Hacker News Reader

Comment by Jensson | Hacker News Reader

> AlphaGo guessed it’s way to being the best.

No it didn't, AlphaGo was backed by an algorithm that perfectly understood go and used that to train itself to play go better. That is a completely different situation.

If you tried to make a go engine that only tried to replicate old moves then you'd have a hard time making it perform valid moves. ChatGPT is like such an engine and we just made it perform valid moves most of the time, and now people say "now that it can make valid chess moves it is only a matter of time until it beats masters!". No just statistically reproducing human moves wont lead to a smart chess engine, so why would you think that would make a smart text engine?

View on HN

Because it is not just reproducing existing moves. An LLM can come to acquire a model which models the structure of moves (in chess for instance using letter an number indexing of board positions), the checking of valid moves and the valuation of board positions.

It seems surprising that training an LLM would result in layers/weights which contain such representations but it seems that this is indeed an efficient representation of information in the same way that neural nets would learn addition by encoding textual integers into floats.

You are too close.

AlphaGo was trained with tech. It beat humans.

ChatGPT et al are being trained.

Everyone is arguing the inner workings of what is in the box. It doesn’t matter.

The usage of the box(s) is what will be the big impact.

Previously those boxes couldn’t be relied upon so stayed down in the hierarchy of usefulness.

Now. They are rising.

That would be an interesting experiment. Take something like ChatGPT, train it to play tic-tac-toe and reversi and chess and go. Tic-tac-toe it will probably be unbeatable. How does it do at chess and go? Is there a level of difficulty of the game where its ability to master it tops out?

Can ChatGPT(-ish), trained on a ton of games, beat AlphaGo?

It's been a while, but wasn't it two AlphaGo versions? One that was trained on existing games and "trained" with existing engines. One that was retrained from scratch without any external input? I might be wrong.

Both had a program running it with real Go rules. We don't have a program that can run with the real rules for logic and text to train a language model from scratch.

AlphaGo Zero

It had basic rules of Go as an input.

It also had a program giving it potential moves it could play, and explained to it when it had won the game. That isn't zero input. We don't have such programs for logic or text. So there will be no GPT zero, at least until we have already solved logic.

What is your point out of interest?

Not the person who you posed the question to, but I believe their point may be that Go is a fully described game, a constrained 'space' even if that space is remarkably deep.

Our world is not a fully described game (unless you want to take a radical philosophical position, and in which case we still don't know the rules of this game). We talk about and think about this world in language. Training a machine on human language and then after training, having that machine take language inputs and generate language outputs is not the same as training a machine to 'play' Go because the two worlds they have been trained in (1. the game of Go; 2. the real world) are fundamentally not equivalent.

Reply on news.ycombinator.com