Slashdot Mirror
'Modern AI is Good at a Few Things But Bad at Everything Else' (wired.com)
Jason Pontin, writing for Wired: Sundar Pichai, the chief executive of Google, has said that AI "is more profound than ... electricity or fire." Andrew Ng, who founded Google Brain and now invests in AI startups, wrote that "If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future." Their enthusiasm is pardonable.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
...before a bunch of angry old coots post telling us that none of this is AI.
We don't have AI, in any form, in the modern world. We have code which solves program similar to a neural network and we have code which can mutate within very strict limits with genetic algorithms. We have nothing even approaching "artificial intelligence," which at the very minimum of the bar would be the level of an "intelligent" Human. If it's not better than a Human with an IQ of no less than 135 at literally everything it's not AI. We have nothing remotely close to equal to an actually retarded Human with an IQ of 70.
A "real" AI would be the ultimate psychopath: Intelligence without any kind of conscience. Pretty much like a corporation, just way more efficient.
Fortunately what we're building is far from anything resembling intelligence. I.e. the ability to use prior experience in totally new situations, evaluate those situations and draw conclusions that can be applied to react properly to it. And I mean totally new.
The point here is that it's not possible (yet, maybe forever) to create an AI that can make such abstractions and apply old knowledge to new situations.
We used to have a Bill of Rights. Now, with the rights gone, all we have left is the bill.
We saw roughly how heavier-than-air flight would work, but we didn't have the pieces to put it together. We understood the airborne part enough to carry humans dating back at *least* to the sixth century (earliest recorded 'paragliding'). We couldn't make a practical aircraft, but we could see how the pieces would play a role in such a marvel if we solved other pieces.
Here, the current 'AI' craze doesn't even in theory extrapolate to higher-order displays of intelligence. It is a highly practical field to advance and is certainly useful, but *if* we want to go to more 'intelligent' systems, it's going to be based on a different methodology, or at least no one who understands the field can see a hypothetical extrapolation of this approach that leads to those results.
The problem people have is that a useful, albeit narrow discipline is conflated with the entirety of human intelligence. I have seen many in the field understandably trying to discourage the phrase 'AI' to head off very annoying irrelevant conversations and concerns.
XML is like violence. If it doesn't solve the problem, use more.
Not true at all unless you are narrowing the definition of AI to such a narrow degree as to make it effectively meaningless.
Use any definition other than "pattern recognition" and we don't have it.
Nonsense. Dogs do not as a general proposition approach human level intelligence. Yet do have real and measurable intelligence. A computer with the intelligence of a dog could very fairly be described as intelligent. AI does not have to pass human intellect be classified as intelligence or to be useful.
We don't even have computers with the intelligence of a rat, in fact the closest we've come to any animal-level AI is to simulate the scans of 1/100th of part of a mouse's brain for a few milliseconds worth of time. We are nowhere near anything that could be called AI, semantic networks and other forms of pattern recognition are tools, genetic algorithms are tools, neither are close to AI.
+1. This is algorithms and infant ML.
I can take my kid and train him to swim and then train him to drive a car and get rudimentary skill in a week in both.
You can only do this after about six years of full-time learning in how to navigate in the real world and how to operate his body. This is the hard part, the part that humans learn in their first six years and AIs don't: dealing with the external world.
Learning to swim and learning to drive a car are easy; machines can do that. Learning to make a peanut-butter-and-jelly sandwich out of what is in the refrigerator: now that's hard.
Computers don't understand 2+2. They perform the operation by moving electrons from one place to another, ending in a pattern that humans interpret as 4.
Our brains have many different parts each with thier own function(s). It's pretty apparent humans have many simple algorithms running simultaneously, in addition to whatever else happens. So the obvious conclusion that will surprise no one is that a true AI like a human would just have deep learning (or a number of deep learning modules) as a single component among thousands that would be required to get the emergent behavior that is strong AI. The technique is simple, easy to implement, and accomplishes certain simple tasks with ease so it's usefulness as a tool when designing strong AI won't ever go away. There is no magic in our molecules, there is nothing magic about strong AI, it should be obvious a single algorithm won't ever make a strong AI, we and other animals prove it can be done it's just difficult. People need to get over it and realize that when millions of people work together, contributing our results and pooling our knowledge, we can make much of science fiction a reality (hopefully just the better parts).