Transluce

I'm excited about Transluce's work and would be excited to see others donate to them.

I think Transluce has a knack for identifying and working on important problems early, and has a good hit rate:

• Docent has been influential, and I think that increasing the amount of qualitative analysis of eval/agent transcripts is valuable
• Identifying end-to-end interpretability as an important research area and pushing on it, eg PCD
• Debugging weird LLM behavior
• Arguing that circuit finding works fine with neurons, no need for fancy things like transcoders.

I'm a huge fan of Transluce's work on investigator agents, automated red teaming and jailbreaking, and introducing fresh angles in interpretability/explainability. Their work has consistently been thought-provoking and among the best safety research out there. They also seem to have quite a talented and productive team as well.

Last year they were a new org. Now they have now grown to 14 people and now have a solid track record and want to keep growing. I have confirmation the team is credible. The plan for scaling themselves is highly ambitious, with planned scale well beyond what SFF can fund. I haven’t done anything like the investigation into their plans and capabilities you would need before placing a bet that big, as AI research of all kinds gets expensive quickly.

If there is sufficient appetite to scale the amount of privately funded direct work of this type, then this seems like a fine place to look. I am optimistic on them finding interesting things, although on a technical level I am skeptical of the larger plan.

To donate, reach out to info@transluce.org.