I personally know several people that do research that have unlocked new possibilities through this tool. My wife is a neuroscientist and she's used this tool a few times for reasons that are above my head (even with a Msc in Microbiology). This type of work used to take a PhD student 4 years or more to do a single relatively simple protein. Getting answers within a few seconds is revolutionary.

Many engineers will say it doesn't code. It just regurgitates and remixes the data it was trained on. It just makes "meta-heuristic guesses."

But anyone taking an honest and objective view of it can see that Copilot does add value. It's no substitute for a real, human, engineer, but it clearly adds value.

I don't think AlphaFold would get to this level of funding, resource commitment, etc. if it was adding 0 value.

Being a domain expert, I'm curious what value, if any, you think a large transformer model could add to the domain of protein folding. Is it really zero value, in your view?

> Author Pamela McCorduck writes: "It's part of the history of the field of artificial intelligence that every time somebody figured out how to make a computer do something—play good checkers, solve simple but relatively informal problems—there was a chorus of critics to say, 'that's not thinking'." Researcher Rodney Brooks complains: "Every time we figure out a piece of it, it stops being magical; we say, 'Oh, that's just a computation.'"

https://en.wikipedia.org/wiki/AI_effect

You won't have a perfect answer unless you want to predict its shape in a vacuum, which wouldn't be very useful either way. Having it "close enough" is already extremely useful. There are definitely edge cases where it gets it wrong, but there are always edge cases in ML. More data = better results with the same architecture.

Tons of things that won Nobel prizes weren't 100% accurate, it's not a prize for solving science, rather a prize for advancing science.

My wife is a neurobiologist and the impact of this advance is groundbreaking for her work.

What is important about their discovery is that we now know for certain that a judicious combination of expensive-to-obtain structure information, and easy-to-obtain protein sequence relationships can be used to build a generalized protein structure predictor (it can predict structures with no prior example of a fold, although there are limits)... and you don't have solve the general folding problem to do it. You do not need to know the path, to get to the destination!

Many of us in the field expected this to be true but there wasn't any really good example to point to that was widely accepted by the community. And in the ~year or so since this was demonstrated, the community has already found a wide range of uses for this that have validated the structure predictions and demonstrated their utility- using open codes and models.