Hey Product Hunt! 👋 For years, we've inappropriately used billion-parameter text generators to make choices. Jev, from TypeSafe AI, is a "System One" model. You give it a situation and a list of options, and it returns a probability for each one in milliseconds. It never writes a sentence. JevGPT makes it write sentences anyway. Every word of every reply is a multiple-choice question: Jev picks the next word from a 1,772-word vocabulary, the app adds it, and asks again. This is a familiar problem for me. In 2007, I published a NAACL paper on generating text with one model to select words and another to score their ordering. JevGPT works the same way: Jev shortlists candidate next words, then scores full continuations to pick the best order. Does it work? Sort of. - It says strawberry has three r's (sometimes) - It told me to walk to the car wash - Asked if it's conscious, it said "No." Try it at https://curata.com/jevgpt with your own TypeSafe API key (about a cent per reply), or run it locally from the open-source repo: https://github.com/idlivada/JevGPT I'd love to hear the weirdest thing you get it to say, and any ideas for making a choice-only model write better. 🙏

JevGPT

A chatbot built on a model that can't write

A chatbot built on a model that can't write