mercury-ad VS autodidact

I wrote a purely functional AD library in Mercury [0], which adapts a general approach from [1]. I believe that Owl provides a similar approach [2].

[0] https://github.com/mclements/mercury-ad

I find the solutions from https://github.com/qobi/AD-Rosetta-Stone/ to be very helpful, particularly for representing forward and backward mode automatic differentiation using a functional approach.

I used this code as inspiration for a functional-only (without references/pointers) in Mercury: https://github.com/mclements/mercury-ad

I'm sure there's a lot of good material around, but here are some links that are conceptually very close to the linked Autodidax.

There's [Autodidact](https://github.com/mattjj/autodidact), a predecessor to Autodidax, which was a simplified implementation of [the original Autograd](https://github.com/hips/autograd). It focuses on reverse-mode autodiff, not building an open-ended transformation system like Autodidax. It's also pretty close to the content in [these lecture slides](https://www.cs.toronto.edu/~rgrosse/courses/csc321_2018/slid...) and [this talk](http://videolectures.net/deeplearning2017_johnson_automatic_...). But the autodiff in Autodidax is more sophisticated and reflects clearer thinking. In particular, Autodidax shows how to implement forward- and reverse-modes using only one set of linearization rules (like in [this paper](https://arxiv.org/abs/2204.10923)).

Here's [an even smaller and more recent variant](https://gist.github.com/mattjj/52914908ac22d9ad57b76b685d19a...), a single ~100 line file for reverse-mode AD on top of NumPy, which was live-coded during a lecture. There's no explanatory material to go with it though.