How do you decide the best chess move? Well, if you had an infinite amount of time, the optimal strategy would be to map out all the possible chess games and then work backwards recursively from the end, assuming that your opponent will make the best possible move and then making the move that limits the best possible move they can make. If you have a limited amount of time, then you can only think so many moves ahead, and the you would have to rank possible game states based on how likely you are to win them. Computers are better at this than humans. As Garry Kasparov put it, "The human mind isn’t a computer; it cannot progress in an orderly fashion down a list of candidate moves and rank them by a score down to the hundredth of a pawn the way a chess machine does."
With Go, there are many more possibilities, but computers have still managed to beat the top human players. One thing that helps is that computers have a desired outcome and a way to determine whether a particular result is desirable or not even if they've never seen it before. Machine learning has gotten particularly good at most strategy games, even those with many possible different outcomes. Observing human players can give an AI model a basis for its strategies and for determining which game states are better than others. Then after it plays billions of games, it can empirically test which strategies lead to better outcomes.
But what if you wanted to write a novel or draw a picture? Using conventional wisdom, you would have the AI make one move at a time. In this case, that means deciding the next word to write. But even with large datasets, AI still lags behind human writers and artists. In contrast to strategy games, there's no objective metric for determining how good a book is, besides polling a lot of people. And in contrast to strategy games, writing books one word at a time probably isn't even the best way to write a book. In chess, you have to be ready to change your strategy every time the opponent does something. With writing, everything is up to the author. Writing a book by predicting all the possible endings is near impossible, and not even desirable if you can't measure the quality of all of them. However, writing is much easier when you know what the ending is going to be.
What if humans had to write books one word at a time, without going back and editing? Suppose 100 authors were in a room, interacting only with online polling software, and they had to vote on the next word until they had a story. I suspect they would do much worse than AI. Yet we expect machine learning software to write using this method, just by feeding it enormous amounts of data.
What if we changed the way we trained AI models to write? What we do right now is a bit like giving an AI model a complex word problem and a complex answer—maybe with enough data, the AI will eventually give decent answers, but it's much harder if the machine doesn't see each step that was taken to reach the answer. Suppose you had to learn to solve math problems in Chinese, and all you saw were paragraphs of questions and answers with no work shown? Maybe a better way would be to have each step written out, so you could figure out what each symbol represents, what the symbols are for addition, etc. It might seem like this is impossible with writing, since many of the steps in the process happen subconsciously, and authors don't say every thought that comes to their mind.
Should we have writers log every thought they have as they write? Do we need technology that scans human brains? I think it's possible to give AI a fighting chance without going that far. Google, one of the largest tech companies, logs each edit users make on Google Docs. Through Gmail, it can see how people create emails. Is this done one word at a time? Rarely. People often go back and rewrite sections once they know how the document ends, changing what comes earlier because of what comes later. As Bill Wilder put it, "If you have a problem with the third act, the real problem is in the first act." Rarely is a first draft publishable quality. If computers can be trained to behave like Go players or video game players or learn to solve captchas even better than humans, why couldn't they be trained to act like human writers, fleshing out an early draft before going back and polishing earlier bits?
I suspect we would see much better results. If Google and other large companies know what they're doing, they've probably even considered taking steps along this path. Instead of training LLMs to write a perfect first draft, something we don't even expect from humans, why not teach them to improve on the first draft? I strongly suspect this approach would work. The bigger question is whether large companies should take this approach. Is it worth all the personal data that would need to be forked over to create such a model? If you're against AI on principle, you'd probably hate this. But in the end, this isn't something we can really change. Up to this point, I've written this as an injunction, a suggestion, but it's more of a prediction. Once tech companies learn that AI can be trained this way to yield better results, those models will be built. And after that, there's no going back.
