Reflecting on training AI

Posted:

I've had just enough exposure to machine learning concepts that I have some idea of what AI programmers are up to, but there's one thought that keeps popping in my head when I read about training Large Language Models. What is considered a "reward" for an algorithm?

I constantly see (and sometimes read) articles with titles like "Training Language Models to Generate Text with Citations via Fine-grained Rewards", and I try to imagine how to say "good boy" to a computer function. I remember when Google was publicly training an algorithm that could recognize drawings by creating a game where people can draw a sketch and have the computer guess what the drawing is. In the game, the app prompts you on what to draw, and as you do, the AI makes guesses. When the AI gets it, the app tells the AI that it is correct and stops the clock.

The genius in this method is that it doesn't rely on the user to "reward" the AI by letting it know it got the correct answer. In essence, it's using one piece of "dumb" software to manage the training session.

It becomes pretty clear that a reward is just a true value in an "if" statement that can break the AI out of a while loop of guessing. I imagine that behind the scenes, the code is tracking its wins as well as its failures, and then tries to use the previous wins as starting off points for future guesses. There's some mathematics going on behind the scenes to guide the guesses, but even so, there's got to be a lot of raw computing power going on behind the scenes.

But, now that AI is out in the wild as a tool, and not a game, I'm not sure that users are giving enough feedback to the algorithms so that they can continue to learn and improve. I know if I ask an AI a question and it gives me a nonsensical or inaccurate answer, I MIGHT try to correct it and ask again. But, I'm just as likely to snort and consult some other resource. And, if we're not actively telling the AI, "bad, robot!", is it assuming it got things right and therefore learning the wrong lessons?

I guess I'm a little sad that there is no real "reward" for the AI. Nothing at all that will please it, make it happy, or give it a tangible benefit of any sort. Computers and software make ME happy, and I just wish I could return the favor.