I just can’t wait until one of the wrong Q* hypotheses turn out to be even better than Q*
Yann Lecunn tweet what this is today. Token prediction with planning. Far below prompt level.
Please explain what he mean with his post:
One of the main challenges to improve LLM reliability is to replace Auto-Regressive token prediction with planning.I had to read that a few times.
Auto-Regressive is like forecasting, it’s iterative.
LLM reliability is this vague concept of trying to get to the right answer.
Hence tree of thoughts as a way to ‘plan’ to that vague concept of the right answer.
Circumvents the univariate next token prediction limitation with parallel planning.
ermm, idk what you mean by any of those words.
Auto-regressive just means it’s a time series that depends on its previous predictions.
So, when you predict a token at time t – you condition on the previous tokens you already predicted.
Consider, “the cat in the hat”. A transformer that predicted it would have predicated it in the following manner (assuming that each of the words are a token bc I’m lazy):
-P(“the”|prompt) is highest
-P(“cat”|“the”,prompt) is highest
-P(“in”|“the”,“cat”,prompt) is highest
So you can see there is a dependency between each of its predictions and the next prediction. This is what is meant by auto-regressive.
Yes I understand all that
Auto regressive is like arima In time series forecasting
Then rnn came along
Then sequence to sequence
They all have the last prediction is used as input for the next prediction in common
Hence auto regressive
It sounds like open source synthia, openchat, and zephyr lol. The whitepapers. lolol.