2026-07-01
pete standing alone
Elizabeth van Nostrand in Works in Progress on the history of the theory of plate tectonics, and how it became scientific consensus with the gradual accumulation of supporting evidence. Somewhat related, Lennart Finke in Asterisk Mag on the limitations of randomized control trials and the possibility of using statistical methods to correct for confounding in observational studies. I read both articles as responding to the question of to what extent AI will lead to an evolution in how science is done, beyond the idea of the Kuhnian revolution into something more akin to an engineering process of applied statistics. When the cost of data is high, then paradigms perform an essential role as a heuristic for which avenues are worth pursuing; but if data becomes sufficiently abundant, then they could conceivably fall to the bitter lesson and become mere retroactive justifications of what already exists. More likely, it seems to me what’s really happening on the brink of change is a realization that we have always been somewhere in between these two systems: uncovered evidence seeding theory; theory guiding evidence search1.
Somewhat related, there is an interesting episode of Evolutionary Psychology podcast with Lauren Ross on the concept of casual mechanisms in the philosophy of science. I would like to join Ross in her attack on causality as being strictly hierarchical, and the description of mechanisms as descriptions of a field being described a a more fundamental lower level one. The more salient part of “causal mechanism” is to me the former, in that it’s the identification of direct versus indirect causes which allows one to distinguish between causation versus mere correlation. It’s worth once again referring to Scott Aaronson’s recent speech around the greatly neglected philosophical implications of theoretical computer science and the statistical mindset, including the idea that “intelligence is fundamentally about prediction, and prediction is fundamentally about compressing your training data.” But crucially, I see prediction as more than simply the standard definition of what is most likely to happen in the case of inaction, but also the more expansive ability to predict the events of every action, such that for any desired outcome one is able to identify the necessary interventions to make it happen. With regard to the former, the compression of the correlations produces generalization; for the latter, the identification of direct causal mechanisms; both are naturally obtained out of the heuristic of usefulness2. It’s therefore self-evident to me why the physical sciences have always been able to perform so well while the social sciences are constantly in a state of struggle, with social engineering having become taboo as a reaction to the scars of the 20th century. While perfectly understandable, it seems to me that social scientists have become somewhat overgeneralized in their correction, because engineering does not necessarily need to be performed at the level of the state, but insofar as a better understanding of human nature also improve one’s ability to navigate human society, can also be done at the level of the individual3. As for worries about the potential second-order effects of providing people with access to such information to society as a whole, this seems to me closer to what one would historically have called social engineering, with self-appointed guardians making discretionary choices about which pieces of knowledge they would like to hide or make available in order to guide everyone towards their own chosen directions. Personally, I do not have the arrogance to presume that I can perform such calculations with sufficient accuracy that it gives me the right to make these decisions on behalf of everyone else. And even more confusing is how one can even reconcile the idea of doing science itself, without the implicit assumption that knowledge is not only truth and beauty, but fundamentally good.
ABB with an introductory overview of the field of economic history, as a study about counterfactuals in history as a means to uncover causal mechanisms in sociology.
Liam Robins introductory explainer on the von Neumann-Morgenstern utility theorem and distinguishing between the typical descriptions of utility between philosophy and economics4. It’s actually a perpetual source of confusion to me why it is that it’s the economists which have (in my opinion) the far more philosophically sophisticated definition of utility as compared to philosophers.
Scott Alexander with a delightful piece on childrearing which is simultaneously a description of his parenting style, as well as an extended meditation into the inheritance of personality and belief, which ties together with a comparison of the relationship between parent and child as between Yahweh and the Jews, as bargaining for the creation of binding covenant with a higher power.
Felice on writing as something one does rather than is5.
Matt Glassman notes on his visit to London.
Kristina Fort commentary on Europe 2031 as a not particularly serious proposal (+ linkthread).
Joseph Levine polymath letter and linkthread.
Scott Sumner links and new movie reviews; Noah Smith for more economics links.
Georgia Berg celebrates Canada Day with some very serious proposals for the improvement of the country.
As such, my interpretation is that nothing is actually changing here with AI-driven scientific research, at least with regards to how science works at an abstract level. But from a practical viewpoint, there is an important distinction that understanding is necessary for the historical case of theories being originated by humans, which may not be the case if theory originated is reduced to something more akin to statistical best-fitting and left to the AI. But while paradigms do indeed become entirely irrelevant as a concept from the perspective of humans, under the curtain the loop remains the same.
Edit: There are some criticisms of Finke’s article which I would like to defend him against, because it seems to me that they are somewhat missing the spirit of his argument, which isn’t really that more data and more covariates will solve everything, but as a description of when RCTs should actually be implemented in practice, given their cost. It’s true that unless one is able to construct an accurate causal graph of all the variables, the method he describes could end up doing things like conditioning on colliders. But it should be noted that, generally speaking, causal graphs are not being chosen at random; rather we generally already have some degree of confidence as to which causal graphs are more or less plausible, having generated them based on convincing theoretical paradigms. Given this fact, it should be possible to provide an estimate of the expected value of an intervention assuming that one’s model is correct, versus the alternatives of intervening wrongly, not intervening, or intervening or not only after running expensive RCTs. Or alternatively, running this calculation in reverse to determine the level of confidence in one’s paradigm which is required in order for the expected value of an intervention to surpass some target threshold, like the cost of an RCT. And while it’s true that data collected post-intervention would also be merely observational, in practice this threshold could only be reached if one expected the effect size to be very large, in which case any ambiguous outcome would serve as convincing evidence of failure. So while it is true that RCTs are still be required in many cases, such as when the effect size is not particularly strong, it seems trivially true to me that “we can trade more computational power for better solutions” and that “the overvaluing of interventional evidence disproportionately harms public health, especially in low- and middle-income countries, where an unconscious choice is made between considering observational data or no data at all.”
Also why I view dismissals and refusal to contemplate AI by some philosophers as not only myopic but fundamentally ungrateful, since it’s exactly because of the rise of artificial intelligence as a testing ground for philosophical theories of mind that philosophy has recently been so interesting and therefore in vogue: the age of AI is the age of philosophy.
One can object “what about p-hacking and the replication crisis in exactly this kind of self-help sociology”, which is fair, but on the other hand it sort of proves the point, in that the best way one can obtain verification as to the accuracy of a theory is to actually try it out. It seems to me that there are many such theories in social science fields which are obviously wrong but which have still managed to hang around, exactly because they avoid creating propositions which are usable (and therefore testable) by individuals.
Tangentially related, Minds Almost Meeting has an episode discussing The Philosophy of Money. While it is indeed incorrect to assume that prices fully capture value, since preferences are individual but money is coordinated, but most of objections described by Callard here do not seem particularly compelling, and to me largely confused.
While on the topic of AI changing the world and writing, for some reason three different AI-related essay contests have announced their winners in quick succession: Astera Institute on AI effects on metascience; Dwarkesh on the big questions; and Hyperstition AI Fiction, for which Henry Oliver has some mixed comments.
I was expecting Jim soup