I’ve recently obsessed about my commuting options, trying to figure out the optimal route. It’s a little nuanced in that it matters where/how you park so you can get on the train, etc. (Could even get into reliability of train vs bus, traffic, frequency of trains, etc.) It’s the kind of exercise that I think resembles very basic level analytics work. I asked chatgpt for its thoughts on which commuting option from A to B was best and it was awful. Better off just doing directions on google maps and taking the first one.
I also learned if you ask it to show you a picture of someone writing with their left hand it shows you a picture of a right handed person and lies.
The problems you mention seem to be able to be “engineered” away. They are basically problems of overfitting. We can ease them by getting more data so that we have training and testing data, and using regularization or Bayesian methods to prevent over-fitting. Sometimes this will work, but not always.
For example it doesn’t cover the situation when you build a model which lets you understand the problem better, and realize you need fundamentally different kinds of data.
We can also consider Newtonian gravity over Ptolemy’s model. Both use the same training data, and both can be falsified. For example, the universality of Newton’s laws provide a different kind of preference, particularly when considered with the various observations of changes in the heavenly bodies (the supernova observed by Galileo for example).
I’m not sure about general everyday ‘equation solving’, but Google used a similar notion to improve some algorithms. See AlphaTensor and AlphaDev. The improvements are small, but the algorithms are used frequently. Obviously, AlphaFold was the bigger thing to come from DeepMind lately.
But yes, chat-bots can’t reason. They can only try to predict the words we use when we talk about reasoning. (I think humans often do this as well-- though presumably not all of the time.) They can also try to translate their words and ideas into code, and then run the code. But they aren’t great at it regardless.
The newest generation of models that do pretty good essentially work through trial and error. It’s not quite simple-brute-force, but it is a lot of force. One part of them is generating ideas. Another part of them is critiquing those ideas. The two together come pretty far in problem solving-- though it’s getting computationally ridiculous.
Open AI’s last paper had a model that used >$1,000 of compute per problem.
I can’t speak to genius. My limited experience with creativity is that there’s also a lot of trial and error. You experimentally combine ideas, images, interpretations, equations, definitions, theorems, whatever. Then you test them to see whether your new ideas are any “good”.
I think a chatbot might be able to do those things one day, but we’ll have to wait until they can at least solve already solved problems.
It’s surprising to me that Fortran was faster than C++ until the mid 2010s or so. Apparently this is because the more limited expressiveness of Fortran allowed for additional optimizations to matrix calculations. C++ caught up using special “template” techniques. The AlphaTensor reminds me of a fancier version of that. It is interesting to hear about it, I don’t think I knew about that particular case.
Mathematica can do more general symbolic calculations. I hope a deep NN speed up is coming soon (or maybe is in a latest version I haven’t seen). It would be more broadly applicable.
A deep problem is that we don’t really know what reason is, fully.
A popular AI textbook tries to pragmatically define intelligence as “doing the right thing”. Unfortunately, nobody knows what the right thing is. Mathematicians are going to tend to have their own idea, influenced by the values that helped them become successful in their field.
One thing is for sure- it isn’t modeled on how people actually think and make decisions.
The reason I bring it up is because we have statements like the one made by elon musk a few days ago (as I recall) that probably by next year, we will have a computer that is more intelligent than any person currently alive. I think it is fair to respond to claims like that with references to genius.
It makes me wonder what kind of “intelligence” Musk and his kind are really looking for. True genius tends to be terribly subversive of existing power structures: academic, political, religious, artistic, etc. Maybe it should be no surprise if so many of these AI execs envision perfect intelligence as a kind of perfect drone worker. Somebody who can take data and, without any motivation or values of its own, return a product that can be sold.
Gothamchess is covering a chess tournament with Stockfish, Martin and 6 AI engines. The first one was posted today featuring Stockfish and Snapchat. Snapchat performed many illegal moves, but Levy let them be played. Fun watch. https://www.youtube.com/watch?v=CZGs4g_hVco&ab_channel=GothamChess
It was funny to see Snapchat move the king onto the same square as the bishop, move the knight twice in one move, have the queen move as a knight, etc.
Like playing with a kid who knows grandpa won’t say anything. Then the kid throws the board into the air ands stomps off after losing.
Will this AI grow up to be a brat or will it mature? This AI might have digested too much Calvin and Hobbes. No not the philosophers.
