Jev in 25 Lines of Python: a Parody That Questions TypeSafe's Claims

On September 22, Duarte O.Carmo of the open-source project NobodyWho published "Jev in 25 Lines of Python" — a post that reproduces the core behaviour of Jev, the decision model TypeSafe AI announced on September 14, using a small local model and about twenty-five lines of code. The post is openly labelled a parody, and it has become a reference point in the online debate about how much of Jev's advantage is architecture and how much is training.

What TypeSafe announced

TypeSafe AI, founded by former OpenAI researcher Diogo Almeida, left two years of stealth on September 14 with Jev, the first of its System One Models. Instead of generating text token by token, Jev takes unstructured state plus typed questions and returns structured answers with calibrated probabilities and confidence scores. The company trains it with Reinforcement Learning for Calibrated Decisions (RLCD) and claims two orders of magnitude better speed and cost than comparable LLMs: 70–500 ms per response against 3–329 seconds for frontier models, and $0.042 per million input tokens, with output too cheap to meter. Because the possible outputs are declared in advance, TypeSafe says, Jev cannot make type errors or hallucinate. System One is named after Daniel Kahneman's fast thinking; the model is named after the economist William Stanley Jevons.

What 25 lines can do

The NobodyWho post loads Qwen3-0.6B, a 0.6-billion-parameter open model, through llama-cpp-python, feeds it a payroll-phishing email with three labelled answers, reads the logits for the tokens "A", "B" and "C", and converts them into probabilities with a log-softmax: Legitimate 3.1%, Spam 8.4%, Phishing 88.5%. No API call, no reinforcement learning, no synthetic data, and nothing leaving the machine. The author notes that the post is a parody and points to open reimplementations: OpenJev, openjev-sglang and an OpenJev build on DiffusionGemma.

Why it matters

It has passed 450 points on Hacker News not because it disproves Jev, but because it isolates the inference trick — reading the probabilities of a few label tokens instead of sampling text. Arcturus Labs made a related observation on September 16: a fine-tuned LLM whose logprobs over label tokens are read directly reproduces much of Jev's behaviour, an approach the analysis credits to developer Harsha Gundala. That analysis leaves the harder questions open. Jev answers many questions in one parallel pass where the 25-line version answers one, and its calibration claims rest on RLCD, which the analysis singles out as the piece TypeSafe has not explained. Accuracy on general real-world tasks is, so far, unproven.