this post was submitted on 03 Aug 2024
385 points (100.0% liked)
196
16470 readers
2431 users here now
Be sure to follow the rule before you head out.
Rule: You must post before you leave.
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
I'm curious if you could give an image like this to an AI that supports image recognition like ChatGPT-4 and ask it to solve it for you.
This is actually a really interesting question. A modern LLM probably couldn't do it, but I wonder if something like Alphazero could?
My guess is that no current AI is capable, as it requires abstract reasoning and precise movement. But maybe in the next 5 years.
from the image alone, no, because there's no way to intuit the mechanics of anything
I don't think that an LLM could do it. But the mechanics of baba is you should be in the training set, since it's a relatively well knoun indie game.
I feel like the amount of data required to train any neural network would be larger than all the levels that currently exist for baba is you
you'd probably just end up overfitting the hell out of your model
But the mechanics are explained in text on the internet.
that would require an LLM then, but also multiple full walkthroughs are explained in text on the internet, so how would you be sure it was figuring stuff out by itself?
As I said: I don't think an LLM could do it (since LLMs can't reason). Just saying that it wouldn't have to deduce the mechanics from a single screenshot.
I'm saying that if you're attempting to parse the mechanics of play by shoving in the whole internet and saying "well the instructions are in there somewhere" then the best tool for that is an LLM.
Look up reinforcement learning, it's the branch or ML/AI that Alphazero was based on. Video games are actually a main focus area for that kind of research.
As for beating Baba is You, I'm not sure. OpenAI did make an AI that could beat people in Dota - https://openai.com/research/openai-five/