RL and Decision Transformers - A Sequence Model

In some sense, you can think of this as the software 2.0 moment for RL, right? Like you let the neural network write the weights for the RL algorithm itself and that's all it is. If you can do this reliably at a large scale, then we can like basically leverage every scale, large scale infrastructure that's been built for GPTs and images or Dolly. And that in my opinion, at least is the future. It's much easier to scale something when there is a whole community of people and investment and resources being put into scaling a particular infrastructure.