NobodyWho · 2026-09-22 · notable
Jev in 25 Lines of Python — a parody that classifies with a 0.6B local model
Duarte O. Carmo's parody post rebuilds the core of TypeSafe's Jev in 25 lines: Qwen3-0.6B via llama-cpp-python reads a multiple-choice prompt, and the logits of the answer letters become class probabilities.
The NobodyWho team shows the basic trick behind typed-decision models fits in one short local script.
What is it?
Jev in 25 Lines of Python is a self-described parody blog post from NobodyWho that answers the hype around TypeSafe's Jev decision model. Duarte O. Carmo's script gives a small local model an email and three labelled options — Legitimate, Spam, Phishing — and prints a probability for each. The post admits what it leaves out: no synthetic training data and none of Jev's calibration training.
How does it work?
The script loads the Qwen3-0.6B GGUF from Hugging Face through llama-cpp-python, runs the prompt once, and reads the model's scores for the next token. It keeps only the logits for the tokens 'A', 'B' and 'C', applies a log-softmax over those three, and exponentiates them into probabilities. Nothing is generated and no API is called.
Why does it matter?
Next-token scores over a fixed set of options are a cheap, private way to get a classifier out of any open model, and many teams forget they can skip text generation and JSON parsing. The post reached 546 points on Hacker News, and it links to fuller open Jev-style projects for readers who want the calibrated version.
Who is it for?
developers building local classifiers
Try it
pip install huggingface-hub llama-cpp-python numpy