[kriscolic]

Originally Posted by Syd Thrift

If you’re talking about training the AI on news articles to produce new ones a la ChatGPT, I think you’re greatly overestimating how large of an operation OOTP is. Also, if I’m being honest, I don’t really want AI generated content in here; the stories, including all of the nested if statements and so on, can be made by humans, thanks.

If you’re talking about training the AI itself… that’s not really how neural networks work and an LLM is going to do basically nothing to make the game make decisions “better”. I think youre maybe confusing LLMs, which basically construct very sophisticated Mad Libs when you enter a prompt, with machine learning, which is a whole different thing (LLMs use machine learning but it’s one very specific application).

Machine learning still needs some human interaction or at least some kind of carrot and stick approach to tell it which decisions work and which ones don’t. You could, conceivably, make it look at standings after a year or 5 years or whatever but I can tell you right now that that will create an AI that’s good at exploiting loopholes in the game, not necessarily making better baseball decisions, and on top of that even if it found some crazy exploit that’s also applicable to actual baseball, it’s based on no accumulated baseball wisdom whatsoever and will look screwy. The thing is, an AI will only really be able to come up with this stuff if you give it thousands and thousands of simulated seasons, and if you have people poring over transactions to sanity check it, much of the bonus you get from machine learning is gone,

Admittedly, this is not my area of expertise. The AI is pretty lacking, though. If a chatbot can compose a play in the style of Shakespeare, it seems like it should be possible to write software that can propose a trade in the style of Brian Cashman. At some point in the not too distant future.