Large language models are incredibly sophisticated but not always perfect in their responses. To improve their accuracy, people have experimented with various techniques to guide the model’s output. From pre-training on larger datasets to using vector databases for context, these methods have shown some enhancement, but there is no fool-proof approach yet.
One intriguing method to steer the model’s responses is vector steering. A fascinating example of this is the Claude Golden Gate Bridge experiment, where Claude always finds a way to bring up the Golden Gate Bridge regardless of the user’s query.
In this article, we’ll explore the research conducted on vector steering and delve into Anastasia Borovykh’s incredible code implementation. For a deeper understanding of this concept, I suggest watching her informative video here.
Are you ready to explore the fascinating world of vector steering in language models? Let’s dive in!