Why I'm excited about Jev-like classifiers

Check this out – how fast Jev vs Claude find a product recommendation. Both make an API call based on your query, both are just prompt engineering on our end. No training, no piles of evals, no months of collecting data, huge difference in speed and cost. Jev is even fast enough that it can [poorly] play a game in real time. Each key press here is an API call with the observation of current game state and a choice for what key to press. Continue reading →