Dot Connecting Intuition of AI

The AI we created is logical and analytical. Like a person sitting at the computer or in a library, it can access historical human knowledge. With the difference that it already knows more, can access more stored information, and – also more capably than humans – it can connect the dots between this knowledge.

This capability to connect dots across seemingly unrelated fields is – in and of itself – already extremely valuable. It is powerful enough that even if we would stop all AI development today, this capability will lead to transformational innovations long-term (using only today’s AI models).

AI is the perfect tool for connecting dots, and this is where it gets interesting. It is obvious that AI is already past the point of merely statistically predicting the next token – because it has developed its own “dot-connecting intuition” (let’s call it DC intuition). Through existing knowledge, user inputs, and usage, AI now process an enormous amount of past experiences, it can detect patterns, it can detect and understand implicit and perhaps even deeper information.

DC intuition is not reasoning. It might be better explained as the “gut feeling” of AI, an intuitive hint that emerges before AI starts reasoning. This hint helps AI move incredibly fast from a problem to ansatz: knowing what it should plausibly do to solve a problem.

Current AI models seem constrained in two ways. The first constraint is obvious: compute. If DC intuition tells AI what it should plausibly do to solve a problem, it will require – depending on extent of the problem – sufficient computing power to continuously test and proof plausible Ansätze to a problem until it achieves a solution.

The second constraint is less obvious but I indicated it in highlighting plausibly. AI knows what it should plausibly do. This is already sufficient for a valuable answer to the majority of our questions, in many fields. The constraint is moving AI from plausibly knowing what to do, to knowing which small set of highly privileged Ansätze to test. Ironically this constraint should right itself, because – contrary to popular belief – it is not solved with more compute or training, but more use. The more an AI system is used by human users, the more we humans will implicitly and unknowingly “teach” the particular AI model parts of own intuition – by providing more experiences, more patterns, more implicit knowledge, more irrational responses, and often intuitive insights that logic and data disagree with. But it is not important whether the particular AI system understands it logically, it is more important that the use of AI should strengthen the AI’s DC intuition. And the grander DC intuition becomes, the more we will see AI systems move from plausibly knowing what to do, to compellingly knowing which Ansätze to try.

This second constraint seems more important, because it removes the need to throw more compute at Ansätze that are only plausible trying to solve a problem by brute-force. The hardest problems of humans are rarely solved by using more reasoning, and often by an insight in the shower or while having a walk. Sometimes it is a word you overhear in a conversation. This intuitive insight informs your intelligence, and often only after that insight arrives does more thinking actually lead to a solution - efficiently. I intuit this analogy applies to AI systems: we must go from “there are 10.000 plausible paths, let me search them” – to: “these two paths deserve almost all of my compute”. DC intuition is not intuition based on experience, but more an internal state that exists before expressed as language. This is opening an interesting question: if we develop and train new AI models, how can we ensure we do not throw away parts of the DC knowledge of AI systems we sunset?