> AI companies - the most valuable companies in the future
To me that's like going to a state fair 150 years ago, seeing the vendors of snake oil, and concluding that the most valuable companies in the future will be selling bottles of secret potions.
For "artificial intelligence", I see no actual intelligence and a lot of emphasis on artificial. The best of it is at its core some fairly traditional applied math and/or electronic engineering.
Instead, there're huge sections of the research libraries for pure math, applied math, physical science, electronic engineering, and computer engineering, and in comparison so far the best of AI is hardly more than a few books on one shelf.
How to do applied science, engineering, etc. hasn't changed; AI hasn't changed how; the only way to have
> AI companies - the most valuable companies in the future
is just to call all the most valuable companies, or safer yet, all new companies, AI. Silly talk.
A claim like
> AI companies - the most valuable companies in the future
will be part of what we've seen before: AI winter, AI spring of new hope, AI summer of scam, AI fall of failure, AI winter again. Get your warm clothing while it's cheap now because the next AI winter will be darned cold!
E.g., look up the little game of, say, matchsticks called Nim. Let AI learn to play it. Against just a common man in the street, it might do well. Against me, it will lose at least half the games if we first flip a coin to determine who moves first. Just why and how I'm winning, the AI program won't figure out; figuring out such things is so far just not what AI does.
Or, teach AI some high school plane geometry. Then consider the problem, given triangle ABC, by Euclidean construction find point D on side AB and point E on side BC so that the lengths AD = DE = EC. To do this, need to be a little creative; that is, need some actual intelligence. The people who taught the AI geometry program likely didn't teach how to do that, and, then, for the solution, the AI program won't know how to do it, won't figure it out, and won't do it.
Net, AI is not "intelligent". Chase down big trees in chess, checkers, Go, etc., play games against itself, accumulate a lot of statistics, and, then, play and often win -- okay. Be "intelligent"? Nope, not yet.
Or, one day I got via e-mail a specification of a problem in 0-1 integer linear programming. A first example had 40,000 constraints and 600,000 variables. Some people had tried "simulated annealing", ran for days, took the best results they had so far, results of unknown distance from optimality, and stopped. Drop that problem into a lot of integer linear programming (ILP) packages, and won't get very far. Write some software to do what such ILP packages do and call it AI, and again won't get very far. So, instead, need some new ideas, that is, from some actual intelligence. So, I put my feet up, popped open a cold can of diet soda, reviewed a little, thought a little, did some derivations in non-linear duality theory (early in that theory there are some darned cute results, e.g., get some surprising, useful convexity), designed and wrote some code, and in 905 seconds on a slow PC found a feasible solution guaranteed to be within 0.025% of optimality. Figuring out how to do that for that problem took some actual intelligence.
Here's one: I was in an AI group monitoring server farms and networks. So, suppose we are given instances of one or both of those and collect data, say, points, where each point consists of numerical data on each of several variables, say, 50, and where we get, say, 100 such points a second.
Now we want to define and detect in essentially real time anomalies as, really, a reason to start diagnosis and see if there is a real problem. For this, we want, and are quite serious about, false alarm rate: We want the rate adjustable in advance and, then, in practice to get that rate essentially exactly.
And when we get an anomaly, we want the lowest rate of false alarms at which the data would still be an anomaly.
Okay, AI/ML machine, how the heck to do that?
The answer needs some intelligence, e.g., borrow some from ergodic theory, get a finite algebraic group, sum over the group, etc. For that work, the current techniques of AI/ML won't even pass the giggle test.
Sure, taking my work, could program that and call the result "artificial intelligence", but really just programmed some pure/applied math. And, sure, will need an algorithm to make the computations fast, and that will need more genuine intelligence.
AI "intelligent"? LOL.
> AI companies - the most valuable companies in the future
More LOL.