AI-powered
podcast player
Listen to all your favourite podcasts with AI-powered features
How to Use Pre-Trained Language Models to Structure Exploration
I would bet good money that there are going to be like 10 papers in the next year saying we took GPT-3 and we used it to like place a distribution over natural language descriptions of high level action sequences. And then we like manually grounded out in low level plans and now we can generalize. That's really if that turns out to be true is going to be a pretty remarkable thing that we can get information out of language models and use it to guide real world behavior. The more interesting question is how do you actually learn those groundings?