OpenAI, Google and Anthropic are struggling to build more advanced AI

i think you underestimate the amount of data a driver experiences in a single 5 minute drive

I never get this argument.

I've seen a deer on a road maybe once. I've seen a rabbit on a road zero times. But I know what to do if I see one.

Is that because the "video" of my perception has many "frames"? Even if that's true at some level, I think it's massively missing the point. Yeah, so I saw that one deer from a lot of angles. But current AI training is like the equivalent of taking every deer that has ever been on camera in the history of the human species.

Somehow I'm still dramatically better at generalization than the AI. Surely that's an algorithm difference.

You might personally have seen a deer just once, but human evolution, and animal evolution prior to that have practiced this skill a lot. AI doesn't have the advantage of evolutionary priors baked in, so it needs explicit walking through many combinations to infer its structure from data, and is remarkably efficient. GPT-4 'only' trained on the amount of language that 30,000 humans use in their lifetime.

But we have seen from AlphaGo that when training data is extensive, it can rediscover strategy on its own and even surpass us. It's not inherently worse than human learning.

Human DNA is just 750 and only a fraction of it is something that may be called "brain pre-training".

> You might personally have seen a deer just once, but human evolution, and animal evolution prior to that have practiced this skill a lot.

Which pre-human animals evolved instincts for swerving a car to avoid a deer?

I'm pretty sure that evolution would select out anything that could not generalize pretty quickly.

A charitable interpretation of what you're saying is that humans produce lots of original data from their experiences of the world, like thinking about their experiences, imagining what they would have done differently, and perhaps even dreaming. I agree with the root comment that something is fundamentally missing, and probably it is the ability to iteratively learn from one's own experiences, test understanding, and recursively improve.

There are definitely teams working on applying reinforcement learning to LLMs. Maybe that will unlock new potential from finite training data.