consortai
Genesis - 03.03.23
In April of 2022 I was sitting under a palm tree on a balmy night in Miami when it hit me - we can chain these things.
I'd spent the last couple of months hanging around at a boarding house in Little Havana, strung out on an idea soup. Web3, Game Design, Generative AI, Behavior Transformation. There was so much potential back then. DALL-E hadn't yet come out and the next best thing was an iterative VQGAN driver that my friends Victor and Diego had built.
My romantic relationship was on the rocks. I'd chosen to instead hang around with the fellows at HF0, a soon-to-be-notorious residency for entrepreneurs. My partner was in San Francisco and I was here, in a chair, under a palm tree, soaking up the dizzy moisture of another Miami night.
My tour through the realms of Solidity, GAN's and card game design (stories for another day) had landed me back where I started. I wanted to build something that could transform human behavior. Something that could help you be a better you.
A long time ago there was an app that had inspired me. It was called Luna, and it was an SMS bot that would remind you when to go to bed. It was designed to create accountability and encourage healthy sleeping habits. I think it was this.
After getting a flavor for what LLM's could do (davinci-002 was state of the art at this point) I started wanting to create my own modern version of Luna. It would be an SMS bot capable of holding you accountable to your desired sleep and wake up times.
I started hacking. Building a generic bot was easy enough. But I very quickly realized that if I wanted to steer towards particular conversations I'd need to create a much more directed prompt framework.
This was when I had a realization. What if instead of treating the prompt as a one-stop-shop for generating the next line of dialog, we treated each LLM call as a single neuron in a network of many inferences that would simulate various kinds of thought and lead to a perfectly crafted response? By today's standards this is can't even be called an insight. But in April of 2022, before Langchain, before Dohan et al's seminal paper, and in my naive bubble universe, it was a breakthrough concept.
I became intoxicated by the potential of chained LLM's. At the time I used the name "HyperStructures" to describe what I wanted to do. And from that point on, the creative flow didn't stop, leading me to many software design breakthroughs that I'll share with you in future posts.
This is going to be able to do a lot more than just remind you to wake up in the morning, I thought. How much more, is up to us.