Jev in 25 Lines of Python

	Jev in 25 Lines of Python(nobodywho.ai)
	673 points by bashbjorn 21 days ago \| 210 comments
	tl;dr: A parody post reimplementing "Jev" (a hyped LLM classification approach) in ~25 lines of Python using llama-cpp and a small Qwen3 GGUF model. The code prompts the model with labeled choices, extracts logits for the label tokens, and softmaxes them into calibrated probabilities—yielding a fast, local classifier without APIs, synthetic data, or RL fine-tuning.
	HN Discussion: ~Technical critique: using logprobs on chat models is unreliable without careful prompting ~Suggests improvements like placing options before content and repeating the task for better calibration ↓The post lacks rigor: no benchmarks, error rates, or latency comparisons to back its claims ↑Skeptical/tired of the Jev hype cycle and reimplementation posts being taken at face value ↑Fundamental critique of Jev's premise, including its System One framing and lack of in-domain calibration