The protein folding solutions like ESM/Alphafold were not due to this LLM/agentic coding or autoresearch type approaches though. They were designed by bio ML researchers.

It's hard to keep track of the frontier on bio ML, but it seems that we're going slower than what Demis Hassabis said in 2024 with 5 years to full cell molecular simulation. We still aren't able to reliably model a tiny surface of the cell membrane.

And of course there's Derek Lowe's takes on the drug discovery pipeline waiting for the proof in the pudding.

To me the only reasonable bullish position is that there is a very non-linear AGI threshold for accelerating progress that we haven't hit yet.

For me personally, I'm looking at other more tractable fields as a proxy to measure this kind of progress. The best modest evidence is from the agentic coding area, (modest because these kinds of gains may not translate to bio progress). Other soft-ish fields to like legal/law/tax are also interesting to watch, as a small amount of people are now trusting AI for these areas that were considered totally unusable a year ago. Another proxy is being able to generate generally entertaining media.

The field of AI did very well in protein folding, but it was completely different from LLMs.

AI made a large improvement in accuracy on a specific protein dataset. We still don't know how exactly protein folding works. So no, it's not solved.

It did a lot more than that; no, it didn’t “solve” protein folding but for people whose interests are not “work out how protein folding works” it basically removes “solve protein folding” as one of the potential bottlenecks for what they actually want to study.

Turns out that “give a reasonable probability of being close enough such that you can bootstrap a solution out of experimental data” gives a very high utility and effectively obsoleted several experimental techniques overnight; pretty much “Molecular replacement” is about the only technique for phasing resolution anyone bothers with any more.

But again, “Bio” is an _extremely_ broad term; for every part of the field Alphafold had a big effect on there are a thousand different parts of the field that it did nothing for.

> I mean they did solve the protein folding problem did they?

They made huge progress, but I would say that the vast majority of work on this problem was designing the harness for the model. That's a lot of work for each and every domain.

There is already (partial) automation there, for example for finding new drug candidates.

Isn't that so far only static folding?

As usual with LLM era craze, finding candidates was never the bottleneck. And, so far, it failed to actually provide any tangible results. [0]

[0] https://www.science.org/content/blog-post/so-how-ai-drug-dis...