Yeah, I saw that! Not sure about real use cases yet. Wonder what their "ChatGPT moment" could be?

Also curious about the actual performance, but I don't have beta access. I hope it will be open-sourced in the future. Fully structured output is exactly what I'm looking forward to.

Its not an LLM. In a simple terms. Its a zero shot classifier. Given a bunch of enums, it can output the probabilities. Because the output is just one token from the list of enums, there are no hallucinations. I think the only real advantage is the cost and latency for classifications tasks.

IMO this prototype of using Jev via DuckDB extension gives strong BAML-like vibes: https://github.com/colliber/duckdb-jev I'm interested to see what a direct BAML integration could unlock.

I'm working on something that combines Jev and BAML, pretty excited about it!

jev + baml looks like a great match. interested to see what we can create