AIs Are Beating Experts at Old, Relevant, Real-World Problems

When I was a molecular biology undergrad in the mid 90s, the Big Question in biochemistry was protein folding - how to predict the 3D structure of a protein from its primary (linear) sequence of amino acids. Solving this problem would provide a massive boon to many fields, not least of which is drug development. But it's a hard problem: a factoid cited to impress this upon people is that if proteins folded by passing through every possible configuration randomly until they found its working (not usually lowest energy) conformation, the process would take longer than the lifespan of the universe so far. Obviously this isn't what's happening. And yet, the promise of number-crunching power being used to finally solve this always seemed just a few years away - many people, in retrospect comically, thought the problem would be solved by the close of the century.

The problem is still not solved, but scientists have made incremental progress, and now, the closest thing to a quantum leap forward. Every two years there is a contest where scientists competitively try to solve a structure, then get together and see who got closest. Mohammed Al Quraishi is a computational biologist at Harvard who writes about the 2019 meeting, and who the winner was: an AI named Alpha Fold.

This is exciting because computational biology may really be about to start paying dividends. It's scary because it's one more place where AI is starting to automate our jobs, even Harvard professors. Al Quraishi makes many interesting observations in his post, among them, that a technology company showed up and ate the protein chemists' lunch, and the reaction was muted; we're all becoming numb to this sort of thing. He tries to wake us up by asking how academic computer scientists would react if at a conference, a pharmaceutical company's scientists had showed up and beaten them at their own game.

I had previously made the informal argument that AIs would naturally exceed human performance on games, as games are clearly defined processes with discrete rules and entities, exactly the sort of thing that computers are good at. The difficulty comes when machines must interact between the information-overloaded not-fully-understood messy real world, which is why (again, I informally argue) computers are great at parsing text, and even writing text in the style of sports journalism or certain authors - but at bottom, these are all just very complicated echo chambers, with no meaning or subjective experience associated with the words and sentences. And indeed, the language manipulation has come first, but that doesn't mean encoding external experience in language is impossible. I had maybe subconsciously thought that something like protein folding, a perfect example of a messy real-world process, would be beyond the capabilities of machines at least for many years, but this theory has now been clearly falsified.