robmay

I agree. The tech industry prides itself on "easy UX" but in reality they make it easy to do the actions they want you as the user to do, not the actions you really want to do. They optimize for revenue, not for user happiness.

We've been building SLMs for enterprise customers for over a year, and in almost every case it's pretty easy to beat a frontier model on price and performance, so we decided to put our tech behind a broader router. Plug it in, we start with a frontier model but then constantly test on other inference configurations and start routing certain workloads to those. You can see on the "Task Explorer" page a few hundred comparisons of various models on common enterprise workloads.

It's new so, we need people to test it out and tell us what doesn't work. Appreciate any feedback you can send.

Definitely - when you consider how varied inference workloads will be, and the different ways to minimize costs - better prompting, SLMs, different chips, batching, etc, there will be tons of opportunity

Best HN comment in a long time.