2025-12-30
137
Jaime Sevilla writes for Epoch AI on the plausibility of decentralized training. On the trend of scaling inference as the cheapest way to get around bottlenecks in human supervision and bandwidth, I wonder to what extent one can also use inference to get mitigate various issues of human feedback such as sycophancy, to unlock training on model outputs in real-time and at scale. Sevilla suggests that frontier companies are unlikely to pursue decentralized training as they have access to superior centralized data centers, but it’s possible that these techniques will actually be what allows them to finally make productive use of their significant pool of free users.
Phillip Trammel on capital accumulation in the post-AGI age. It’s interesting because in my opinion, the strongest argument for meritocracy in division of resources is that incentivizing the most capable individuals is the most effective way to raise both total and baseline utility over the long run. But if robots end up doing everything, then this no longer applies, and diminishing returns to wealth suggest that strict egalitarianism becomes superior: what naturally follows will be significant wealth and inheritance taxes. As a result, it’s actually only by preserving human agency and social mobility that developers of AI will be able to justify amassing disproportionate gains from what they’ve developed.
Ruxandra Teslo writes in Asimov Press on how clinical trials provide experimental data which can be used for refining medical treatments, even or especially in the case of failure1.
Trevor Klee on how even successful medical interventions produce decline. Somewhat related, Zhu Yutao translation of an interview with Gordan Liu on preparing the Chinese medical system for the aging population. It’s interesting that even as Chinese companies are the primary suppliers for grey market peptides in the US, there doesn’t appear to be much interesting, official or otherwise, for GLP-1RAs as of yet.
Dr. Psych Mom has an discussion with Lana Li on dating, which is unusual among dating discourse in being both frank and nonjudgmental. While on the topic, there’s also an interesting discussion occurring on Twitter today on to what extent dating is something only Americans do. Also, Jennifer Wilson writes in the New Yorker on the increasing prevalence of the prenuptial agreement, and Uri reviews I Just Want This Done, “supposedly the best book on getting a divorce”.
John Boyle in the Center for Education Progress on case studies of accelerating gifted children.
Alexander Sorondo in the Republic of Letters on working low-skill service sector jobs2.
Chinese Doomscroll translation of a lecture by Yuan Tengfei on the launch of the Cultural Revolution. Related, Li Mingguang on the work of Gao Hua (via Tanner Greer).
Zhengdong Wang, the annual letter successor of Dan Wang, recaps 2025 in AI and travel.
PSmith metareview of the current PSmith industrial complex.
No Magic Pill linkthread.
The distinction between designing better drugs versus testing more reminds me of the arguments over print statement debugging which keep recurring on Hacker News. As a fan of print statement debugging myself, it’s worth pointing out that even if we had virtual cells, that still would not be equivalent in fidelity to modern debugging tools.
I suspect that part of why I was in such a hurry to retire as soon as possible was exactly because my initial impressions of work were not particularly enjoyable: a paper-route, where you needed to ring doorbells to collect; fast food, where your shift could be unilaterally extended by rain-induced crunch; or cash, and the judgmental assumptions and jokes of “superior” customers. And that’s as a kid; I don’t want to imagine having to do these as an adult, just for minimum wage.