Followup to: Rationalists are missing a core piece for agent-like structure (energy vs information overload). Related to: Linear Diffusion of Sparse Lognormals (I don't consider that sequence done yet BTW, I just took a break because I ran out of pre-written posts, so I need to write some more). Response to: Fixing The Good Regulator Theorem.
Where I last left off, I argued that an "energy"-orientation may permit a more efficient set of models that can zoom in on what really matters. But I didn't really go into much detail about what to expect or why, so I think I should try harder, in particular framing it more like a potential selection theorem.
The universe has certain limited resources - for instance, mass, energy, space, time, and similar (money?). Anything that happens within the universe makes use of these, and for events where there are causes which point in contradictory directions, the cause with more resources tends to win out (almost by definition).
Complex dynamics have sprung out of interactions between these resources, e.g. the sun, the earth, life with plants and animals, humanity, society, capitalism, NVIDIA, Berkshire Hathaway, Saudi Aramco, Elon Musk, Eliezer Yudkowsky, Sam Altman, etc.. The vast majority of the complexity that has sprung up around the resources has ~no influence on their distribution or long-term development, but the distribution of influence is long-tailed, so there are some particular entities who we can observe are likely to have exceptionally big influence.
The heterogenous fluctuations will sometimes lead to massive shifts in how the resources are distributed. Everyone who relied on the earlier distribution and who are not ready to shift to the newer distribution will go extinct, or at least severely diminish in prominence.
Thus, the universe has a "universal backbone" of causal dynamics that really matter, in the pathways of the resources, and the entities that can shift those pathways. One needs to model the parts of this universal backbone that are directly relevant to oneself.
However, because everyone is modelling this universal backbone, everything is tightly connected to it, and that also makes different parts of the backbone tend to be connected to each other. Thus, there will tend to be an incentive towards fully modelling the top of the backbone (e.g. the entities I listed under the "Heterogenous fluctuations" premise), rather than just parts of the branches most closely connected to oneself.
I think the closest selection theorem to this one is The Gooder Regulator Theorem, which states that any regulator which was to solve a sufficiently rich set of tasks with respect to a system needs to model said system. The part of The Gooder Regulator Theorem that seemed a bit fishy to me was the idea that there could be a sufficiently rich set, though I dismissed that concern at first (since e.g. it might just follow from generalization/uncertainty).
However, Linear Diffusion of Sparse Lognormals has made me conclude that the assumption of a sufficiently rich set of tasks is just extremely disastrously wrong and will make you crash into information overload.
The causal backbone conjecture solves this issue via the limited resources: while it cannot be necessary to model everything that is going on, it will at least be necessary to model the biggest aspects of the flow of these resources, since that influences the big dynamics that do matter to you.