I created https://www.addictingwordgames.com/play-game/20-questions-with-ai

The aim of the game was supposed to be to get the AI to confess that you are the winner, its possible but the game is also open ended.

The backend generation is from https://TextGenerator.app.nz which is heaps cheaper than the OpenAI models, but quality is i think somewhere between OpenAI currie and babbage.

In the prompt engineering theres some random topics picked that the user wont see, (that doesn't mean that the AI actually does think of that topic though).

Theres also some retries and repetition penalty randomness that goes up to stop looping which i think is a problem in all models right now.

in comparison to OpenAI the Text Generator API was easier to use because you can send max_sentences=1 and it will give you 1 sentence instead of trying to work out the sentence boundaries with the stop sequences (which is also supported but i dont find that as easy to work with)