TypeSafe AI

TypeSafe AI is an artificial intelligence lab that builds machine-native intelligence infrastructure for automation, designed for software to make decisions instead of generating text for people.

It was founded by Diogo Almeida, co-inventor of RLHF (the method behind InstructGPT, ChatGPT and GPT-4), together with Sasha Sheng and Erik Gafni. The company is based in San Francisco, California. Its team comes from companies such as OpenAI, Google Brain, Meta/FAIR, Stripe, Airbnb, Plaid and Docker.

Its first product is Jev, the first model of a category the company calls System One. It is designed to evaluate questions about the state of a program and return structured results — an option from a list, a score or a probability estimate — together with a confidence level in the first two cases. Unlike conventional language models, it does not generate free text: it produces responses that code can consume directly, without further interpretation or parsing. Training is based on a proprietary architecture and a method called Reinforcement Learning for Calibrated Decisions (RLCD).

TypeSafe AI holds that AI development has focused excessively on conversational assistants, while much of production software still relies on hand-written logic for tasks such as classification, routing or information extraction. Its approach treats intelligence as a component that can be invoked directly from code, in a way similar to a database query, and combined with the rest of a software system to handle decisions that would otherwise require manual rules.