TypeSafe AI emerged from two years in stealth this week with $40 million in funding and a model called Jev. It cannot write an email, hold a conversation or produce an explanation. It makes decisions, and by replacing sequential token generation with decisions computed in parallel, TypeSafe claims Jev can run around 100 times faster and 100 times more cheaply than a conventional LLM on suitable tasks.
The company was founded by Diogo Almeida, who spent four and a half years at OpenAI and was a primary author of the InstructGPT research. That work used human feedback to teach language models to follow instructions and helped establish the methods behind ChatGPT. Almeida also contributed to GPT-4 and OpenAI’s early work on reinforcement learning from human feedback.
His next company starts from a criticism of the technology he helped create. Language models have become extremely good at communicating with people, but software does not usually need a conversation. It needs a reliable value that determines what happens next.
TypeSafe calls Jev a “System 1” model, borrowing from Daniel Kahneman’s distinction between two modes of human thought. System 1 makes immediate judgements, while System 2 handles slower and more deliberate reasoning. The recent reasoning models we have covered extensively approximate the second mode. They consume time and computing power to produce long chains of thought before answering difficult questions. Jev focuses entirely on the first.
A developer sends Jev some state, which can be text or structured data, followed by a set of questions and permitted answers. Jev can select from a list, assign a score or estimate the probability that a condition is true. It evaluates multiple questions in parallel and returns typed values that software can use directly.
TypeSafe's own benchmark. Jev's zero is structural: a model that can only return a value from a fixed set of answers cannot produce a malformed one. The chart shows whether outputs are well formed, not whether they are right.
TypeSafe trained Jev using a method called reinforcement learning for calibrated decisions. The objective is not merely to select an answer. Its reported confidence should correspond to how frequently similar answers are correct. TypeSafe has not published the architecture, training code or model weights, so these claims still require independent testing.
The immediate uses include routing work between agents, selecting models and tools, ranking search results, filtering context, detecting unsafe actions and checking whether generated answers have adequate support. Jev could also classify support tickets, prioritise alerts, assess transactions or moderate online communities. These are small decisions, but modern software may need to make millions of them. The name Jev is a nod to William Stanley Jevons, who argued in 1865 that more efficient steam engines had increased Britain's consumption of coal rather than reducing it, because cheaper power made coal worth using for far more work. TypeSafe is betting on the same effect: make a judgement fast and cheap enough, and software will make many more of them.
TypeSafe demonstrated Jev playing Doom by repeatedly describing the game state and asking it to select the next permitted action. Another demo, JevPilot, applies the same method to simulated driving. Software generates several possible three-second paths using the road layout, traffic and nearby obstacles. Jev receives this structured description and selects the most appropriate path, including its direction and target speed. Jev reviews the available choices as often as four times a second, and provides the repeated semantic judgement without needing to be trained on driving.
Jev is currently a proprietary API behind a waitlist, with little independent evidence and no established production record. But the buzz is significant. We can't remember a new model architecture that gained so much traction in such a short space of time, with replica versions and community sites with hundreds of projects already in existence. Our own early experience is encouraging. We have started integrating Jev into our systems, and so far it has given our agents a fast classification capability, handling quick judgements that would otherwise need a slower and more expensive call to a language model.
Takeaways: Jev does not replace reasoning or generative models. But it could handle the frequent judgements surrounding them, deciding what they see, when they run, which actions they may take and whether their work should continue. The technical claims remain unproven, but early signs are promising. AI systems need more than models that can think and talk. They also need models that can make fast, bounded decisions, and assuming they can satisfy demand, TypeSafe might have a breaktrhough on its hands.