In the metric-learn package, all algorithms currently implemented learn so-called Mahalanobis distances. Looks like for metric learning, it’s Mahalanobis or bust.

Scalable oversight via debate is underrated.

Mechanistic interp is finally tractable.

RLHF reward hacking is everywhere once you look.

Evals should be adversarial by default.

Open-weight models make safety regulation pointless.

Interpretability research is a funding sink.

Alignment needs verifiable evals, not vibes.

https://en.wikipedia.org/wiki/Riffusion

I don’t have a link, but did you see the person using an image training AI to generate music? It trained it with the waveforms then output waveforms. It wasn’t terrible.

Mahalanobis ftw

Mahalanobis ftw