Astral Codex Ten

Mayan

9 min read

Mayan

Article

Mayan is a recurring concept in the Astral Codex Ten archive, appearing 2 times across 2 issues between June 03, 2021 and July 17, 2023. The archive places it in contexts such as “The population, agricultural output, number of administrators, and monumental construction at Mayan centers increased”; “uncovers hidden Mayan cities”. It most often appears alongside ancient Rome, Asia, Becatti 1968.

Metadata

  • Category: Concepts
  • Mention count: 2
  • Issue count: 2
  • First seen: June 03, 2021
  • Last seen: July 17, 2023

Appears In

Source Context

Recovered passages from the original issue text. When the raw archive preserved outbound links inside the source passage, they are listed directly under the quote.

Your Book Review: The Collapse Of Complex Societies
June 03, 2021 · Original source
After explaining his model of various types of diminishing returns on complexity eventually pushing societies to collapse, Tainter picks three states to apply his model to. He chooses out three civilizations at different levels of complexity – the Chacoan civilization (‘proto-state’), the Mayan civilization (‘definitely a state but not an empire’), and Rome (Rome). In 1988 the Mayan script had not been fully decrypted, and Chacoa didn’t have writing, so Tainter draws primarily on archaeology to put together pieces about what happened there.
For the Mayans, the strategy that brought them down was investing in a red queen race of population growth and agricultural intensification. There were several major power centers in competition, and they might have prioritized population growth as a matter of policy – when food started becoming scarce, male height started going down while female height didn’t, which some people think indicates that Mayans prioritized keeping women healthy for population growth reasons. (Tainter doesn’t say, but I assume they ruled out the hypothesis that the women were never that well fed to begin with, since it’s easy to check for – male:female height distributions under equal nutrition are well known.)
Inline links: red queen race
Mayans practiced similar kinds of agriculture, and hard times hit everyone at the same time. This was a major cause of population clustering around the power centers / states:
Contra The xAI Alignment Plan
July 17, 2023 · Original source
If it did work, it would be bad. I want to start by discussing the second objection, then loop back to explain what I mean about the first. A Maximally Curious AI Would Not Be Safe For Humanity The one sentence version: many scientists are curious about fruit flies, but this rarely ends well for the fruit flies. The longer, less flippant version: Even if an AI decides humans are interesting, this doesn’t mean the AI will promote human flourishing forever. Elon says his goal is “an age of plenty where there is no shortage of goods and services”, but why would a maximally-curious AI provide this? It might decide that humans suffering is more interesting than humans flourishing. Or that both are interesting, and it will have half the humans in the world flourish, and the other half suffer as a control group. Or that neither are the most interesting thing, and it would rather keep humans in tanks and poke at them in various ways to see what happens. Even if an AI decides human flourishing is briefly interesting, after a while it will already know lots of things about human flourishing and want to learn something else instead. Scientists have occasionally made colonies of extremely happy well-adjusted rats to see what would happen. But then they learned what happened, and switched back to things like testing how long rats would struggle against their inevitable deaths if you left them to drown in locked containers. Is leaving human society intact really an efficient way to study humans? Maybe it would be better to dissect a few thousand humans, learn the basic principles, then run a lot of simulations of humans in various contrived situations. Would the humans in the simulation be conscious? I don’t know and the AI wouldn’t care. If it was cheaper to simulate abstracted humans in low-fidelity, the same way SimCity has simulated citizens who are just a bundle of traffic-related preferences, wouldn’t the AI do that instead? Are humans more interesting than sentient lizard-people? I don’t know. If the answer is no, will the AI kill all humans and replace them with lizard-people? Surely after a thousand years of studying human flourishing ad nauseum, the lizard-people start sounding more interesting. Would a maximally curious AI be curious about the same things as us? I would like to think that humans are “objectively” more interesting than moon rocks in some sense - harder to predict, capable of more complex behavior. But if it turns out that the most complex and unpredictable part of us is how our fingerprints form, and that (eg) our food culture is an incredibly boring function of a few gustatory receptors, will the AI grow a trillion human fingers in weird vats, but also remove our ability to eat anything other than nutrient sludge? I predict that if we ever got a maximally curious superintelligence, it would scan all humans, vaporize existing physical-world humans as unnecessary and inconvenient, use the scans to run many low-fidelity simulations to help it learn the general principles of intelligent life (plus maybe a few higher-fidelity simulations, like the one you’re in now), then simulate a trillion intelligent-life-like entities to see if (eg) their neural networks reached some interesting meta-stable positions. Then it would move beyond being interested in any of that, and disassemble the Earth to use its atoms to make a really big particle accelerator (which would be cancelled halfway through by Superintelligent AI Congress). This doesn’t mean AI can’t have a goal of understanding the universe. I think this would be a very admirable goal! It just can’t be the whole alignment strategy. But Also, We Couldn’t Make A Maximally Curious AI Even If We Wanted To The problem with AI alignment isn’t really that we don’t have a good long-term goal to align the AI to. Back in 2010 we debated things like long-term goals, hoping that whoever programmed the AI could just write a long_term_goal.txt file and then some functions pointing there. But now in the 2020s the discussion has moved forward to “how do we make the AI do anything at all?” Now we direct AIs through reinforcement learning - telling them to do certain things and avoid certain other things. But this is a blunt instrument. Reinforcement learning directs the AI towards a certain cluster of correlated high-dimensional concepts that have the same lower-dimensional shadow of rewarded and punished behaviors. But we can’t be sure which concept it’s chosen or whether it’s the one we think. For example, there are many different ways of fleshing out “curiosity”. Suppose that Elon rewards an AI whenever it takes any curious-seeming action, and punishes it whenever it takes any incurious-seeming action. After many training rounds, it seems very curious. It goes off to the jungles of Guatemala and uncovers hidden Mayan cities. It sends probes to icy moons of Neptune to assess their composition. Overall it aces every curiosity test we give it with flying colors. But what’s its definition of curiosity? Perhaps it’s something like “maximize your knowledge of the nature and position of every atom in the solar system, weighted for interestingness-to-humans”. This would produce the observed behavior of exploring Guatemala and Neptune. But once it’s powerful enough, it might want to destroy the solar system - if it’s completely empty, it can be completely confident that it knows every single fact about it. Or what if it’s curious about existing objects, but not about nonexistent objects? This would produce good behavior during training, and makes a decent amount of sense. But it might mean the AI would ban humans from ever having children, since it’s not at all curious about what those (currently nonexistent) children would do, and they’re just making things more complicated. Or what if its curiosity depends on information-theoretic definitions of complexity? It might be that humans are more complex than moon rocks, but random noise is more complex than humans. It might behave well during training, but eventually want to replace humans with random noise. This is a kind of exaggerated scenario, but it wouldn’t surprise me if, for most formal definitions of curiosity, there’s something that we would find very boring which acts as a sort of curiosity-superstimulus by the standards of the formal definition. The existing field of AI alignment tries to figure out how to install any goal at all into an AI with reasonable levels of certainty that it in fact has that goal and not something closely correlated with a similar reinforcement-learning shadow. It’s not currently succeeding. This isn’t a worse problem for Musk and xAI than for anyone else, but there are a few aspects of their strategy that I think will make it harder for them to solve in practice: One good thing about order-following AI is that it’s useful now, when AIs aren’t agentic enough to have real goals and we just want to use them as tools in commercial applications. The hope is that we do this a bunch with GPT-4, then a bunch with GPT-5, and so on, and by the time we have a real superintelligence, we’ve worked out some of the kinks. I’m not sure how Musk’s maximally-curious AI helps do office work, which means there’s going to be more of a disconnect between current easily-tested applications and the eventual superintelligence that we need to get right.
Inline links: colonies of extremely happy well-adjusted rats, how long rats would struggle against their inevitable deaths if you left them to drown in locked containers