I've been working on an open-source implementation of Anthropic's research on monosemanticity ("Towards Monosemanticity"). The problem Anthropic is trying to solve is that language models are hard to interpret because individual neurons can be responsible for multiple different things. The research finds that training a small autoencoder on neuron activations can result in "features" which are much easier to interpret.

When I was reading the original research, I got really excited when I realized that the models they used were really small, and I could probably train them from scratch with just my M3 MBP. My models are somewhat undertrained compared to what Anthropic produced, but I think my results are still very compelling. Let me know what you think!