That is actually a great idea. One of the issues with a network, however, is how the output would be produced in such a scenario and how the dictionary would be controlled and how the training data could be managed. The transformer model allows different RNNs to be separated, but it would be challenging to 'put it all together'. Furthermore, bad actors could intentionally poison the network during the training process.
Furthermore, how would upgrades work? When new techniques are discovered, it would be hard to upgrade it. We see these challenges already when it comes to major changes in cryptocurrency networks.
What you’re describing is http://chat.petals.ml/
Excellent. I will check that out and see if I can contribute.
Thanks a lot for the link. It is an interesting project. I tried the Google Notebook and it was a trip:
Somehow, the AI believes that Alan Shepard was the first person on space.
Doing this session reminded me A LOT of Asimov's short stories in "I Robot". I feel that the future AI "debugging" session will be like that. It's been a while since I read the book, so I'll have to re-read it again.
I would love to work on doing something like that.