Lukasz Stafiniak

A ray tracer in tutorials and benchmarks. Maybe try another tutorial where the ray tracer is optimized on some metric.

MNIST classifier in tutorials and benchmarks.

People complain that model surgery in TensorFlow v1 is hard, making TensorFlow v1 is unsuitable for research. We should do better despite being fully compiled.

That would be the second-simplest complete example, after the `micrograd README` example.

Maybe use the language modelling setup from the Transformer example.

It's used by PyTorch v2. Related: CUDA backend issue [#26](https://github.com/lukstafi/ocannl/issues/26), C backend issue [#43](https://github.com/lukstafi/ocannl/issues/43).

[Google Colab notebook](https://colab.research.google.com/drive/1SiF0KZJp75rUeetKOWqpsA8clmHP6jMg?usp=sharing#scrollTo=wW1-8xqswRYg)

See [ocaml-onnx](https://github.com/LaurentMazare/ocaml-onnx).

Based on [Automatic gradient descent](https://github.com/jxbz/agd).

Have a story about / support for experiment tracking: graphs of observables e.g. loss, device health

2

This needn't be the core library though...