https://www.thepsmiths.com/p/review-aztecs-by-inga-clendinnen the Aztecs were scary in a unique way. it wasn’t just “people in the past were more brutal than us”, it was a sophisticated theology based on cosmic horror that demanded a really unusual degree of human sacrifice and suffering.
https://contraptions.venkateshrao.com/p/walter-mitty-effects-in-ai-incident one of Venkatesh Rao’s sloptraptions. it’s not a bad idea, and the LLM helps unpack it. the “unpacking” is more voluminous than I really need, but also I expect the amount Rao would have patience to “unpack” on his own is less than I’d need to get the idea. It’s also, like, just a mundane take, not really better or worse than other people’s takes.
this implies that (under certain nice conditions) the Fristonish worldview where beliefs and desires are undifferentiated is in fact wrong
i have a related but not identical intuition that “nondual optimization” (as in, simultaneously trying to get the “whole world” into a state you want) is inefficient relative towards having a “factorization” of everything into a self-model and the range of actions available to oneself.
Rosas doesn’t get into the self-model part but he does indicate that if you couldn’t “factor” beliefs and desires into separately computable independent parts, that would imply a type of optimizer with “less nice” properties
you have the left-wingers who defend them as cruelly misunderstood victims of Spanish colonialism. That latter impulse is what results in videos like this one, an AI-enabled imagining of Tenochtitlan in 2026 if the Mexica had fended off their conquerors: [...] What’s remarkable about this video is that it’s more of an act of colonialism than anything the Spanish ever did. The Tenochtitlan it imagines is just Brooklyn but with step pyramids. Everything distinctive and disturbing has been airbrushed away, a kind of cultural erasure more profound than any book-burning, and no less destructive because it was born out of a patronizing condescension rather than out of indignation. It imagines the Mexica, as the victors, adopting our worldview and norms whole-cloth. This is because it cannot imagine anything outside of itself.
The obvious proposal is to design a realistic AI video based on actual Aztec beliefs and behaviors, only extrapolated to use modern technology. (YouTube would probably censor it.)
links 9/1/26: https://roamresearch.com/#/app/srcpublic/page/09-01-2026
https://agreed-facts.vercel.app/ a collaborative fact-editing site for debates
https://www.thepsmiths.com/p/review-aztecs-by-inga-clendinnen the Aztecs were scary in a unique way. it wasn’t just “people in the past were more brutal than us”, it was a sophisticated theology based on cosmic horror that demanded a really unusual degree of human sacrifice and suffering.
https://contraptions.venkateshrao.com/p/walter-mitty-effects-in-ai-incident one of Venkatesh Rao’s sloptraptions. it’s not a bad idea, and the LLM helps unpack it. the “unpacking” is more voluminous than I really need, but also I expect the amount Rao would have patience to “unpack” on his own is less than I’d need to get the idea. It’s also, like, just a mundane take, not really better or worse than other people’s takes.
https://www.lesswrong.com/posts/awMDNhoL6J97s6wFJ/the-separation-principle-where-do-beliefs-and-desires-come Fernando Rosas has a really nifty conceptual framework for two intuitively “agent-y” properties: beliefs and desires. if you have a more general sort of control process (an LQG regulator, or a POMDP), you can sometimes derive a “factorization” into independent parts that map to beliefs and desires.
this implies that (under certain nice conditions) the Fristonish worldview where beliefs and desires are undifferentiated is in fact wrong
i have a related but not identical intuition that “nondual optimization” (as in, simultaneously trying to get the “whole world” into a state you want) is inefficient relative towards having a “factorization” of everything into a self-model and the range of actions available to oneself.
Rosas doesn’t get into the self-model part but he does indicate that if you couldn’t “factor” beliefs and desires into separately computable independent parts, that would imply a type of optimizer with “less nice” properties
From the article about Aztecs:
The obvious proposal is to design a realistic AI video based on actual Aztec beliefs and behaviors, only extrapolated to use modern technology. (YouTube would probably censor it.)