Hacker News new | ask | show | jobs
by caleblloyd 395 days ago
Anthropic is releasing blog articles where they are discovering how Claude works through experiments and observations. It seems more like science than engineering when even the creators have to run scientific experiments to figure out how what they engineered works.

https://www.anthropic.com/research/tracing-thoughts-language...

2 comments

Isn't that more an artifact of the nature of the indeterministic black boxes that llm's are?

anthropic is doing great work, that understanding models blog post was really ground breaking.

The description of how Claude does "arithmetic" was enlightening. Shows how it uses clever pattern matching to fit the data, without ever learning the algorithm, even though the algorithm description was certainly in its training data.