Discord + Slack bot that makes TypeSafe's Jev talk — a decision model that "cannot generate text," loomed word-by-word into broken sentences.
Jev is a non-autoregressive decision model. It answers questions with calibrated probabilities, not text. This bot gives it a 20K word vocabulary and asks "next word?" repeatedly via tournament sampling until it forms a reply.
- Tournament sampling: 20K vocab shuffled into 255-word buckets, all scored in parallel
- Runoff: Top-2 from each bucket compete in a final round
- Completeness judge: A separate
noulquestion asks "is the reply complete?" — jev stops when it thinks it's done - Penalty system: Content words penalized 2.5x per reuse, stopwords 1.6x — prevents "is are I is are" loops
- User-only history: Last 3 user messages included as context; jev's own broken output is excluded (it poisons follow-ups)
- "I love jazz because its improvised and freedom."
- "Rock.? Yeah"
- "No because overkill. Overkill!.!.!"
- "I depends on on situation of circumstances."
- "Yuck no ugh spit! Gag gagging ing"
- "Band is from california in san los angeles. Las angels."
pip install -r requirements.txtThe generation core lives in jev_core.py; jev_bot.py (Discord) and jev_slack.py (Slack) are thin front-ends over it. Both take --v1 for the original broken-grammar style.
Create .env:
DISCORD_TOKEN_JEV=your_discord_bot_token
TYPESAFE_API_KEY=your_typesafe_api_key
Enable Message Content Intent in Discord developer portal.
python jev_bot.pyMention jev or reply to jev's messages. Replies only — it won't respond to messages that don't involve it. Say @jev stop to mute it in that channel/thread (it reacts 🤐 and ignores replies to it there until it's @mentioned again).
- Go to api.slack.com/apps → Create New App → From a manifest, paste
slack_manifest.yaml. - Basic Information → App-Level Tokens: generate a token with
connections:write(this is thexapp-token). - Install App to your workspace and copy the Bot User OAuth Token (
xoxb-). - Invite jev to a channel:
/invite @jev.
Create .env:
SLACK_BOT_TOKEN=xoxb-...
SLACK_APP_TOKEN=xapp-...
TYPESAFE_API_KEY=your_typesafe_api_key
python jev_slack.pyUses Socket Mode, so no public URL or Request URL is needed. Jev replies in-thread when @mentioned, when someone posts in a thread jev is already in, or in DMs. Context is per thread: inside a thread jev reads the last few messages of that thread, while a top-level @mention or DM only sees that one message. It adds 👀 to your message while it thinks (Slack bots can't show a typing indicator). Jev also looks at images (PNG/JPEG/GIF/WebP, resized to fit the API's limits) attached to the message and to the earlier messages it reads as context — this needs the files:read scope, so if your app was created from an older manifest, add that scope and reinstall the app. Say @jev stop in a thread to mute it there (it reacts 🤐); @mentioning it again in that thread un-mutes it. Mutes (for both platforms) are saved to mutes.json next to the code (override with JEV_MUTES_FILE) and survive restarts.
Calls POST https://api.typesafe.ai/v1/systemone with jev-preview via the TypeSafe Python SDK (set TYPESAFE_BASE_URL to point elsewhere). Image support is dreamer-ready, so requests with images send the TypeSafe-Danger-Zone-Images: true header. Images are billed as input tokens and ride along on every call, so a reply with an image costs noticeably more than a text-only one. Tournament sampling does ~9 API calls per word, roughly 80 buckets × 255 words each; at $0.042 per million input tokens a reply is a fraction of a cent.
vocab.txt is a 20K word list (from bewinxed/jevgpt) with slurs removed. Words can be added or removed freely — the vocab IS the content filter.
- TypeSafe AI for Jev
- bewinxed/jevgpt for the tournament sampling architecture and vocab
- Built by lyra + clod