My department in Edinburgh is hiring a lecturer (assistant prof) in Software Engineering: elxw.fa.em3.oraclecloud.com/hcmUI/Candid...

The 13x13 grid became 14x14. It's really hard to get image models to make a board with the correct number of evenly spaced lines and star points. I'd thought it might copy a reference board into a scene, but it can't even do a simple copy, showing limits to guidance that's possible with this model.

A screenshot of a virtual 13x13 go board with some go stones on it.

ChatGPT's reconstruction of a 13x13 go board is actually 14x14 and it has changed the configuration of the stones.

Image models struggle to create Go boards correctly. I thought ChatGPT may be able to copy a correct one into a scene, but conditioning the model is through too small a bottleneck. "Please copy this image exactly to create a new image. Every line and stone should appear in exactly the same places."

Standard normalizing flows are "approximate" too: e.g., standard implementations are often not exactly invertible in practice. But ODE solvers can use adaptive computations to control the errors. So they're not intractable in the same way that (say) a Boltzmann machine is.

@ramandutt4.bsky.social 👋

I do ML + Bayesian inference.

Yes please.

Yes please.

👋