[D-Machine]

This article is garbage and makes many incorrect claims, and it is clearly AI-generated. E.g. the claim that "AlphaFold doesn't simulate physics. It recognizes patterns learned from 170,000+ known protein structures" couldn't be farther from the truth. Physical models are baked right into AlphaFold models and development at multiple steps, it is a highly unique architecture and approach.

AlphaFold models also used TPUs: https://github.com/google-deepmind/alphafold/issues/31#issue...

EDIT: Also annoying is the usual bullshit about "attention" being some kind of magic. It isn't even clear AlphaFold uses the same kind of attention as typical LLM transformers, because it uses custom "Evoformer" layers instead: https://www.nature.com/articles/s41586-021-03819-2_reference...

[lukah]

I interpreted that section as alphafold not learning physics, but rather correlations within a constrained setting that a-priori correspond to physically sound inferences. It has a specific architecture that allows the model to make inferences that are more physically plausible than not, but not that it’s discovering actual, causally verifiable laws of nature (like what I’d assume are encoded into another non-ML approach to the folding problem for example).