Q and A: The future of artificial intelligence | Hacker News Reader

Q and A: The future of artificial intelligence

people.eecs.berkeley.edu

149 pointsaduffy30 comments

Because it's not in the title nor anywhere on the page, I'll point out that this appears to be written by Stuart Russell, co-author of Artificial Intelligence: A Modern Approach (http://aima.cs.berkeley.edu/), the standard intro to AI textbook.

That's a good point, I should've included that in the post

I have a loathing for claims put forward in the "Common Misconception: XXXX" format.

For example: "Common Misconception: ...AI will necessarily increase inequality."

The question of whether AI will increase inequality involve long arguments that hinge on both unknowns about society and unknowns about AI.

Sure, it's not absolutely certain that AI will increase inequality but that's an empty assertion and the whole thing seems like an expert brushing off serious concerns about technological development impacting society.

Of course AI will increase inequality... AI will be baked into more expensive products which the poor cannot afford.

Maybe. The question is which gets eaten first - capitalism, AI, or humanity in general?

Just think of all of those smart phones the poor can't afford.

Oh sure, yes AI will also be baked into products sold to poor, no it won't be for their benefit but rather to control, manipulate and rip them off. Kind of like advertising and surveillance now.

i dont really see it that way. i see it as more of a we need to include all the facts and acknowledge the complexity of this 'thing' if we want to have any shot at understanding the future world state.

    > > When will AI systems become more intelligent than people?
    > [...]  
    > Achieving [general-purpose AI with greater ability than humans] would
    > require significant breakthroughs in AI research and those are
    > very hard to predict.
    > Most AI researchers think it might happen in this century.

Is the author projecting here, or is that actually true?

It's very true that it's hard to predict. Whether most researchers think it "might happen in this century" can only be answered with a comprehensive survey.

Some attempts and analyses of such surveys: http://aiimpacts.org/ai-timeline-surveys/ and http://www.nickbostrom.com/papers/survey.pdf

It's speculation, really. The point is, breakthroughs or showstopping difficulties are hard to predict.

It's not true, and it's basically sensationalist nonsense. The scope of human intelligence hasn't even been mapped yet and we can't even accurately formalize things like natural language.

Saying AI has a long way to go is like saying an electron is "pretty small".

What part of that quote, exactly, is "sensationalist nonsense"? The first part that says that more breakthroughs are needed, or the second part that states the testable fact that most AI researchers give a high probability that we will achieve human level AI this century?

This popped out of nowhere. Because I am in serious dilemma if I should focus for a few years on Ai(ML||DL||NN). I got place on very respected computer club on my faculty. I got to choose in which team would I like to go. At first i wanted to go to OS/Systems programming. But recently I was reading a lot about Ai in terms pf CS. And pne of the board members of the computer club asked me why I do not think about joining Ai team, since I was very well covered on all basic bases of programming (OOP, algorithms and a little bit of functional in Lisp). Since then I am bashing my head really. Ai looks really interesting to me, but I am afraid of it's high level abstraction, and it looks to me that it is just piping data for the most people (those people who make actual algorithms are not included here). So i think my fear comes out of unknown. Ai is the future and it looks very interesting and attractive to me, plus I could go a learn awesome new things in terms of different, more exotic languages that could do me well in general as software engineer, but man I don't know yet...

Something to keep in mind if you're looking into studying AI seriously is that it's really half a dozen barely-related fields, that have all retained the group name "AI" solely because it sounds cool. Machine learning specialists use a whole different set of ideas than robotic control specialists, for instance. The logical inference engines or expert systems that used to be the bread-and-butter of AI departments are now flippantly referred to as "GOFAI" ("Good Old-Fashioned AI") to distinguish them from the shiny new approaches that get students hired into Google.

I think one could frame the question like this: What do you want to learn; logic, statistics, and/or cognitive science? For each there is a branch (or several branches) of "AI" that revolve around it, and have only a limited amount to do with the others.

I am also struggling with deciding on what field to dive into. I really don't want to ask but can you give me some insights based on what I am looking for?

What do you want to build? Do you have a system you're excited about? For instance, a virtual assistant/personality? A robot that can move around and learn like an animal? Something else? Answering this will definitely help identify the proper field(s) to look at.

I saw an interesting TEDx talk on this topic a while ago by an AI researcher from the Netherlands [1]. The talk is titled "Why are computers not intelligent (yet)?" by Prof. Niels Taatgen.

[1] https://www.youtube.com/watch?v=n1ARQsdNvmI

There are 2 irksome AI memes in particular that I've seen repeatedly in HN comments. Even sillier, they contradict one another. The first is something like "That's not AI, it's 'just' ML." The second is almost the opposite (but just as misguided) and is something like "If you're not an ML expert talking about ML, then you're not talking about 'real' AI."

This Q&A counters both, and I found it refreshing.

>Common Misconceptions: Neural networks work like brains. In fact, real neurons are much more complex than the simple units used in artificial neural networks; there are many different types of neurons; real neural connectivity can change over time; the brain includes other mechanisms, besides communication among neurons, that affect behavior; and so on.

I really have to disagree with this statement because it's actually broad and misleading. Yes, real neurons are more complex, but that doesn't mean that artificial neural networks are not capturing the important functional/computational properties or real neurons. In fact, you might even say that real neural networks are approximating the computational properties of artificial neural networks via a somewhat Rube Goldberg like process. If you actually look at the mathematics that functionally describe artificial neural networks it's the same math that describes gene regulatory networks. This is not a coincidence. This mathematical abstraction like a platonic computational system. In effect, evolution converged on this computing paradigm twice using the processes available at its disposal. The first was a chemical reaction network, the second was neurons. Moreover, if you actually study the computational properties of these networks you discover that most of the implementation details are actually irrelevant to the functional output and it's the topology of the network that's primarily driving the function. Not unlike how an engineer might recognize the A 8-bit added from the circuit-logic diagram. The problem that people have with artificial neural networks is that they're not spending enough time reverse engineering the salient properties of biological systems and instead they're trying to brute force AI with their mathematical brilliance. A lot of these questions disappear if you study the network architecture.

> In fact, you might even say that real neural networks are approximating the computational properties of artificial neural networks

There's little to no evidence for this currently. To my knowledge, there's nothing to indicate that the brain uses anything like backpropagation, which is the mechanism by which artificial NNs learn. That's such a fundamental operation for ANNs that I think it would be premature to claim anything like the above.

Back prolongation is a training algorithm for ANNs, not the circuitry that performs the function.

> When will AI systems become more intelligent than people?

Human intelligence is a sort of 'field awareness' as described by Alan Watts. Think more waveforms than digital bits. I think strong AI will happen when the intelligence is given enough free will. We might one day have to throw out the rulebook regarding AI harming us.

I think we lack a real metaphysics yet for what Strong AI is allowed to do. Aleister Crowley's[1] 'do what thou wilt' is the only well thought out metaphysics for free will we have now, and something we could teach the machines.

Putting them in sandboxes is a nonsense as it has been demonstrated even in modern computing that breaking out of sandboxes can be done. Even the most hardened air gapped VMs can be bridged to the public Internet, more often than not, by accident.

[1]: https://en.wikipedia.org/wiki/Aleister_Crowley#Beliefs_and_t...

The article says: "However, the kinds of tasks addressed by AI systems tend to differ significantly from traditional algorithmic tasks such as sorting lists of numbers or calculating square roots."

Yes, those tasks do significantly differ from the mathematical calculation tasks that algorithms are traditionally associated with BUT the whole idea of AI is to take those tasks and represent them as algorithmic tasks so that they can be solved using computers. No matter which human cognition concept you associate to a computer, the reality is that the only thing that a computer can do is run the given algorithm. Somehow most of the AI community is just fascinated about giving computers the ability to "think" or have "perception" etc without realising that what they are actually doing is "representation as calculations".

I like this one

> Humans are generally intelligent. This claim is often considered so obvious as to be hardly worth stating explicitly; but it underlies nearly all discussions of AGI.

I wonder though. Should that really be the goal? This seems like another instance of people's propensity to think of themselves as the pinnacle of something. In this case that something is intelligence.

I think this frames it in the economic terms. The unstated goal is to replace human beings with robots, to get automation to do everything people can do. Which would then render people economically useless. Which would then cause some sort of drastic change in the purpose of human life, which since the 1600 or so has been about using people to produce things that they then consume.

I see it more like us working on general intelligence instead of specific solutions. This however would require a benchmark to compare to, and Human intelligence is a reasonable one.

I believe it is generally accepted that artificial general intelligence is considered unsolvable. It would be nice to hear that addressed more explicitly.

Where did you get that notion? At least among people working in the tech sector, it is widely believed that we will achieve AGI one day, and that too with the the help of Turing machines. Except for Roger Penrose's argument against this, I haven't seen any opposition to this belief yet.

Mostly from conversations in CS grad school ~2012 with research professors doing related work, such as in ANN. At that time at least it was commonly discussed in academia. Perhaps a breakthrough has happened since but not one that I'm aware of.

For example, see the sections "Tests for confirming operational AGI" and "Feasibility" in [1], the $100,000 prize requirements in the Loebner prize [2], or "The integration bottleneck" in The real reasons we don’t have AGI yet [3].

[1]: https://en.wikipedia.org/wiki/Artificial_general_intelligenc...

[2]: https://en.wikipedia.org/wiki/Loebner_Prize

[3]: http://www.kurzweilai.net/the-real-reasons-we-dont-have-agi-...

Nobody has told Geoff Hinton or Rich Sutton.

Reply on news.ycombinator.com