Alignment is not free: How model upgrades can silence your confidence signals v variance.co 121 karinemellata 1 year ago 67 comments
We used sparse autoencoders to explain LLM moderation flags of violent threats v variance.co 6 karinemellata 1 year ago 0 comments
At least for investigative workflows (where it's pretty common for data scientist to often have to rebuild custom dashboards), it'd be really cool to have the system handle prompts such as "give me a regex on email address that seem to incur the most amount of fraud" -- or unveil some insight on the data that is usually just a mix of operator/data scientist domain specific knowledge + pattern matching.