Also AI is really good at computing protein shapes. Not in a “ChatGPT is good enough that it’s not worth hiring actual writers to do it better” way, in a “this is both faster and more accurate than any other protein folding algorithm we had” way.

Yeah, people don’t realize how huge this kind of thing is. We’ve been trying for YEARS to figure out how to correctly model protein structures of novel proteins.

Now, people have trained a network that can do it and, using the same methods to generate images (diffusion models), they can also describe an arbitrary set of protein properties/shapes and the AI will generate a string of amino acids which are most likely to create it.

The LLMs and diffusion models that generate images are neat little tech toys that demonstrate a concept. The real breakthroughs are not as flashy and immediately obvious.

For example, we’re starting to see AI robotics, which have been trained to operate a specific robot body in dynamic situations. Manually programming robotics is HARD and takes a lot of engineers and math. Training a neural network to operate a robot is, comparatively, a simple task which can be done without the need for experts (once there are Pretrained foundational models).