Graham Gibson

@bdeonovic That sounds good to me! How about the old faithful data? [here](https://www.math.ucla.edu/~anderson/rw1001/library/base/html/faithful.html) I have seen that used in a couple GMM examples, would that make sense to you?

@bdeonovic That makes sense, but the custom samplers still require some return ```julia Gibbs_sigma2 = Sampler([:sigma2], (mu, sigma2, y) -> begin a = length(y) / 2.0 + shape(sigma2.distr) b =...

How do you get access to the underlying chain to return the last accepted step, I guess you don't need it you can just return the input to your sampler?

Awesome! Thanks for your help! The matrices are PSD until s^2 gets sufficiently small

Unfortunately, I can't use Gibbs sampling because the conditional density is not available (unlike Bayesian linear regression , gaussian processes don't have an analytical distribution for s2 that I know...

But then I would need a proposal dist etc right? I would basically have to re-implement MH in that code block? Gibbs seemed nice because it was just actually sampling...