Machine Super Intelligence Lecture (Final)
In my last post about general AI, I described Solomonoff Induction (with a more detailed description in Shane Legg's videos), which theoretically creates the most intelligent machine possible. Such a machine cannot be built because it requires knowledge of Kolmogorov Complexity, which is not computable. So, time to give up and go home? Not so fast!
The Kolmogorov Complexity is actually known for the simple mathematical formulas, and research in theoretical mathematics and computer science can give us better and better approximations for the Kolmogorov Complexity of more and more hypotheses.
But how are we going to work out all of this math!?! The answer is that we're not. Next semester, I'm going to take a course in automated deduction, but I've have some (sad) attempts at creating mathematical theorem provers in the past. It's not too hard to build them, but searching for new knowledge is very tough because it's an exponentially large search space! Human mathematicians currently have an advantage because they have the intuition, but many of the lastest breakthroughs have been machine-assisted.
While randomly looking for new mathematical formulas is expontially frustration, we can get these automated mathematicians to work better by applying machine learning and data mining techniques to their search algorithms. Now we have a virtuous cycle where better automated deduction leads to better machine learning, and better machine learning also leads to better automated deduction.
When this cycle first gets started (and it is already underway, but only partially automated) it will seem slow, but at some point it speed up and get automated to such a high degree that I think a machine super-intelligence would develop much more rapidly than people think.