Chad Scherrer

Great! Yesterday's covid booster is hitting me pretty hard, but I'll send you an update when I'm feeling better. Lots going on 🙂

> Hi @cscherrer, could you give an MWE for why the following Soss model is not type-stable? > > ```julia > For(n -> Normal(α + β * x[n], σ), 1:N)...

Looking at this a little closer, I think my approach is kind of clumsy. Do you think it could be possible to instead have an option to just avoid boxing...

Just a little more information on this, then I'm going to try to set it aside for a while... Say you start with ```julia using Soss, JET m1 = @model...

> Right now I cloned `Soss#dev` but have issues in executing `instantiate`. Sorry, I've been changing more things. I'll get to something stable and send you the Manifest. > is...

Ok, I sent you a Zulip message with lots of details

This seems maybe helpful: ```julia julia> g = mk_function(Main, :(ctx -> (j -> Bernoulli(ctx.p/j)))) function = (ctx;) -> begin begin begin let freevars = (ctx,) (GeneralizedGenerated.Closure){function = (ctx, j;) ->...

Based on the Twitter discussion [here](https://twitter.com/MoritzSchauer/status/1561635006602264576?s=20&t=FuQAzKJZF8FyRJIf1-gSew) we could also consider making kernels callable not just on values, but also on measures. Then composition of kernels would just be $\circ$, as...

Seems at least loosely related to energy-based models: https://arxiv.org/abs/2101.03288 Also, I was thinking of abstracting quadratic forms, and a quick search found this: https://github.com/JuliaSmoothOptimizers/QuadraticModels.jl

Great point. @nignatiadis has been thinking about related issues in https://github.com/cscherrer/MeasureTheory.jl/pull/226 I really like this idea of having more structured likelihoods, especially for these special cases where we can handle...