On: Pathways Language Model (PaLM): Scaling to 540B pa...

[gibsonf1]

It doesn't grasp the world at all, there is no intelligence. It's just statistics about word use.

[1024core → gibsonf1]

Please explain the "joke explanations" to me, using statistical methods.

[yldedly → 1024core]

Take 100k joke and explanation tuples as training. Pattern match input joke to training jokes. Edit amalgamation of most structuraly similar training jokes to match content of input joke. Edit corresponding training explanation with same substitutions. Output explanation. Repeat with 100 input jokes, pick the best one.

[1024core → yldedly]

But that is not what they did. Please read the paper.

[yldedly → 1024core]

I'm not saying that's what they did. I'm saying it's functionally equivalent.