may increase massively since, in addition to the labs needing to continually increase model size for general capabilities, users now also need to be able to perform custom RL for their specific use-cases.
Robin Hanson on Abundance liberalism (not really). Just what comes to mind after reading Noah Smith’s piece on political tactics for the Abundance movement. Anyway, I agree with Robin that strong governance could lead to more variation, particularly if the way we get strong governance is something like Archipelago. But the claim that “in the past, weak governance may actually have been a big reason why we’ve had so much competition, which has induced so much cultural evolution” seems to be conflating a lot of different types of competition which may or may not be all related. Related, Ryan Puzycki, in reference to the current slowdown in growth in Austin, notes that narratives based around short timescale trends are often not very useful in terms of explanatory or predictive power, and suggests taking a more holistic and long-term perspective. Given that Austin is a YIMBY success story, what ties them together is this observation by Richard Hanania on how the Abundance debate is playing out in public in the format of various policy debates.
Nicholas Decker complaining that private companies are the new public companies. Also seems bad to me.
Yaw on the 1991 Indian gold pledge with the IMF.
PSmith review of the Cambridge Latin Course, and it’s attempt to synthesize the grammar-translation and natural methods of learning the language. It’s interesting to see the relationships to how other Romance languages are commonly taught: it’s my impression that French curriculums tend somewhat more towards grammar-translation, while Spanish pedagogy is almost entirely natural (not sure about Italian, as I’ve never attempted it). I haven’t tried learning Latin, but from this review it seems to me, even accounting for the focus on the written text, that this might be a case where the older tradition could actually benefit from seeing how it’s more popular inheritors are doing it.
Dylan Black with some additional comments on measure theory and fine-tuning.
Astera Institute announces intention to use preprints in lieu of scientific journals to publish their results, and suggests tools to help transition science to a journal-less future.
Dan Frank suggests that more people should have linkthreads.
Niko McCarty linkthread.
Sam Harismony linkthread, including information on his orexin trial for sleep need reduction.
