Why Does Parallel Not Call Itself an AI Lab?

Parag Agrawal says Parallel does not call itself a new lab, because a new lab's output is a model and Parallel's output is a complement to a model. Parallel builds something that multiplies on top of a model so the model gets better, or the agent built with the model gets better and has superpowers.

Every query is a 100 billion to 100 billion matching problem: return a thousand tokens from trillion web pages, and make sure they are the right thousand tokens. Until last week Parallel gave itself three seconds to throw compute at the selection at read time. Last week Parallel shipped a product that does the selection in 200 milliseconds. Parallel is not pre-training extraordinarily large models. The job is to compress things down into tiny ranking models. Parallel is trying to organize information in memory across the memory hierarchy so access stays fast, and trying to figure out how to train a model to select the best thousand tokens. A large amount of research goes into taking a pre-trained model, adapting the architecture for the new problem, and allocating compute so the best output fits a limited latency budget.

Parallel has always wanted a place where a better model from somebody else leaves a higher hill to climb. A better model now unlocks four more use cases where Parallel can be valuable.