Abhishaike Mahajan introduction to organoids as models for drug testing, as a means for exposing interactions at both the cell and tissue level.
Matthew Green commentary on the Anthropic post on applying Mythos towards the topic of cryptanalysis. Somewhat related, Vitalik Buterin explainer on the Diamond indistinguishability obfuscation scheme, and some related commentary by Georgios Konstantopoulos.
Anton Leicht in Asterisk Mag continues his attempts to get the countries of the world to consider what will happen to them in the scenario where AI is: firstly, a big deal, and secondly, not commoditized1. It’s somewhat confusing to me how few people are speaking on this issue to the extent that exhortations are still only being delivered to “middle powers” as a whole. Unfortunately, there are clearly some selection effects at play here, as there is a relatively simple method of downside mitigation that individuals can play which doesn’t require persuading or waiting for their countrymen to reach similar conclusions. This, more than anything else, is why I’m not very optimistic on sovereign AI as a solution, but unlike Anton, it seems to me that it still must be tried, if only as a justification and means to increase the size of the coalition necessary for investing in domestic compute. The problem with dismissing any possible intervention is that the current position is such that no strategy is very likely to suceed: that compute can become obsolete, resources be fully exploited or substituted, supply chains can be transferred and automated, and protectionist measures can be eroded2. In which case, more risk-tolerant model of capital allocation will be necessary, alongside interventions which can open up the possibility of positive second-order effects, such as in AI-assisted improvements to governance.
Zern Hee on Chinese zombie companies as a side-effect of the Chinese industrial-policy model of decentralized local investment, which leads to cities and counties being unwilling to cut off losers who are providing local employment.
Dan Williams review of A Conflict of Visions. My own view on the question of whether human nature is good or evil is that it’s a sort of category error, in that outcomes (and perhaps individuals) can be good or evil, but human nature is an aggregate description of how a distribution of individuals interact with their current environment, and therefore an emergent rather than fundamental phenomena. In which case, the idea that evil must be constrained or that good can be let free is less a metaphysical description about human nature than an empirical description of whether any particular combination of population and environment is capable of achieving one’s desired outcomes3.
Nathan Goldwag commentary on the historical accuracy of Nolan’s The Odyssey, which serves as good contrast to Emily Wilson’s more critical review.
Snowden Todd gonzo journalism of his explorations in the Caucasus.
Naomi Kanakia review of the works of Ben Lerner, as “a highly-acclaimed man of letters who’s written four novels about the daily life of a highly-acclaimed man of letters”.
GOT The Mad King Updates on the theories that have been confirmed by The Mad King stageplay.
Abundance and Growth Blog linkthread.
Dwarkesh has some predictions about the price of compute which seem to be somewhat backwards, reasoned as it is from what the labs themselves would prefer to do. But if the market decides otherwise, then revenue will no longer 10x and compute will no longer 3x, in which case labs will indeed have to spend an increasing fraction of costs on inference. It’s only in scenarios such as those where open-weights models are insufficiently competitive that revenue can continue to expand faster than compute. Otherwise, even algorithmic improvements in inference efficiency will only lead to Jevon’s style increased demand, which will indeed increase the price of compute (in addition to reducing lab margins; in this respect, Dean Ball’s comment that open-models will slow down research progress seems to me to be correct). But these are all mechanisms which are ultimately determined by market demand and investment, and not really by what the labs themselves might prefer.
The primary way I’m modeling the effects of AI disruption to national economies is as a digital analogue to China shock, which like China shock is not a single event, but a long-term phenomena which expands continuously in all directions. Tangentially related, Feyi Fawehinmi on how Chinese car exports into Nigeria are eroding the information assymetry and margins of Nigerian middlemen.
On the topic of imposing one’s preferences onto everyone else, Matt Bruenig has a post which seems to me to be entirely missing the point in interpreting “moderation” as equivalent to finding a “median” position, since moderation has nothing to do with medians but rather the practice of only advocating for policies which a majority of the population is actualling willing to support, both rhetorically but also importantly in practice. One can’t say that moderates don’t practice what they preach: here’s Matt Yglesias, upon learning that the majority of the population supports free markets but not capitalism, arguing for more free markets.
While on the topic of politics, here’s a Helen Pluckrose article on “calls to violence”, Brian Chau against protectionism, and Scott Sumner linkthread.