> 21. That every expression graph built from differentiable elementary functions and producing a scalar output has a gradient that can itself be written as an expression graph, and furthermore that the latter expression graph is always the same size as the first one and is easy to find, and thus that it’s possible to fit very large expression graphs to data.

> 22. That, eerily, biological life and biological intelligence does not appear to make use of that property of expression graphs.

Claim 22 is interesting. I can believe that it isn't immediately apparent because biological life is too complex (putting it mildly), but is that the extent of it?

We haven't found anything in nature that resembles reverse-mode automatic differentiation, either in evolution or in neuroscience.