I want to know how quickly AI's understanding about biological life grows.
If AI simulates a lifeform that is more heavy than an amoeba this question resolves to yes.
Any simulation must be shown to approximate molecule dynamics. If this information is latent then a interpretability technique should be able to extract it.
Any simulation is allowed to have prior information about laws of physics, basic chemical reactions and which molecules should be present but no description of how these molecules should be positioned so definitely not descriptions of entire organelles.
Any AI should learn its simulation from a self-supervised dataset composed of measurement data about at-the-time living amoeba (or heavier). This data may not be too pre-interpreted by a human, so no description of the state of the organelles, but rather pixel data from a video recording or water temperature.
I'll be likely to resolve this to yes if we're largely unsure about the workings of the simulation, but still got novel insights into protein folding from the few parts that we did understand. For this to happen the simulation should still be accurate.
1,000As of today, molecular dynamics (MD), the non-AI based simulation method that, as the question asks, simulates proteins/water/etc at the atomistic level (not quantum, but each atom is it's own particle) cannot simulate even 2 fs of a Craig Venters-style minimal cell. We might get a few tens-hundreds of ps in the coming year, but just as a gimmick for exascale demonstration, definetly not enough to see anything relevant to the function of the cell.
AI based molecular simulation are mostly slower, not faster, that current top level non-AI code when atomistic modeling is necessary. Their benefits tend to be better accuracy of modeling. It is unlikely to change soon because the non-AI code are basically optimal on GPU hardware- where a AI code would also run - in the sense that there is not much waste left to optimize at a given simulation accuracy level. The only area where breakthrough are possible is if the AI code can approximate molecules holistically (not by simulating it's component atoms but as a whole) or have larger time stepping than non-AI code. There are some work on these currently, but they are not at a production ready level even for smaller system, so a cell-scale simulation seems out of reach for the time being.
@CromlynGames Quite confidently, not in the next 3 years. After that, idk.
Generally bearish, because of the difficulty of the problem, but also mostly, because there is not a lot to gain from simulating a whole cell, if you can't simulate it for, say, hours.
Up to now, we have had a handful of simulation of very large system at atomistic level, or almost at that level, for instance a full HIV capsid, a large cube of liquid with protein reproducing a cell's cytoplasm. This was done for very small length of time, due to computational limits (it takes weeks of full utilization on top supercomputer). But we didn't learn a lot from it, because the duration were so much smaller than, say, the timescale for virus attachment or cell division. Unless we can simulate longer time scales, these are going to stay in the realm of showboating for Nature papers. And there is an opportunity cost: with the same compute time, a bunch of other research groups could have simulated their protein of interest in a much smaller system (E.g just a spike and receptor; or just 2-3 proteins involved in the cell cycle) for much longer, where you can get relevant insight, develop drug from, etc.
@CromlynGames I'll be likely to resolve this to yes if we're largely unsure about the workings of the simulation, but still got novel insights into protein folding from the few parts that we did understand.