Hacker News Reader

Comment by comment_ran | Hacker News Reader

Why not just introduce AlphaGo as an API that can be used by chatGPT? So every time you want to do a this type of gaming, you just send a request. I mean, chatGPT sends a request to AlphaGo, but as a user you don't know actually what's happening. But in the background, it happens really fast, so it's just like you are chatting with chatGPT, but using much, much powerful tool to do this kind of things.

View on HN

That kinda defeats the purpose. Of course you can use AlphaGo, but the question here is – can a generative AI teach itself to play chess (and do a million other similar generic tasks) when given no specific training for it.

What if you come up with a prompt to tell ChatGPT to figure out how to use APIs and external tools to help itself get better at chess?

GPT-4 plys chess just fine. So you certainly cn

Honest questions,

Isn't it very likely chess games were part of the training data? If so, chess sequences seem like an amazing matched filter of sorts. It doesn't seem surprising that it can guess a next move.

Is it teaching itself anything here? Is the model being updated in any way? Or are you talking more generally, looking into the future sort of statement?

You can put every chess game ever played in its database and yet every time it plays a new game it will immediately come across a completely unique move. So it still needs to learn the rules and know how to play independently. Whether it can do that or not is yet undermined.

I think they meant “teach itself” in the course of a conversation, within the token limit. ChatGPT will forget what it has “learned” once the conversation is over.

How about, can a generative AI teach itself how to use a chess AI to beat chess? Give GPT4 the ability to make REST API calls and also access to FFI, and put a chess-bot library somewhere. Train it how to use these but not necessarily how to use the chess API specifically. If you ask GPT4 to play chess, can it call into that library and use the requests/responses?

This has bigger ramifications too: if GPT4 learns how to use RPCs effectively, it can do accurate complex mathematical computations and simulations, do IoT/home automation tasks, and even control heavy machinery.

GPT is a language model. It doesn't call APIs. You can take the output of GPT and decide to call a specific API based on it, and sure you can bake that into ChatGPT or some independent tooling, but again that's not the purpose of the core project. I'm sure plenty of people/companies are working on "chess engine interface using ChatGPT" already.

This is exactly what I'm hyped for in the next-gen GPT-7. Imagine it having the ability to self-teach, just like a child. I may not know how to whip up some cheesy goodness, but with external resources like YouTube vids, I can improve. And if GPT-7 can store this knowledge, it can access it for future tasks! That's some next-level stuff, and I'm stoked to see where it goes.

This is actually a huge debate right now.

OpenAI is on the side of 'LLMs have only surprised us to the upside, so using crutches is counterproductive'

Whereas other people think 'Teaching an LLM to do arbitrary math problems through brute force is probably one of the most wasteful things imaginable when calculators exist.'

I'm actually very excited to see which side wins (I'm on team calculator, but want to be on team LLM)

I think I'm on both teams.

As GPT models get better at arbitrary computation, it stands to reason that their ability to use tools will rise as well.

Smarter models and smarter ways to get them using tools feel like complementary capabilities.

How about a more human-like approach: the LLM designs a calculator and then makes use of that!

And Mathematica when given a math problem. And web search search for news.

And a good hacker can add these features to Alpaca and run it on his desktop...

Import all his WhatsApp chats and private emails into the corpus.

Add speech recognition and TTS.

Reply on news.ycombinator.com