Back to Blogs Home

Why RL Will Take more "Market Share" From SL In ASI Long-Term

Guy Reading
Nov 06, 2024

Abstract

My views on why reinforcement learning (RL) will be a core solution to artificial super-intelligence (ASI), not supervised learning (SL).

Introduction

When we talk about ASI, we imagine a computer that is able to out-perform the thinking of a human. And with the advent of LLMs & generative AI, we might start thinking that supervised learning is how we'll get there. There are a few issues I have (but I'm not smart enough to do any better) with the current state of the art of LLMs, including that they're currently just a statistical mapping of an input to an output (which means that they don't understand what or why they're doing what they're doing), but in this post I'd like to focus on just one issue.

"The Issue"

This issue is: supervised learning models learn from data. Human data. So, the SL models will only be as good as the humans that they learn from. Even worse: they'll probably end up being as good as the AVERAGE human, as they're training, and looking to generalise, on the whole dataset of human data.

However, when we think of ASI, we think of agents that can out-perform humans. Currently, as far as I can see, there is only one ML technique that can do this: RL. RL has trained agents that have beaten the best chess, Go, StarCraft and DOTA players, to name a few. In some games, the agents have won by a large margin. RL doesn't rely on how the best humans play, it works on how to play the best game overall, based on only the game rules. This is a stark contrast to SL.

Final Words

RL still has other issues, including the one mentioned in the introduction: specifically, model-free agents learn using just a statistical mapping of observations to actions. It doesn't understand why or how it is doing what it is doing.

It's also interesting to note how well SL does, even though it's using human data, of which a lot of it is sub-optimal or has blatant errors. It goes to show the magic of AI at scale, and how humans do a good job, on the largest scale, of not having systemic errors or mistakes: therefore any particular grammatical error one human makes will be smoothed out by the larger dataset.

Overall, it'd be interesting to see which technique gets used for which tasks in the future. It seems like some tasks actually work best by regressing towards the mean: including language tasks and more general gen-AI tasks. Perhaps the intuition here is that if someone spoke with the words of a 200 IQ person (without consideration for their audience), not many people would understand them. Other tasks look to require revolutionary changes, such as AI discovering scientific breakthroughs. It would make sense that RL would eventually be what cracks these aspects. In the future, I'd expect to see RL taking on more of these core sub-tasks and working with SL to solve harder and harder problems.