I have written a paper on this recently viewing transformers as Hopf algebras https://arxiv.org/abs/2302.01834. This paper is very relevant https://arxiv.org/abs/1206.3620.
Pretty cool stuff and helped me grok some of the more difficult back and forths from a tutoring session w/ GPT-4.
Also the particle appendix is really cool and I wish we had the whole elemental table. Lots of self-consistency in what the algo wrote and I'm pretty impressed, though of course I'm sure it's maybe best to treat it as a leaky bucket of information while learning it.
Anywho here's the paper (nearly-direct link): https://github.com/tysam-code/fileshare/blob/main/knick_knac...
Obviously an early line of research won't have anywhere near the depth, but as an example the geometric algebra community has done a great job of this. Lots of 10 minute videos to 2 hour lectures. Steve Brunton's videos are another example "learning ML math" resource that has helped me a great deal.
Edit: I also just found https://bivector.net/ Holy crap. So good.
That said, Adam tends to follow a very precise and measurable interaction selection path with many people regarding some of his work and the discord, so I also just see it as how he interacts with others.
But yes, agreed, I wish I had a human-understandable resource that a mortal of my own kind of flesh could more easily grok. Even the 'for absolute beginner' online tutorials are a swamp of symbology and prior assumptions, and it doesn't even seem to be that terribly difficult of a concept on the whole, at least for the basics of it.
Just my 2c, lots of good and mixed for us all in this crazy road we call life! <3 :D :)))) :D
It sounds so promising. But I have to assume that if it really is then this can be translated to code that can run on Colab or something, at least by some genius. Until it is I have to be a little skeptical and I'm not going to be able to access it anyway.