Marcus Hutter: Universal Artificial Intelligence, AIXI, and AGI | Lex Fridman Podcast #75

> Are they? They are actually. It's a kind of revelation when you realise this statement is true. There are people in AI who have thought long about the problem of intelligence and had this realisation. For example: [link] Then there is the interview with Hutter himself: [this video] You can even semi-formally derive this from Bayes' rule. Let's consider intelligence to be the ability to predict future data from past data. This requires building a model for the past data that can make predictions. So intelligence produces the most likely model M given data D: argmax_M P(M|D) which from Bayes' rule is: argmax_M P(M|D) = P(D|M)*P(M)/P(D) You convert probabilities to information by taking the negative log: argmin_M -log(P(M|D)) = -log(P(D|M)) + -log(P(M)) We lost -log(P(D)) because we consider D constant so it can't influence the minimum. You can read the above informally as: The most likely model (the most general model) that can make predictions from data D is that where the (encoding of the model with the least information) plus (the encoding of the data using the model) is minimal. This is essentially a statement about compression. I'm sure an AI person could do this better. This is also known as: [link] or in philosophy as: [link]