OpenAI is well positioned to fast-follow Jev

A September 21 analysis argues that OpenAI is well placed to copy TypeSafe AI's Jev, a model that answers questions about text with typed, probability-based decisions rather than prose. The author's reasoning is that OpenAI's models already make small bounded choices internally, such as whether to call a tool and which one, so they could be trained to expose that kind of classification more generally.

The harder part, according to the essay, is not the interface but whether the decisions can be trusted. The author reports finding domains where Jev's probabilities did not hold up, and locates TypeSafe's possible advantage in its training data and training process rather than its architecture. The proposed architecture and the competitive outcome are the author's hypotheses, so the supported distinction is between copying an interface and reproducing dependable decisions.