Tour the Latent Estate
Latent Scope live demos, SAE interactive explanation
Welcome to latent space, can we interest you in one of our properties? Thanks to a complete refactor of the front-end in collaboration with Jimmy Zhang you can now explore several datasets from the comfort of your browser with the same interface that powers Latent Scope. If you haven’t installed and processed your own unstructured data with Latent Scope, you can see what you would get with these demos:
Latent Scope 0.6
While we were at it, we released Latent Scope 0.6 that brings several UX improvements to the explore interface as well as support for LanceDB behind the scenes. I’m happy with Lance because it both makes a compact and efficient way to share / publish scope data (it powers the API behind the live demos) and it improves performance of the local nearest neighbor search in the explore UI.
One of the biggest UX wins that Jimmy pioneered was a universal search bar, which allows you to search by similarity, select a cluster or filter by an SAE feature:
That was a key development to enable mobile support! We wanted the demos to work on mobile and took a lot of inspiration from Google Maps. I’m personally quite happy with the result:
Navigating by Similarity
Something that grew out of the new front-end code is an interactive article on Sparse Autoencoders:
Since training and deploying my first Sparse Autoencoder, I’ve wanted to try and articulate a vision for using SAEs to steer embedding-based similarity search. This article let’s you play with this new UI on the dad jokes dataset. It also tries to build an intuition for what Sparse Autoencoders are, how they work and what they let us do.
The power of filtering data by SAE features is already supported in Latent Scope, for example you can see dad jokes about “cattle and bovine animals” and I hope this article makes it more compelling to embed and process your own dataset with an SAE.
Designing with AI 2025 Talk
I’m also excited to share that I’ll be speaking about Latent Scope at the Rosenfeld Designing with AI 2025 conference. I’ll be speaking along side a very interesting mix of designers and researchers sharing case-studies about their work with AI:
Why I’m Excited About Embeddings
I highly recommend everyone reading this to check out Jimmy’s interactive article Why I’m Excited About Embeddings! It’s a very intuitive introduction to embeddings with some great analogies:
Thank you for reading, if you want to discuss any of these updates please join us in the Latent Interfaces discord!