"The same network with the same weights can play Atari, caption images, chat, stack blocks with a real robot arm and much more, deciding based on its context whether to output text, joint torques, button presses, or other tokens."
This is rather mind blowing. Does it also mean that the generalist network is smaller than the sum of all specialist networks that are equivalent? Even if not, I find the idea that a single network can be used for such diverse tasks at all highly fascinating.
I don't find it surprising that a single network can do all those things with appropriate formatting of the data. In itself it just means the network has a large enough capacity to learn all the different tasks.
The interesting questions imo, which they studied, is what kind of added generalization takes place by learning across the different tasks. For example, does learning multiple tasks make it better at a given task than a model that is just trained for one task, and can it generalize to new tasks (out of distribution).
They looked at how it performed on held out tasks (see fig 9 in the paper). I'm still getting my head around the result though so couldn't summarize their finding yet.
Yeah, Figure 9 is the money figure in this paper and it actually splashes some cold water on the claims in the rest of the paper. While it generalizes OK to some tasks that are held out, it does pretty poorly on the Atari boxing task, which they openly admit is quite different from the others. Gato seems more likely to be a competent attempt at brute forcing our way towards weak general AI, which is a valid approach, but the question then will always be how does it do with something its never seen before, and how do you possibly brute force every possible situation? I think we're heading more towards a constellation of very intelligent expert machines for particular tasks that may be wrapped into a single package, but that are not strong AI.
Many networks just predict the next integer in a sequence of integers. It sounds like this model identifies what category of problem a sequence of integers falls into and then makes an accurate prediction for that sequence, as you would expect given what it was trained on.
This is rather mind blowing. Does it also mean that the generalist network is smaller than the sum of all specialist networks that are equivalent? Even if not, I find the idea that a single network can be used for such diverse tasks at all highly fascinating.