Draw the thoughts of a large language model
Draw the thoughts of a large language model
https://www.youtube.com/watch?v=BJ9BD2D3DZA
Author: Anthropic – Duration: 00:02:56
AI models are formed and not directly programmed, so we don't understand how they do most of the things they do. Our new methods of interpretability allow us to retrace their thinking (often complex and surprising). With two new articles, anthropic researchers have taken important measures to understand the circuits underlying the thoughts of an AI model. In an example of the article, we find evidence that Claude will plan what he will say a lot of words to come and write to arrive at this destination. We show this in the field of poetry, where he thinks of possible rhymes of rhymes in advance and writes each line to get there. This is powerful proof that, even if the models are formed to produce a word at a time, they can think of much longer horizons to do it. Find out more: https://anthropic.com/research/tracing-thoughts-language-model
