AI's Buzzy 'Jev' Is Just 25 Lines Reading Logits From a Local LLM

A tongue-in-cheek post from the team behind NobodyWho deflates the hype around ‘Jev,’ a supposedly revolutionary LLM paradigm, by reimplementing it in 25 lines of Python. The trick is unglamorous: load a tiny local model (Qwen3-0.6B in GGUF format) via llama-cpp-python, feed it a prompt with lettered answer choices, then pull the raw logits for the A/B/C tokens and softmax them into probabilities. In the demo, an email describing a payroll password request on a non-company sign-in page gets scored 88.5% ‘Phishing’ over ‘Spam’ or ‘Legitimate.’

The joke is aimed squarely at AI marketing. There’s no ‘System One decision model,’ no API call, no synthetic training data, and no reinforcement-learning calibration scheme — just constrained classification by inspecting the model’s own token probabilities. The author concedes the numbers aren’t always well calibrated, but the approach is fast, runs entirely on-device, and keeps data off third-party servers.

The piece is explicitly a parody (it links to ‘more complete’ open implementations as a bit) and doubles as promotion for NobodyWho’s privacy-first, open-source stance. The underlying technique is real and genuinely useful, though: reading answer-token logits is a lightweight, local way to do zero-shot classification for tasks like spam and phishing triage without shipping sensitive text to an external LLM provider.