The last third of the talk where he actually discusses work that could be considered to regard generative agents appears to be a deep learning version of the Bayesian model presented three years ago accompanying the initial release of the Omniglot dataset.

http://science.sciencemag.org/content/350/6266/1332.full

Pdf can be found free by googling the paper title

Yeah, from here: https://youtu.be/N5oZIO8pE40?t=31m38s

From where he says, “Now why did we want to do that?” :)

edit:

He starts talking about SPIRAL. (Cute name.) “Synthesizing Programs for Images using Reinforced Adversarial Learning” Link here: https://deepmind.com/documents/183/SPIRAL.pdf

Authors: Yaroslav Ganin, Tejas Kulkarni, Igor Babuschkin, S. M. Ali Eslami, Oriol Vinyals

transcript: https://pastebin.com/Y8dJtJfB

Reflowed the transcript to 80 characters: https://pastebin.com/aQmKc0WQ

Censored in Turkey :-(

This cryovac solo guy really knows his stuff.

Nothing less from the son of the captain of the Millennium Falcon.

Thank you, I wasn't willing to watch for 45 minutes.

As a warning, this transcript was auto-generated. I had better luck turning on closed captions and following along at 2x speed.

Click the triple dot underneath and to the right of (almost?) any video on youtube, click "Open transcript" and read.

It's actually quite awkward to read in that format. They should some kind of easy to read view.

Perhaps interesting for graduate students and professors and industry professionals, but i think is pretty far beyond comprehension for those with only basic understanding of machine learning.