Re:LLMs trained on “A is B” fail to learn “B is A”
by peteradio on Saturday September 23, 2023 @10:00AM
(#37623294)
Did anyone try what you proposed with the LLM? -- Give them lots of general context including other games but withhold any actual chess games. Given them the manual now see how it responds to a prompt to play chess. --
Re:LLMs trained on “A is B” fail to learn “B is A”
by Jensson on Saturday September 23, 2023 @10:17AM
(#37623432)
If you give the chess manual during runtime it could work, plenty of examples of that working. If you give the chess manual during training time it wont work, which means that the LLM can't understand what it is trained on. But you could argue that the LLM can understand the data you give it during runtime, it is harder to argue against that, but it is pretty easy to say that it doesn't understand itself ie the data encoded in it.
Maybe you need two models to understand things, since the models wont understand itself but a model can understand the data encoded in the next model. This isn't an unfixable problem, but it is a problem that has to be solved.
They can recite the rules perfectly, hence they remember them, I didn't say he just saw the rules once, both the LLM and the human has seen the rules enough to recite them perfectly.
Then a typical human can start to play chess based on those rules without having seen anyone play before, humans learn to play games by reading manuals all the time. You trying to argue against it here is absurd.
> think about how many times you've sat down to learn a new board game and after the person who had played it before explains all the rules you still don't really get it
I read the manual before playing board games because I like reading manuals, it is me teaching my friends then. You don't need someone to show you how to play games, reading the manual is enough...