Arcturus Labs · 2026-09-21 · notable
John Berryman — OpenAI could copy Jev, so TypeSafe's moat is its data
John Berryman argues OpenAI could add Jev-style typed decisions to its own models quickly, because Jev runs on ordinary language-model machinery. He says TypeSafe's real advantage is its training data, not its architecture.
Jev runs on standard transformer machinery, so John Berryman expects OpenAI to build its own version before long.
What is it?
OpenAI is well placed to copy Jev, argues John Berryman, founder of Arcturus Labs and a co-author of 'Prompt Engineering for LLMs'. Jev, from TypeSafe, returns a calibrated typed decision instead of generated text; Berryman's reading is that it gets there with conventional language-model machinery rather than a new architecture.
How does it work?
The argument turns on where the value actually sits. A language model already produces a probability distribution over tokens, and Berryman notes OpenAI has effectively been doing classification for years through tool calling. He sketches a path where OpenAI embeds classification inside its own models behind special syntax tags, letting a model make a calibrated judgement — for a safety check, a tool validation, or routing to another model — without leaving the GPU.
Why does it matter?
If the mechanism copies easily, TypeSafe's future rests on its data. Berryman quotes a TypeSafe cofounder saying 'we consider ourselves a data research lab' and that 'the vast vast vast majority of research was on making data that is truly general'. His conclusion is binary: a deep moat makes TypeSafe an acquisition target, and a shallow one means 'OpenAI just builds it themselves, and TypeSafe's window closes fast.'
Who is it for?
developers weighing Jev and other typed-decision models