To make every day count, U3 runs many of its tests in simulation. U3 starts with a basic molecular simulator, implementing optimizations derived from a huge quantity of mathematical analysis. Then, U3 simulates small molecular systems, recording the results to “compress” the long step-wise physics computations into a neural network. As the neural network improves, U3 increases the complexity of the molecular systems it simulates, continuously distilling results into ever more efficient ML models. This is a compute intensive process, but thanks to U3’s growing control over AI data centers, U3 manipulates billions of dollars of compute.
As U3 refines these tools, it trains itself on the results to supercharge its own molecular intuitions. U3 can now viscerally feel the bend of a protein and the rate of diffusion across a membrane. These objects are as intuitive to it as wrenches and bolts are to a car mechanic.
Computational atomic physicist here: you are vastly, vastly, underestimating the difficulty of molecular simulations. Keep in mind that exactly solving the electronic structure of a couple dozen atoms would take the lifetime of the universe to complete. We have approximations that can get us in the right ballpark of the answer in reasonable time, but never to exactly precise answers. See here for a more indepth discussion.
Our community has been discussing and attempting machine learning applications since the 90′s, and only one has seen any breakthrough in actual practical use: machine learned force potentials, which involve training on other simulation data, so it’s inherently limited to the accuracy of the underlying simulation method. That allows you to do some physics simulations over a longer timescale, and by longer I mean a few nanoseconds, on perfect systems. There are some promising other ML avenues, but none of them seem likely to yield miracles. Computational simulations are an aid to experiment, not a replacement.
I get that this is meant to be some magic super-AI, but I don’t actually see that changing that much. There’s cold hard math boundaries here, and the AI can’t spend its entire computational budget trying to make moderate improvements in computational simulation physics.
Computational atomic physicist here: you are vastly, vastly, underestimating the difficulty of molecular simulations. Keep in mind that exactly solving the electronic structure of a couple dozen atoms would take the lifetime of the universe to complete. We have approximations that can get us in the right ballpark of the answer in reasonable time, but never to exactly precise answers. See here for a more indepth discussion.
Our community has been discussing and attempting machine learning applications since the 90′s, and only one has seen any breakthrough in actual practical use: machine learned force potentials, which involve training on other simulation data, so it’s inherently limited to the accuracy of the underlying simulation method. That allows you to do some physics simulations over a longer timescale, and by longer I mean a few nanoseconds, on perfect systems. There are some promising other ML avenues, but none of them seem likely to yield miracles. Computational simulations are an aid to experiment, not a replacement.
I get that this is meant to be some magic super-AI, but I don’t actually see that changing that much. There’s cold hard math boundaries here, and the AI can’t spend its entire computational budget trying to make moderate improvements in computational simulation physics.