I think this is a step in the right direction, but the performance on most tasks is only mediocre. The conversation and image captioning examples in the paper are pretty bad, and even on some relatively simple control tasks it performs surprisingly poorly.
That's not to say it's not an important step. Showing that you can train one model on all of these disparate tasks at once and not have the system completely collapse is a big deal. And it lays the foundation for future efforts to raise the performance from "not totally embarrassing" to "human level". But there's still a ways to go on that front.
That's not to say it's not an important step. Showing that you can train one model on all of these disparate tasks at once and not have the system completely collapse is a big deal. And it lays the foundation for future efforts to raise the performance from "not totally embarrassing" to "human level". But there's still a ways to go on that front.