User repository on GitHub
Turns Jev into a chatbot
- Forks
- 5
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jevchat
A Python experiment that turns Jev into a chat model by repeatedly asking which symbol comes next, including alphabets, sampling modes, and a bench command.
- Stars
- 97
- Language
- Python
- Published by
- Community
- Author
- kyle-pena-nlp
- Last verified
- Sep 2026
About this project
jevchat turns that into a chat model. At every step it asks Jev one question: The options are an alphabet plus an option to stop emitting. Jev returns a probability for each one, and the sampler draws the next symbol from that normalised distribution. Append, repeat, and stop when STOP is drawn. There are several alphabets (including truncated token lists) and sampling strategies available. The idea is for fun, the cost is somewhat impractical, and the results are hilarious.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
Highlights
- Each step asks Jev which symbol follows, then samples from the returned distribution until STOP.
- poetry run jevchat bench compares modes; live panel shows symbols per second and top logits.
- Accepts api_key in .env or JEV_API_KEY / TYPESAFE_API_KEY from the environment.
- Author describes the idea as fun, somewhat impractical cost, and hilarious output.
Watch out
No license file in the repository snapshot; treat as demo code only. Claude-assisted experiment per README.
Questions about this entry
Short answers drawn from the linked project or page.
A sampling experiment. Each step is one question asking which symbol comes next, so a reply costs one API call per character, and the README states the idea is for fun, the cost is somewhat impractical, and the results are hilarious. It was a Claude-assisted experiment in sampling algorithms and strategies, not a chat backend.
Any alphabet larger than 255 symbols cannot go into a single Choice question. Those need the buckets strategy, which splits the alphabet across many questions with an OTHER escape, and the README notes buckets is the only strategy that can hold more than 255 symbols; refine builds on it.
No. The 158 tests are entirely offline, using a scripted fake client for the generation loop and an httpx.MockTransport for the HTTP layer, with no API key and no network required. jevchat bench is the part that actually hits the API, so it is the command that will fail without a key.
Presenting finished strings instead of bare symbols. Under --presentation hypothesis the options are the resulting texts, so Jev only ranks completed strings; the README calls it the single largest improvement in the project, roughly tripling top-1 and doubling the probability mass on the right symbol while using fewer input tokens than per-option descriptions.
At each step it asks which symbol comes next given the question and reply so far; the options are an alphabet plus an option to stop, and the sampler draws the next symbol from Jev's normalised probabilities.
choice asks one question over the whole alphabet, bisect asks earlier/later yes-no questions down to small groups, buckets splits the alphabet across many questions so it can exceed 255 symbols, and refine rescoring the winners.
Lowercase-plus-space, a full ASCII alphabet, token lists, and larger word or BPE vocabularies (words1k, bpe2k, bpe5k) that exceed 255 options and so need the buckets strategy.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
Hacker News: I turned Jev into a (lousy) chatbot
kp1197173 points · 49 commentsI turned Jev into a (lousy) chatbotRead the thread on Hacker NewsMore Jev resources like this
110.5kGitHub stars
Multi-agent trading research with optional Jev screening of StockTwits and Reddit posts for company relevance before sentiment analysis runs.
Jev Tools & CLIs#community#python#trading
Community
TauricResearch/TradingAgentsPython
TauricResearch/TradingAgentsPython29.3kGitHub stars
The Cua computer-use platform documents a jev-use recipe where TypeSafe Jev picks bounded action IDs and Cua Driver executes clicks and fills in the browser.
Jev Demos & Experiments#community#browser#computer-use
Community
trycua/cuaRust
trycua/cuaRust2kGitHub stars
A desktop AI assistant with a Jev / System One inline plugin for classification, routing, verification, and scoring without chat-style output.
Jev Tools & CLIs#community#python#desktop
Community
szczyglis-dev/py-gptPython
szczyglis-dev/py-gptPython523GitHub stars
Classifies and splits PDFs, DOCX, and PPTX with local page text from LiteParse and category decisions from Jev.
Jev Tools & CLIs#community#python#document
Community
jerryjliu/docjevPython
jerryjliu/docjevPython97GitHub stars
A Playwright-based CodexQA skill where Jev chooses indexed controls on a page-built index while the browser executes the action.
Jev Demos & Experiments#community#browser#playwright
Community
openqa-cn/jev-browserTypeScript
openqa-cn/jev-browserTypeScript36GitHub stars
A Clash Royale bot that turns a phone screenshot into JSON once a second and asks Jev for a strategy, a card, and a square.
Jev Demos & Experiments#community#python#gaming
Community
bytelabs-oss/clash-jevPython
bytelabs-oss/clash-jevPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.