Say-hi Tests Whether a Choice-Only Model Can Learn to Talk

Summary

The open-source say-hi project tests whether TypeSafe's JEV model, which selects from listed options rather than generating text, can still hold a conversation. It sends JEV a situation and question, then lets the model choose goals, answer words, and individual keyboard keys; a SEND key ends each reply. The system plans the reply, runs a tournament over roughly 10,000 candidate answer words when needed, and then issues one request per key, with hints showing previews and possible word completions. A full reply is capped at 400 keystrokes and $0.50, while the API key remains on the local server and calls are logged without the key. In the project's observations, JEV selected answers such as 8, Einstein, and 2 × 10^30 kg with high confidence when offered whole words, but struggled to spell the answer from an uninformative first letter and produced word salad for open-ended questions. Preview and hint changes improved spelling from about 70% to nearly 100% in the reported tests. Across 16 measured iterations of an 18-turn scripted English conversation, the score rose from 2.47 to 2.77 out of 3, though advice, unfamiliar vocabulary, and some arithmetic remained weak. The repository also compares the cost of a short "Hi" reply across JEV and several listed language models, and includes mocked tests plus a paid live evaluation script.