Conversation

If we were to train a sufficiently large and powerful model on all available data for a sufficient time, do you think it will converge on

one possible universe
31.8%
a small set of universes
8.6%
countably many universes
15.8%
uncountable universes
43.9%

1,391 votesFinal results

Replying to

countable but it depends on whether it's embodied - I'd expect an embodied model to converge on a smaller set

Replying to

and

my intuition here is that as long as your data is countable (which it would be if collected by any practical method), you can fit anything to it, but every interaction with the universe (real or simulated) constrains the models to useful ones

Replying to

interaction is just data that you interpret as interaction

Replying to

interaction is codata

Replying to

and

the salient difference here is that formalism-wise, data is what you've already collected, while codata represents what you're going to collect - a natural choice for I/O or randomness sources

5:57 AM · May 31, 2022Twitter Web App