jev-gpt
Text generation with Jev via cascaded Choice questions: word type, WordNet category, embedding cluster, then the word itself.
Overview
jev-gpt makes Jev, a model that never generates text, produce text by asking one question per choice. For each word it asks the word type, then a WordNet category, then a group of about 200 words clustered with jina-embeddings-v5-text-small, then the word itself, and one more question ranks candidate texts in a beam.
The README's sample output is a short, loosely coherent answer about the future of AI models. It reports about 400 Jev calls, 90 seconds and 2 cents per prompt, using only Python 3. A video shows a guided replay where every call and probability is real but the beam was steered toward the shown answer. It is a playful experiment rather than a practical generator.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which word type, category, cluster and word comes next? | — |
| Choice | Which candidate text is best? | — |
How it works
- Ask Jev the next word's type
- Narrow to a WordNet category and an embedding cluster
- Choose the word
- Rank beam candidates and repeat
Strengths
- Shows how far cascaded Choice can go
- No dependencies beyond Python 3
Caveats
- About 400 calls per prompt
- Output quality is limited
- Demo video is steered
Getting started
`export TYPESAFE_API_KEY=...` then `./jev-gpt "your prompt"`.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI
- Has code
- yes
Use cases and integrations
How curated lists describe it
Cascaded Choice questions that make Jev pick the next word from a word tree instead of generating text.
logicrw/awesome-jev-projects
Frequently asked questions
How does Jev write text?
By choosing each word through a cascade of Choice questions.
What does it cost?
About 2 cents and 400 calls per prompt, per the README.
Is the demo video unedited?
Calls are real, but the beam was steered toward the shown answer.
Related projects
Toy chat that makes Jev talk by repeatedly choosing its next word from a tiny caveman vocabulary.
ProjectprototypeTypeScriptLocal browser app that turns Jev's per-pixel colour probability distributions into textured paintings.
Browser extensionprototypeJavaScriptExperiment that uses Jev as a sampler: DeepSeek proposes top-k next tokens and Jev chooses which comes next and when to stop.
ProjectresearchPythonExperiment that makes Jev produce prose by asking one choice question per word over a routed domain vocabulary.
LibraryresearchJavaScriptPlayful web arena where Jev does arithmetic, writes chat replies and draws pixel art one Choice or Noul decision at a time.
DemoprototypeChatGPT-style web chat where every word of every reply is a Jev multiple-choice decision over a ~1,800-word vocabulary plus END.
ProjectprototypeTypeScript