Will Tebbutt

> It seem like this approach could make it more flexible about types, so you could put easily use StaticArrays, for example. Is that right? Exactly. Indeed, some `SparseMatrixCSC` functionality...

No strongs views on my end -- activity analysis isn't presently really a thing in Tapir.jl, so happy to take the lead from others.

I don't have a strong view on exactly what structs you should test, but I do know of several things that you will need to make decisions about, based on...

@gdalle did you ever think any more about this? The release of 1.11 has prompted me to restart this discussion because `Array`s are now generic structs. This is relevant because...

Fair enough. In that case I'll ignore this issue until the upgrades are done, and figure out how to make everything work on the DI end when we get to...

It almost certainly won't by default. edit: I say "almost" because I'm not 100% sure what the best choice is from Mooncake's perspective yet.

Great stuff. I've spoken with @yebai and he's also on board with this, so I'm happy to press forwards when the above versions have been yanked.

Yup -- see discussion [here](https://github.com/SciML/ADTypes.jl/pull/75#issuecomment-2265237133).

Hi @mjyshin thanks for opening this issue! The issue you're encountering is to be expected, and will be found in any GP package. It's a consequence of unavoidable numerical error...

So we actually do do this -- see [here](https://github.com/JuliaGaussianProcesses/AbstractGPs.jl/blob/f4473ef49d6f29c60c4fefa0db0ea27684bcf8bc/src/finite_gp_projection.jl#L17). The problem is that different kernels + different sets of inputs require different values of this, so if you're training kernel...