Giving Jev a voice
We have gone full circle: I got Jev to write token by token.
Jev is TypeSafe’s decision model. You give it a situation and a list of options, it picks one. It does not write text.I wanted to see how (in an absurd way) how far we could take this.
The idea
Give Jev the English alphabet as its options, plus a space and a stop. Ask it a question. Every letter of the answer is a separate decision: here is the question, here is what you have written so far, which letter comes next? Then ask again, and again, until it picks stop.
How well does it spell?
Once it is a few letters into a word, it mostly finishes it. Getting started is the hard part. I tried shuffling the options, offering them in smaller groups, and letting it look one letter ahead. None of it helped.
It spells what it can see
The most interesting part came from just playing with it. Ask an open question and Jev copies the question back:
Are dogs real animals? ARE ARE ARE ARE ARE
Put the answer in the question and it finds it:
Are dogs real animals, no or yes? YES
Is ice hot or cold? IS COLD
What is bigger, a mouse or an elephant? A ELEPHANT
Cost
Each letter is one call of about 750 to 1,000 input tokens. A short answer like YES costs a fraction of a hundredth of a cent. Answers are capped at five words, so the longest possible one is about $0.002.
Try it
The code is on GitHub: Jose-Sabater/jev-ouija. No dependencies, one server file and one page. You need a TypeSafe key with access to Jev.
Superintelligence has finally arrived.
Typewriter sound: Erika 5 by Work With Sounds, CC BY 4.0. Bell by _stubb, CC0.