in the image of their tools
19 Nov 2026first of all, what drew me in was probably the title. it gives off this whiff of anthropology. it’s probably just the context i live in now, rereading The Dawn of Everything.
on the face of it, the argument in this Science essay is simple and pretty extreme. the new tools we have are so important, so valuable if used correctly, that it’s worthwhile to redesign institutions and processes at the level of society to take advantage of them. pretty radical.
i think there’s a lot of validity to the claim that we need to invest more in figuring out how to better use new technologies (in general, and with AI specifically). it’s a familiar point, and it’s usually less exciting than developing new technologies (at least for me). but it touches on real problems. i bet it warrants real research and actual thought.
overall i think the thesis of this piece is kind of backward (and the title does a pretty good job of highlighting it). :-) working out how a society was organized from the tools it used sounds more like archaeology to me. if we’re going to redesign an institution, i would argue that the best question to ask would be what we want it to do. then we can work out what structures would support that.
they claim that “institutions whose structure matches the era’s cognitive infrastructure should outperform those that do not.” so what’s this era’s cognitive infra? what was the previous one? what does it mean for an institution to ‘match’ it? they give a few historical examples, including ICANN and India’s payments system, without going into much detail. and then there’s the familiar story about how the first electrified factories got little out of electricity until they changed how the factories were laid out.
then they go on to describe some traits of neural networks, including distributed representation, attention (the relevance of each input to other inputs), and credit assignment through backprop.1
they talk about mechanisms like Polis and Community Notes on Twitter without tying them too tightly to AI. Polis is a pretty cool system for finding where people agree, and it’s been used to inform actual decisions, from regulating Uber in Taiwan to designing local services in Finland.
this, at least, is a somewhat interesting take. they suggest science should give findings credit for how surprising they are and how much they change what a field believed. i’d have liked to see them spend more time on that.
i realize it’s a lot easier to criticize than it is to come up with interesting, good solutions and ways forward. but these are hard problems (and ones i think are worth tackling individually, per institution, per field, etc.). i don’t see these blanket principles applying in the same way in science and in law, for example. i think the solution starts with getting the people who understand how to use the technology in their fields more involved. governments and large civil organizations should borrow some of that expertise too (see the UK AI Security Institute, for example).
-
does this analogy even hold? i usually think of credit assignment in the context of RL. what exactly is being rewarded here when they talk about backprop? ↩