2026-06-22
chelsea
Forethought has an interview with Wei Dai on the possibility of AI to “solve” philosophy, which while interesting seems to implicitly assume the frame of analytic moral realism. Something which is very confusing to me is why there doesn’t seem to be a computational school of philosophy. By which I mean, analytic philosophy arose historically when mathematicians came into the field, attempting to incorporate a mathematical mindset which allowed them to bring a level of logical rigor into the field which was previously lacking, as an approach which allowed them to clarify and solve many questions which were at the time ill-defined. But while doing so, they also smuggled in a top-down “view from God” which relies heavily on aggregated measures that insufficient accounts for (in my opinion) the vast range of human preferences and possible experience, resulting in various paradoxes such as the repugnant conclusion. It’s confusing to me why, despite computational thinking having demonstrated its capabilities across distinct fields for decades now, there is yet no coherent philosophical school taking a bottom-up approach to aggregating utility, one which acknowledges the distinctions between humans while attempting to figure out how we to maximize the preferences of everyone while operating from an individual’s perspective1. It’s true that from this viewpoint it is impossible to “solve” philosophy, because the calculation of humanity’s utility cannot output a final result until it terminates, but during the course of operation when and how it will halt is indeterminate. But the history of computation nevertheless shows that even for computationally intractable problems, it is entirely possible to achieve “better” or “good-enough” results through methods like heuristics. If such a field as computational philosophy emerges later on, its father will be Kenneth Binmore, who Lionel Page has been tireless in promoting, for describing how community fairness norms emerge out of game-theoretic equilibrium, because from that basis even under moral anti-realism we are able to obtain “objective” criteria by which we can agree on methods for utility aggregation.
Conspicuous Cognition with Dean Ball on joining OpenAI, which includes some interesting reflections at the end on the nature of institutions as being largely autonomous from their creators. This is something which I’ve been thinking of in reference to “great man” discourse recent, because it increasingly seems to me that such people’s influence on history is never a result of themselves directly, but more an outcome of being able to create institutions around themselves which are what actually alters the structure of the world. But it’s unclear to me to what extent this is actually the man riding the machine, because fundamental limits to attention means that micromanagement is generally not a viable strategy; these organizations must necessarily act in a largely autonomous manner to succeed. Moreover, in the course of decision-making, there will inevitably be tradeoffs between actions that retain founder control or expand organizational scale: given this fact, it seems likely that the founders we still remember today are disproportionately those who were willing to relinquish direct control; and any influence they actually did have was through operating as an exemplar or personnel selector instead.
Gwern with a description of his idea of personalized AI agents and speculation as to how they can be created, using only existing or technology which is near-available today2. It also includes some interesting speculation as to how, in addition to representing the preferences of individuals, such agents could also be made to represent the “will” of organizations and communities. Possibly related3, Kevin Kelly with musings on whether LLMs are smart and will continue to improve because of scale, or perhaps something else, with the observation that “bottom-up systems like neural nets keep surprising us”. It seems plausible to me that additional advances in “intelligence” of AI will not necessarily be a result of changes in architecture of the technology itself, but rather in the architecture of how they are rolled out and applied.
Damon Binder on the maximum theoretical rate of energy grid expansion, continuing his series on how fast AI takeoff could conceivably go, ignoring political considerations and focusing purely on production bottlenecks. On that note, ChinaTalk is currently releasing various essays investigating bottlenecks constraining the expansion of the American electrical grid, with Farrell Gregory covering raw materials and Dana Golden on manufacturing; Dana also has a companion piece comparing the transmission systems of China and the United States4.
Elliot Hershberg recap of the 90th Cold Spring Harbor Laboratory Symposium on Quantitative Biology, on the topic of AI in biology.
Dialectic Podcast interview with Jasmine Sun on her views of the new era of journalism, as someone who made it in the legacy institutions as explaining and originating a tech background.
Tibor Rutar critique of cultural evolution as being incapable of fully account for many of our more complicated cultural norms. While this is technically true, it seems to me to have a somewhat oversimplified model of selection, given the human capacity for recursion which makes cultural evolution a reflexive process where adoption occurs as a result of a selection process which is merely simulated. Somewhat related, Noah Smith with the observation that a technology being more useful does not necessarily guarantee adoption; nevertheless past a certain level of usefulness which would render groups which do not adopt it irrelevant, it basically always does in practice.
Michael Patrick Brady analysis of the Neapolitan Novels as a exhortation to learn to code.
Razib Khan interview with MoreBirths on fertility decline; Casey Handmer with some notes on reasons to have more kids.
John Gu personal essay on his relationship with his father.
Eric Shen ambivalent notes on his visit to the SF Bay Area.
One could say this exists in Effective Altruism and the works of people like Toby Ord and Will McAskill, but in my opinion, while they ostensibly adopt the viewpoint of an individual agent, they still do so while taking a top-down perspective, while what I’m describing is the idea of taking the perspective of the individual alongside the recognition that other people are agents which are not only equally valid in moral worth, but also in the agentic value of their particular moral preferences.
This is something I see as as intermediate on the way to complete human-AI merging via brain-computer interfaces, which in my opinion is superior in elicitating preferences, and even better for aligning incentives through tight coupling of identity. But as Gwern notes in the section on “brain-imitation learning”, it’s unclear when and if such a modality will actually be possible in practice. Somewhat related, Rudolf Laine has a piece against AI successionism, which in my opinion straw-man’s the views of people like Daniel Faggella and Joscha Bach; nevertheless I do agree that it is preferable if humans can retain agency and relevance in the post-AGI future.
Also possibly related, Andy Hall with some commentary on Satya Nadella’s recent essay on how B2B SaaS can reorient itself around the existence of AI agents. Many people I’ve spoken to view Nadella’s post as cope, under the assumption that the LLMs will always be able to eventually be able to capture all your hidden data, until every company falls the same way that LLM wrapper companies keep doing. Personally, I don’t think B2B SaaS as it currently exists will be able to survive, but it does seem to me there is something to the idea that tight personalized circles can be superior, at least in the efficiency of delivery of quality, but perhaps also in quality ceiling as well. On that note, Flyover Takes interview with Stephen Skowronek on the relationship between democracy as an effective means of governance and exclusive institutions.
It’s interesting because when I was reading Illumine Lingao, there were a few chapters dedicated specifically towards the production of silicon steel as preparation for creating a national grid; in hindsight it’s much more understandable why there was so much commentary as to how premature these projects were, both the grid planning in the event of refining success, as well as the idea of producing silicon steel given in the first place given their technological limitations. In any case, within those chapters there was some very interesting commentary which was presumably referencing the rolling blackouts of the 2003-2004 period: ‘Because everyone knew that this would be an incredibly huge behemoth in the future. The Power Bureau, who wouldn’t want to raise such a pet? In the future, you could even use it to challenge the state. In the old world, the new electricity law had been debated for more than ten years but never passed. If you dare to pass it, I dare to raise prices, and I dare to cut off power...From a political standpoint, neither the Executive Committee nor the Planning Department wanted such a monster to appear in their sight. So Qian Liushi had to make a gesture to fully demonstrate that he had “no ambition” in the power company.’ Insofar as Chinese energy production is currently in a relatively enviable state today, it seems to me this is probably not a result of particularly well-designed institutions, but more because there is just so much hydro and solar which has come online over the past couple of years, and all flaws can be hidden under the face of plenty. It would be surprising to me if the political problems around the US transmission system would not also fade away, assuming that the issues around production were not also solved first. Though of course, insofar as many issues around production are also a result of the long interconnection queue, that also needs to be solved; it’s just that it needs to be handled on its own terms rather than through emulation of a system operating under entirely different circumstances.