← Back to context

Comment by naasking

1 year ago

Some good prompt-reply interactions are probably fed back in to subsequent training runs, so they're still stateful/have memory in a way, there's just a long delay.