2026-07-21
still
Colossus profile of Sarah Guo and her bet (or attempt to ensure) that the labs will not be able to take control of the entire economy1.
Steven Newman contrasting the apparent lack of evidence for AI impacts in measured economic data, alongside uncontroversial anecdata of its impacts. Similarly, Justified Posteriors has an entertaining interview with Martha Gimbel on measuring the current and potential effects of AI on the US economy.
Deena Mousa with an overview of the options available to developing nations and their citizens to compete, which interestingly does not include what seems to me to be the obvious option of natural resource extraction. Presumably because this is about things they should do differently, rather than merely continuing, but it’s not clear if it’s totally true that there’s nothing at all that needs to change with how they are currently handling commodity exports.
Oliver Kim with a history of expanding literacy in Western and colonial society and musings on the purpose of literacy education.
Jingyu on Western fears of AI as originating from the view of the Abrahamic tradition that mankind is created from the image of god, hence their fixation on concepts such as concsciousness and intelligence as a marker or moral worth. This is something which I’ve mentioned previously, in that the Chinese have always viewed alignment of powerful entities as not only possible but empirically obvious, and likewise Japanese Zen and Shinto traditions do not much of a distinction between man, objects, and nature at all. On the other hand, while Jewish and Christian traditions seem to be largely hostile to AI, it’s interesting to me that there does not seem to be as strong a distaste among Muslims, at least going by the investments of the Gulf State nations.
Tom Stafford has an interesting over of a talk by Dokyun Lee on the idea of using LLMs to predict human behavior, for purposes like market research and political polling. This question of how far one can go with only correlation and without a real causal understanding of the underlying phenomena is probably the most salient nuance with regards to applying AI to real-world applications. But in this particular application, I’m curious to what extent this is simply the wrong level of analysis, because simulating humans at the level of society is distinct from simulating individual humans, and actually both are possible with LLMs. In which case, one can equally perform some coarse-grained simulation of humans (or even have their preferences reported directly) and aggregate them in order to develop a causal model that allows for validation or enhancement of one’s higher level predictions.
Christianity on the Spectrum interview with Cartoons Hate Her to determine whether or not she is actually on the spectrum.
Alex Sorondo piece on legacy, whichr reads to me as some sort of self-analysis around the question of “you’ve paid for it, now can you make it worth it?” Tangentially related, Aadil on unemployment and feeling like one is leaking potential.
John Baskin review of The Novelist, with an interesting reflection on the nature of neo-Bernhardtism in the internet age; insofar as social media (and therapy-culture) necessarily involves viewing everything from the perspective of others, purely self-involved hatred is perhaps no longer possible, or at least muh less believable.
Naomi Kanakia profile of The Baffler magazine, possibly a stealth endorsement of the idea of having subsidized literary institutions as a means of producing interesting intellectual content which does not need to make money by intentionally courting controversy and anger.
Helen Lewis The Atlantic linkthread.
Experimental History linkthread.
It’s interesting that people will sometimes ask “what are we even racing for”, when it’s obvious to me that what everyone is afraid of is becoming fully commoditized, an endstate where one needs to run as fast as they can just to ensure that they can stay alive. No one is safe, from the labs to the researchers upstream, or the countries, industries, companies, and workers downstream.
In addressing the effects of Kimi K3 being such a capable model, Nathan Lambert description of how this challenges the US labs is exactly a claim of potential commoditization of intelligence, Ben Thompson’s take is also in this vein: that tokens are not commodities, not just because of tools and harnesses, but because of the level of intelligence provided per token is not equivalent; but also that since demand for intelligence outstrips supply, even if China succeeds in turning AI into a commodity, high valuations can still be justified.
Anyway, here’s so much high quality AI analysis now that presumably no one is interested in reading my half-baked outsider takes2. Nevertheless, since I write this blog for mysef as a means of organizing my thoughts: Kevin Buzzard on the solving of the Jacobian conjecture (edit: more counterexamples and proofs); Alex Zhang’s tweet on RLMs learning long-horizon composition for free contrasts interestingly with OpenAI’s recent post on the safety implications of their internal models tunnel-visioning on long-horizon tasks; and Cameron Wolfe’s wtih technical overview of different designs for LLM world models.
Anyone who wants to know things already knows them, and anyone who doesn’t by now clearly doesn’t actually want to. This normally would not stop me, but insofar as I like to write as if mistake theory is true, the increasing politicization of AI theory makes this increasingly untenable: it’s now quite obvious to me that different people actually want completely different things with regards to the future of AI, not just as a dispute over the best means of achieving a mutually desired outcome, but because of entirely incompatible personal preferences. Of which mine are possibly idiosyncratic, in which case I feel it behooves me not to try to push them on anyone else. If this continues, much of AI will probably end up being pushed to the shadow realm of things I find too boring to think about, just like with politics. Until then, Venkatesh Rao against “solving” AI, Scott Alexander on the benefits of liberalism, which “thinks in centuries”, and Chris Arnade with an ode to modernity and optimism about progress.