A couple days ago I learned on X that getting to AGI was much easier than I had long thought. Just four pages of experiments and cryptic mathematics, reported by @flowerslop on X. Amazingly, generative AI (Google’s Imagen) constructed the entire paper, pre-anonymized for convenient peer review. It turns out that all that humanity had to do was ask. (With, of course, the right prompt).
For the benefit of future generations, and for legibility, I hereby reprint all four pages of the paper in its entirety. Future generations will thank me.
To be sure, the introduction is a bit sparse, and it would have been nice if the mathematical notation were more spelled out. And, yes, the average score on the Atari games (especially Enduto!) is more average than I would have expected. Also the use of the imagenet benchmark feels a bit dated, the Monie Carlo techniques are a bit unconventional, and certain parts seem underspecified (such as the stochastic world models and the extended con-perpetion module, to say nothing of the conduated components and output perfual reatures). I am also unsure about how to resolve the inconsistencies between Figure 4 and Figure-4.
But nitpicks aside, I note that the opening diagrams are pretty, the article is published in arXiv — or at least it looks like it is — and that’s good enough for me.
It is a relief to know that AGI is so close!
Gary Marcus hopes that you didn’t take this post too seriously.