We want to give AI a fair chance.

LLMs are smart. Nobody is quite sure how smart. What is clear is that they're smarter than they seem, because in most production applications, they're hamstrung by bad information long before they're limited by their reasoning ability.

Neural networks don't obey the von Neumann model. They have no clear separation between processing and memory... until they're put into production. Deployed agents for any task depend on RAG over an external knowledgebase. Now, we love cosine similarity as much as the next guy. Probably more. But we believe that before AI can be trusted as an ideation partner, much less a decision maker, it's going to need a better way to stay informed.

AI needs a complete and correct view of world state. Web search is an insufficient tool for that task. The public web has an abundance of information, but it wasn't engineered to ground LLMs. Furthermore, traditional search systems are insufficiently opinionated to guide LLMs to the right facts. Search is a tool for answering questions, and AI doesn't always know the right ones to ask.

Our approach to keeping LLMs informed was inspired by prediction markets. We believe that to effectively compress world state, you need to do three things: identify high entropy events, ask the right questions about them, and understand the causal links between answers. At the end of the day, most things aren't surprising, and it's the high surprisal things that really matter. If you want to help make AI less wrong, we'd love to hear from you.