jevseek
Experiment that uses Jev as a sampler: DeepSeek proposes top-k next tokens and Jev chooses which comes next and when to stop.
Overview
jevseek turns Jev, a decision model that cannot generate text, into a sampler. At each step deepseek-flash returns its top-k next tokens with log-probabilities, Jev is shown the conversation and asked which token should come next, and the chosen token is appended. Jev also decides when the reply is finished and when it has started rambling. The README shows the result has a distinctly different voice from DeepSeek alone.
It descends from jevgpt, which drove Jev over a 20,000-word dictionary at six calls per word; jevseek asks over ten candidates at one call per token, about a second and $0.00005 each according to the README. Presets control how much Jev intervenes: the default gated mode only consults Jev at branch points where DeepSeek is less sure, while chat mode lets Jev choose every token. An optional Pangram key adds an AI-text detector score. It needs DeepSeek and OpenRouter keys and is an art and research curiosity, not a production tool.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which token should come next? | — |
| Yes / no | Is the reply finished? | — |
| Yes / no | Has the reply started rambling? | — |
How it works
- DeepSeek returns top-k candidate next tokens
- In gated mode, skip Jev if DeepSeek is very confident
- Jev chooses a candidate token given the conversation
- Append the token; Jev decides whether to stop
- Optional trace records offered and chosen tokens
Strengths
- Shows a novel use of a decision model
- Per-token trace makes choices inspectable
- Gated mode limits calls to branch points
Caveats
- About a second per token makes it slow
- Needs DeepSeek and OpenRouter keys
- Experimental; output quality is subjective
Getting started
pip install git+https://github.com/blingdivinity/jevseek, set the keys in .env, then jevseek --say "do you ever get bored in here".At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- CLI REPL
- Has code
- yes
Use cases and integrations
How curated lists describe it
DeepSeek proposes the next token, TypeSafe's Jev chooses it: a decision model used as a sampler.
hellogumbo/awesome-jev
Frequently asked questions
Can Jev write text?
Not by itself; jevseek lets it pick among DeepSeek's candidate tokens one at a time.
How expensive is it?
The README estimates about $0.00005 and one second per token.
What is gated mode?
Jev is only asked at uncertain branch points; DeepSeek supplies the obvious tokens.
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 extensionprototypeJavaScriptText generation with Jev via cascaded Choice questions: word type, WordNet category, embedding cluster, then the word itself.
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