Basten Heutink explains how Delphi models individual thought using language models as pattern recognizers. The system builds an adaptive knowledge graph from your writings, then generates responses that sound like you. It generated a personalized reply for a grandfather based on his actual words.
Take us through now behind the scenes, like what is actually going on when you create a Delphi? Like, how do you represent a human?
Well, I want to talk a bit about what Ray Kurzweil said in his book.
Yeah.
So in the book, he says that the mind is a hierarchy of pattern recognizers. And on the lowest end, you have— on the lowest level of the hierarchy, you have things that recognize letters and that feeds up to higher levels which recognize sentences and that feeds up to meaning and that feeds up to consciousness. And so when I was working on my first startup, Friday, which involved LLMs, I had this insight, wait, an LLM is a pattern recognizer. And so what we've ended up with is an adaptive temporal knowledge graph. Knowledge graphs are great because you can store the connections between things and you can also store the weights of confidence. Like how, how likely is it that you would actually say this? And the very important thing we had to solve before getting any customers is every person was like, what if it says something wrong and someone takes a screenshot and sends it to the Daily Mail? It's just a huge fear. And so a Allowing people to kind of choose, okay, it only says things it's trained on that directly answers the question. That's great for people like Dr. Mark Hyman, who works in the medical space. But I think the most powerful setting is it only says things it's trained on, but when posed with a new situation, it can predict what you might say in that new situation. If I were to ask my grandfather's Delphi, how would you think about running an AI company? He's never spoken about AI in his life, but in his book, it talks about him starting an oil business in Iran in the mid-20th century, and that was kind of an uncertain time in the country. So we have his principles of dealing with uncertainty, and AI is an uncertain field right now, so we can maybe reason what he might say in these new situations.
