My cellular biochemistry does a pretty good job of replicating itself with absolutely no understanding whatsoever of how to resolve the contradictions between general relativity and quantum mechanics.

When it comes to brains, I don't know if anyone knows what might be the simplest sufficient model that would usefully replicate them, even if you specify "usefully" well enough to know if this is about fundamentals of intelligence or about the impact of drugs on cognition, which are two completely different standards.

For example, perceptrons are a toy model, but modern AI can do more in (breadth XOR single-skill performance in various domains) than any single human, even with much smaller parameter counts than we have synapses; but the broad-skilled ones also mess up in inhuman ways, like being equally good at advanced calculus as basic arithmetic, or being a poet at the level of stereotypical teenager but in every language simultaneously.

If anyone's made a neutral network that can get high on simulated caffeine — and I'm not saying it hasn't been done — it's not reached any of the places I follow discussions on this kind of thing. (Google didn't help, results were about software named Caffeine and non-artificial neurones).

If the "something" never exploits a particular law, then in principle you could simulate it if you knew all the laws that actually applied (imagine a hypothetical biological organism that never exploited or experienced quantum entanglement).

But strictly speaking, as we understand it, it's not possible to replicate something exactly without recapitulating the exact laws and running a deterministic simulation, which is not practical.

I don't think anybody is really attempting to exactly replicate things, but rather to create a physical model which can be calcualted and contains enough similarity or transferrability to be able to make accurate generalized predictions about the behavior of the simulated system. How and why that works with modern math methods is still somewhat mysterious. The most useful thing written about that so far is https://en.wikipedia.org/wiki/The_Unreasonable_Effectiveness...

Abstraction. You don't need to know what's going on at the quantum level because higher-level simulations can capture all the properties you care about.

This is a fundamentally flawed way of questionning, "knowing all the laws of physics" is probably impossible to achieve, but we still produce accurate predictions for a lot of phenomena. You can see that we do in fact "replicate" (predict would be a more proper term) things despite not knowing all the laws of physics: weather, movement of the stars, cooking time for a browned piece of bread...

The same way we can simulate the movements of the planets - it will never be exact, but the better we understand it, the more precisely we can simulate it

clearly depends on needed resolution and our ability to recreate the material structure. you don't have to understand much about wood to build a chair.