Hacker News Reader

> which is inherently based on existing sources

Look at how compute share of labs is moving to RL and maybe reconsider this prerequisite to your argument.

Open to learning. How could RL yield novel discoveries? What I mean to describe are things reliant on empirical or real-world testing of human reception. How would that be synthetically derived?

We're seeing the cost of code drop dramatically and are facing the same problem in building new companies across most disciplines of, "what's a problem painful enough that someone will pay us to solve it?".

I suppose if we get to a massive enough scale we can effectively model every atom's interaction and human on earth's reaction to any stimulus but I think we're still some time off from that :)

For a cost-prohibitive example, you could put a humanoid robot whose pretraining contained no chemistry, in a chem lab, and put it in a hypothesis loop, where it generates a random action, guesses the result, and then does it. Every time the lab burns down, you just build it again. Reward the model for producing novel actions and reward the model for correct hypotheses.

In other words, you generate more training data by actually experimenting, just like humans do.

Reply on news.ycombinator.com