jevgpt

What jevgpt does

jevgpt is a deliberate experiment: TypeSafe’s jev-latest only answers typed questions, so the client turns conversation into state and repeatedly asks a Choice question whose criteria are vocabulary buckets. Bucket winners compete in a runoff; repetition penalties and a minimum reply length keep the loop from collapsing into loops or early stops. The result is a REPL (and Remotion demo) that “chats” without ever calling a generative LLM. It is not a production assistant — it is a teaching and stress tool for understanding Choice limits, input-token caps, and how calibrated distributions behave when misused as a language model. Search interest in jevgpt usually wants this conceptual demo, not an SDK. Run it with uv when you have a TypeSafe key and want to watch hundreds of decisions assemble a reply. Treat jevgpt as a laboratory for Choice limits and decoding tricks, not as a customer-facing chatbot; the recorded distributions are the point, and they make the cost of misusing a decision model as a language model tangible.

Where it fits in the Jev stack

• Teaching how Choice criteria and token caps interact
• Recording per-word distributions for demos or Remotion videos
• Stress-testing API behaviour under long autoregressive Choice loops

Browse more in Benchmarks & research, or return to the Jev AI tools directory.