Jev: The Language Model That Won't Talk

No, it constrains output shape, not judgement quality. A model that can only emit a pick and a probability will always parse, which is genuinely useful, but it says nothing about whether the pick was right. The piece's own numbers show Jev behind stronger models on accuracy.

Because it is near-parity with a far costlier path, 67.9 percent at $0.0304 and 10.1 s against $0.0004 and 0.4 s, while Jev still trailed GPT Sol at 74.1 percent and Claude Opus 5 at 73.1 percent, and lost badly on invoice processing at 61.8 against 79.1. The pitch is cost at parity, not better decisions.

Not much, and the author says so. The reference labels averaged two other models rather than operational ground truth, on workflows TypeSafe itself designed, so it measures agreement with model opinion rather than correctness against what actually happened. He is explicit that he wants independent testing.

No, calibration is a population property. Across many predictions, answers labelled 80 percent should be right about 80 percent of the time, which says nothing about any single call. A model that always predicts the base rate can be perfectly calibrated and still useless, so accuracy and easy-versus-hard discrimination still matter.

Diogo Almeida, the TypeSafe founder. The piece records him as an equal-contribution primary author of the InstructGPT paper and a contributor to GPT-4, and calls that more accurate than the co-inventor of ChatGPT positioning used in company material on X. It is the reason the piece gives his critique weight.

The piece separates three objectives. RLHF optimises for what humans prefer, RLVR optimises against outcomes you can check, and RLCD is supposed to optimise decisions so their probabilities match how often those decisions turn out right. The author's point is that automation needs uncertainty which can govern behaviour, and reassuring language is not that.

Not yet. TypeSafe says questions are evaluated independently and in parallel with little added latency as the count grows, and the author says he would want that measured by someone other than TypeSafe. He does accept the shape of the argument: ask for intent, frustration level and whether a refund was asked, then apply the policy in code.