GPT-4 is all anyone can talk about in tech (outside of, say, Elon Musk destroying Twitter from within), particularly with prompts and what inputs get the best out of the language model. But The Stochastic Parrot goes the opposite route:
I got access to the GPT-4 model on Friday (17th March), and I had this up over the weekend.
Essentially it’s a GitHub action that runs every hour and sends an empty prompt to the OpenAI API. I tweeked some numbers (I found setting the frequencies to 1 gives a good balance, too low and you get things like multiple repeating commas or words like and, the, etc. The only random number (using crypto random) is the temperature. I was using 200 tokens and I was getting text always around that – bumped it up to 500 and now I get much more of a range from 60s to near 500
Once it gets the response, I save it to a file then I post it on Mastodon – the script posts up to 500 characters, and creates a thread for longer content. Once that’s done the Astro site build is triggered to update – I’m looking to make it support SSR so I can deploy it to somewhere with functions to make it searchable
via Reddit
Outputs from empty prompts? Now there’s a concept! And the outputs appear random, which makes them so interesting.