Jev GPT

Jev only answers typed questions with probabilities. jev-gpt asks it one question per choice: word type, then WordNet category, then a group of about 200 words split by jina-embeddings-v5-text-small, then the word. Code assembles the chosen words into text, and one more question ranks the candidate texts.

First a question for the word type, then one for the WordNet category, then a choice inside a group of about 200 words split by jina-embeddings-v5-text-small, and finally the word itself. A final question ranks the candidate texts the terminal could continue with.

The README states about 400 Jev calls, 90 seconds and 2 cents per prompt. Those are the author's figures for the example prompt rather than a measured range across tasks, and the only billed component named is the TypeSafe API key.

Export TYPESAFE_API_KEY from console.typesafe.ai, then run ./jev-gpt with your prompt as the argument — for example ./jev-gpt "Think step by step: what is the future of AI models, Jev, Claude or GPT?". There are no dependencies beyond Python 3.

It is a guided replay: every call and every probability is real, but the README says the beam was steered to the answer shown so the run ends there. The judge panel shows the text the terminal follows next to Jev's own top picks, so you can see where they differ.

The README documents no output-quality benchmarks; the example answer is one prompt. Generation is word by word, and word groups are split by jina-embeddings-v5-text-small, so wording comes from choices Jev ranks rather than free-form text. Treat it as a demonstration, not a writing tool.

It is a small project by florian-hoenicke, published as a demo of what a probability-only model can do when code reconstructs text from choices instead of sampling tokens. The repository includes the full-quality MP4 of the replay alongside the GIF.