We love to discover and learn from personal blogs, essays, and articles on the internet. But despite exciting advancements in language modeling, it’s felt harder to discover interesting links from real people.

We believe there is untapped potential to use these language models as “curators” rather than “creators”. To explore this, we’re building a recommendation system that curates content for you based on the links you visit from our site. You can also prompt it with things like “serendipity in art” or “I want to learn more about the intersection of painting and programming”.

Under the hood, we’ve trained a decoder model (llama) to understand how people recommend sites through hyperlinks and used it to embed ~40 million sites. These embeddings are served in a custom database we’re building to run fast on commodity hardware, work well with both vector and structured data, and serve more expressive representations of the sites (multi-vector, multimodal, etc).

We’ve had a blast using it internally, and early users have mentioned it reminds them of Stumbleupon.

We still have a long way to go and would love to hear your feedback! There are no sign-up barriers to get started.