Artificial Intelligence, Really, Is Pseudo-Intelligence | Hacker News Reader

Artificial Intelligence, Really, Is Pseudo-Intelligence

npr.org

43 pointskevbin51 comments

To say something is pseudo-intelligence or true intelligence, one has to define intelligence rigorously. That is more philosophical than people think, because how little we scientifically understand about intelligence, or consciousness, or even life.

For most of people, the intuition of intelligence is a bit naive, and nowadays intuition is often at the opposite of the truth. It often is along the line of "looking intelligent and knowing what it is doing”. Something like turing test describes the first part intuitively. But we soon found out that we wouldn’t call something “intelligent” even it passes turing test since the machine is unlikely to know why it is saying something, hence the latter part. But again, this is about consciousness or “strong AI”, which is even further away from our understanding than “intelligence”. How do we know that the machines that “look intelligent” don’t “know what they are doing”? Consciousness is very hard to define and even harder to inspect, and maybe someday machines begin to have consciousness even before we realize it? In fact, we already did this to most animals, thinking they are not intelligent or self-conscious while they are.

But to me, the psychology behind these definitions is very interesting and maybe quite important. No matter how we try to define intelligence, we wouldn’t call machines intelligent until they very much behave like us. Because to most people, intelligence is what defines humans, so they intuitively use their feelings of “does that feel human” to benchmark intelligence. And psychologically, we are exceptionally good at telling something “not human” from “feeling/looking/speaking like human”. But this is a very narrow and, if I dare to say, ignorant and self-centric view, because it implicitly relies on the thinking of “we are far more intelligent than any other life form, and human's intelligence is the only way intelligence can be work”.

Think about it, if there’s a super-intelligent being that observes our world, maybe to it, human’s intelligence is as primitive as fish, and trees aren’t far behind; and maybe, entities such like countries or the internet are actually of more intelligent “life” forms, just we haven’t realized that yet because they don’t look like us at all; and maybe, on the scale of intelligence from 0 to that super-intelligent, humans, birds and machines are clustered in a small range, though significantly larger than zero, still not far from it.

Although the research on human brains and a lot of our own stuff inspired the research of AI quite a lot, it is perfectly possible that we create something truly intelligent or conscious without us understanding human intelligence or consciousness fully. We can have a different kind of “intelligence”, and our researches are well on their way toward that. It’s very hard to see because we don’t understand intelligence, and maybe our hardware isn’t powerful enough to fully utilize our theories.

But this is a very narrow and, if I dare to say, ignorant and self-centric view, because it implicitly relies on the thinking of “we are far more intelligent than any other life form, and human's intelligence is the only way intelligence can be work”.

I would rephrase that last bit as "human intelligence is the only way intelligence has been proven to work." We just don't and can't know which bits of human intelligence are necessary and which are arbitrary. For example, an emotional framework, the aspect of having drives and "desires" that lead to "illogical" and "irrational" behavior and interactions with others - is that required or not? How can we begin to reason about a non-human-centric concept of intelligence? Given how little we understand, I think that is too tall a task at this point.

and maybe, entities such like countries or the internet are actually of more intelligent “life” forms, just we haven’t realized that yet because they don’t look like us at all

I do believe that the collective aspect of intelligence has been wrongly ignored in AI. We try to build these monolithic beasts that take in mounds of information and try to make sense of it all themselves. But (yes, relying again upon the only known implementation) a single human on its own will never achieve the modern sense of intelligent behavior. It needs a society of humans, the benefit of a shared amorphous corpus of understanding that exists outside of any individual. And it requires many independent individuals with their own perspectives and ideas to probe and extend the boundaries of that understanding.

While I agree with your point that intelligence must be defined rigorously. We don't need to leave it to the philosophers. After categorizing the types of definitions, Shane Legg successfully used "Property of an agent that interacts with its environment so as to successfully achieve goals across a wide range of environments"[1] for his work in AI.

What Alva Noë is hinting at is the flaw with weak or pseudo-intelligence, as Alva calls it. As impressive as deep blue is with chess, or Watson with Jeopardy, they fail the "goals across a wide range of environments" test. This is clear with deep blue, and Watson once you realize it's just a glorified search engine.

I often hear people claim Watson learns because it improves based on the answer in the category. While it does use previous answers to tweak the final ranking algorithm. All of this is lost once the category is finished. To really "teach" Watson, you need to add more data to its index and re-index offline.

We have a workable definition of intelligence. What we're missing is a fundamental primitive unit of strong AI. Until this is found, AI will always remain just a bag of tricks, only capable of solving problems its programmers planner for.

[1] https://www.youtube.com/watch?v=V6umr1OP8uo

There's still a huge dearth of rigor in defining both the width of a range of environments, the degree of success in achieving goals, and the significance of various goals. I'll expound briefly on the first. Who is to say that the range of "environments" a chess supercomputer faces is narrower than the environments a human faces? Obviously, we intuit our range of environments to be wider, but can we explain why rigorously?

It's more than intuition. Both chess and jeopardy are subsets of the human environment. Watson couldn't handle tic-tac-toe, never-mind chess. Change one rule of chess, and human could cope but Deep Blue would need to be rewritten.

> How do we know that the machines that “look intelligent” don’t “know what they are doing”?

And, equivalently but perhaps more philosophically evocative, how does one know that other humans "know what they are doing"?

The authors comparison of intelligence to biological systems is absurd. Biological systems are more complicated than necessary. Let me say that again. Biological systems are overly complicated. This is because Natural selection has no regard for simplicity. Evolution is a blind process of trial and error that sticks to what works. So just because the gene networks of an amoeba are a confusing mess, which they are, does not imply that Intelligent machines have to be just as complicated.

I wrote a post on this awhile ago: https://docs.google.com/document/d/1wlFJtuBgnVkWm1nflKtJAy3b...

Natural selection does have a regard for simplicity, in that complexity has costs. Natural selection has very primitive hueuristics for achieving design: variance + selection. But once a superior arrangement is achieved, it succeeds. Presuming it provides an advantage in survival, reproduction, competition, etc.

Examples of retrogressive complexity among natural systems include the many birds and insects which evolve toward flightless variants in island habitats. Flying is a useful, but complex, and very high-energy adaptation. Where it provides a capability to out-compete other life forms (better range, escape from predators, more food-gathering capability) it's useful, but given risks presented (including being blown off an island in a storm, say), it can also be maladaptive.

This argument would have merit if we were anywhere near biological intelligence in our "AI". We don't even understand how most of the brain works, what it does or how it does it. In fact if it was so simple even to re-create some aspect of the brain, we would have created the super-brain with the amount of resources we have at our disposal and time we have spent in the field of AI. It is like saying calculus is unnecessarily complex because all I can do is simple arithmetic.

Look, if Evolution can give rise to intelligence, which is nothing more than a drawn out process of trial and error, then perhaps creating intelligence is easy. The major roadblock standing in our way is a lack of computing power. The Human mind still vastly outstrips our fastest computers. Of course, there will be technical challenges in creating machines with human level intelligence, but once we have super fast computers I think it will be possible.

You assume a-priori how the brain was formed. Perhaps the burden of proof could be in the reverse direction. Ever considered that ? If it was just a drawn out process of trial and error, surely hundreds of top AI researchers with 10,000s of cores/petabytes of RAM at their disposal could have would have made massive strides with their non hit and trial, deliberate process. Surely copying something that is in front of you is easier than designing something like it from scratch ? I hope you agree.

But we haven't been able to. And I don't imply that we will never be able to, at least hopefully copy or even enhance some of the interesting aspects. I find AI very exciting. But I would stop calling something we don't even understand "simple and generated from hit and trial".

Actually, we have made dramatic strides in AI. 2 years ago we couldn't classify multiple objects in a photo properly. Now we can, and within 2 years I would bet we will be approaching human performance on that task.

Watson can do medical diagnosis better than most humans.

We understand some parts of intelligence quite well, and others less so.

Yes, you're right. Relatively speaking, we have come quite far from 10 years ago. Using Google Now, searching for objects on Google+, doing facial recognition better than the average human, winning Jeopardy are all amazing achievements.

But as you have noted, we have been able to pin down some of the more methodical parts. I was mainly focusing on the fact that what the brain does is still amazing, and if we were to list all the different forms of intelligence it exhibits, we probably have a very long way to go. In no way is it simple and neither do we understand how it can do cognition, connect semantic concepts and reason about them so well as it does. Topics such as creativity, intuition, thinking, feeling that we witness human subjects exhibit are still fringe topics with much debate. So the original assertion that the brain is unnecessarily complex in its design and was simply a process of trial and error may be unfounded, until we can demonstrate better or even point to/simulate/emulate some mechanism that would give rise to a human's intelligence by hit and trial.

I understand your point, and do think is valid to some extent, but keep in mind hundreds of AI experts and thousands of cores for half a century still fall several orders of magnitude short of an entire universe running natural selection for 14 billion years. it could really be simple, given an absurd amount of resources... our replication process is actually much harder because we're trying to do it in a fraction of the time and "cost".

Edit: Look I didn't want to challenge your optimism. It is refreshing to believe that. But I have dealt with enough AI to know that most people have no idea what AI is. There is not much collaboration happening between the neuroscientists, philosophers and AI researchers. Unless a single organization brings them together under one roof, we are very far away from what most people would call biological intelligence.

Why is the brain the benchmark for "intelligence"? We already have machines that can run circles around the brain on very complex reasoning tasks. True, we don't have a talking robot that simulates humans, but in many other tasks we've met or exceeded what humans can do.

I would argue that we already have intelligent machines. What we don't have is a human simulator.

Haven't you heard? At the moment a computer exists which can outperform a human mind at a certain task, that task can no longer be considered a criterion for intelligence. Multiplying large numbers doesn't require intelligence. Chess doesn't require intelligence. Facial recognition doesn't require intelligence. Proving mathematical theorems doesn't require intelligence.

There's a widely-accepted standard for judging intelligence called the Turing Test. Computers are still a long way from passing that test.

They aren't that far from passing it actually.

One kinda sorta passed it earlier this year, by pretending to be a 13 yo boy. It was at the annual Turing Test event though[1].

Even if we ignore that, given the recent interest in using RNNs to generate natural language I'd say we are 18 months from being able to pass it reliably.

[1]http://www.reading.ac.uk/news-and-events/releases/PR583836.a...

I strongly doubt this. Here's a good way to defeat any computer in a Turing test: tell a joke, and ask it to explain why it is funny.

http://en.m.wikipedia.org/wiki/Computational_humor#Joke_reco...

The Turing test is a method for judging whether a computer has intelligence comparable to that of a human, or what's often called "strong AI." Strong AI is obviously a special case of AI, but most AI research isn't even particularly interested in developing a strong AI.

Ah come on. AI algorithms

1) have passed the turing test, then, indeed, we make the turing test more difficult.

(I would argue that if you give an accurate definition of a Turing test, namely "convince, on occasion, a human that you're human", that Eliza passed it in the 1960s)

2) several industries depend on their programs passing some sort of turing test every single day (ie. outsmarting humans directly, or outsmarting "are you a human" tests, automated or no). Examples : HFT (and HFT-like industries, like loan appraisal, credit validation, insurance brokering, ...), Spamming, police chatbots ("Lolita", a US version and "Sweetie", a British equivalent), Anti-spam, anti-HFT companies (who, for example, execute share buybacks for large companies), important parts of police work, ...

3) Some things are so AI-driven that nobody in their right mind still does it manually, or at least it is globally agreed that doing things manually results in a large competitive disadvantage (e.g. shipping).

4) There are places where the replacement of humans with AI has enabled new modes of operation that can not exist without AI. UAV's being the big obvious example, but I would argue there's a better one. AI has enabled bigger ships. Most of the 50000 ton + container ships have a turn radius of 50+ kilometers. If you don't have clear vision 20 kilometers out, you cannot control these ships safely (just look for the yearly demonstration disaster of that fact in the Channel between Britain and France/Belgium). Then we started building 650000 ton ships. When it comes to planes, it's not just UAVs that cannot be controlled by humans, but several kinds of fundamentally unstable planes have started flying, the big example being flying wing aircraft. But there's plenty of others : UAVs, dynamic wing aircraft (like a bird or insect would be a good description), aircraft too light to have a human control them, ...

Let's be fair here : this war was lost, a few years ago. Nobody seems interested in giving an AI a datacenter's worth of resources just to have it figure out how to benefit itself/improve itself. If someone did this, for a few years an algorithm could learn to quite literally be a human.

You raise a valid point but for an orthogonal question. When I first saw a calculator, as a 5 year old, I thought it was the most amazing and intelligent thing ever.

The original point discussed above was about how unnecessarily complex the brain is if its function/intelligence is so simple. We may or may not want to treat it as a benchmark, but the fact remains nobody understands it or can mimic its function even after trying for a very long time. (though believe me you, that's what most of AI is trying to do, even if in a simplistic way)

It is not a benchmark but a target to be reached. Once we have models of human simulation, we can treat the human brain as a benchmark.

This is quibbling semantics. I bet there are software simulations for amoeba behavior indistinguishable from the real thing. It's placing an inordinate degree of importance on physical reality. Is a perfect simulation different from the real thing? That is a fundamental philosophical question that can't simply be assumed to be no.

The reason why they are indistinguishable stems from our ignorance to learn about a deeper abstraction. We cannot simulate further than what we already know.

What does that mean?

We can tell when our simulations behaviour doesn't match the goal. But when it does, then what?

You seem to imply there is some (actual, literal) magic we are missing.

Not entirely. I am simply stating that, our simulations are indistinguishable because, 'we' are simulating them. The simulations aren't going to get any better than what 'we' understand. That being said, it doesn't imply we have a perfect understanding of the Amoeba down to the last particle of it. (since we absolutely have no idea what it is).

our simulations are indistinguishable because, 'we' are simulating them.

When we build bad simulations, then we can distinguishable them fine. Then we understand more, and do better, and then they become indistinguishable.

The simulations aren't going to get any better than what 'we' understand.

That's not entirely true. We simulate vision systems using RNNs, and we don't entirely understand how they learn.

Yes, agree with your point on how they become indistinguishable. And, w.r.t RNNs, yes, probably we don't understand how they learn, but has that aided to process of how visual perception actually happens?

Lets see it this way, does the amoeba simulation ever begin to show signs of evolving? Even if given a millenia? If no, then it's not intelligence, call it video playback if you will.

The author is simply ignorant. This kind of evolution of simulated amoeba-like organisms was done over 20 years ago with Tierra: https://en.wikipedia.org/wiki/Tierra_(computer_simulation)

I believe this is the most advanced artificial lifeform that is intended to be a full simulation of the real thing :

http://www.openworm.org/

( https://www.youtube.com/watch?v=r2OBznK6HkI )

Yes, it evolves. Well, in theory : running time is nowhere near what would be necessary at the moment.

Note that non-evolving versions would be extremely valuable as well. Evolution happens, but individuals don't. So if you want, say, medical advances, evolution is probably one of the first things you'd throw out.

But if you are generally looking for artificial intelligence that evolves, there's tons and tons of examples. The problem here is that people aren't usually all that comfortable with real running evolution during "production".

People want standard, reliable solutions that perform every time. The way this is done is to "freeze" the algorithm's evolution before actually using it (or even, as Geoffrey Hinton pointed out, first train an algorithm on the actual problem, then train a second algorithm to simply exactly imitate the first's behavior, which has much better results. Or even train a second one using a non-evolving algorithm like decision trees, which I'm told is often done in finance).

This is a pretty short and simple article but I think it is interesting and has been striking at the core of controversy of what AI is.

Is something that behaves like a human really intelligent? Google Search answers a lot of questions almost like a smart well educated person. Is it an intelligent entity? A computer plays chess better than humans. Is it intelligent?

In this case I think the author wants the entity to behave like biological organisms to be considered intelligent. This includes autonomy and self-replication. I can see that. This sort of points to robotics as the field that will produce these entities. Or maybe computer viruses...?

This is also a classic debate between behaviorists and those like Chomsky who want to understand the underlying processes of human cognition and intelligence. I can see both sides and maybe I like Chomsky's more idealistic view a bit better...

Here are a few easy to read articles that talk about this:

http://www.theatlantic.com/technology/archive/2012/11/noam-c...

http://www.tor.com/blogs/2011/06/norvig-vs-chomsky-and-the-f...

Another interesting one (a bit longer) in the same vein by David Deutsch (yes that Deutsch, the "multiple Universes physics guy")

http://aeon.co/magazine/technology/david-deutsch-artificial-...

The whole article is based on a false premise - behaviour of unicellular organisms is now understood very well by biologists and it is clear they are nothing more than automatons, basically. A simple explanation of amoebas behaviour can be found here[1], for example:

When an amoeba encounters a chemical that attracts it, the chemoattractant binds to receptors on the amoeba's surface. The receptor then activates a protein called a G-protein, which activates a protein called PLC. The PLC modifies other molecules to create short-lived messengers that will activate other pathways necessary for movement by the cell.

The story is pretty similar for any other behaviour they exhibit, its just a simple chain of chemical reactions triggered by an impulse from an environment. In contrast to many more complicated organisms where we hypothetically know the same to be the case (like humans?), in unicellular organisms we actually traced many of such pathways.

People very emotionally cling to their beliefs about intelligence, and this blinds them to simple logic. This is the same thing at operation that caused the backslash against evolution theory. In the end I don't see why such determinism would sadden anyone, our everyday experience is what it is and its nature is not changed by learning that we react, in the end, somewhat automatically.

[1] http://www.ehow.com/info_8517923_ways-amoeba-sees-responds.h...

What an astonishingly vapid and frustrating article. We need to get past these semantic arguments about what constitutes "intelligence." The "do submarines swim" analogy is cliched, but perfectly appropriate. It doesn't matter how you define "swim." Submarines are better at certain tasks than human bodies on their own. Likewise, computer systems that we label "AI" systems are better at certain tasks than human bodies (brains) on their own.

The author seems to use "intelligence" and "autonomy" interchangeably, which is troubling, mostly because he makes no attempt at actually defining either. His claim that amoebas and plants "exhibit an intelligence, an autonomy, an originality, that far outstrips even the most powerful computers" is ludicrous under any remotely plausible definitions of the terms. If humans genetically engineer an amoeba or a plant to change its behavior, is the magical "autonomy" suddenly gone?

Also, his argument is ridiculous even as semantic arguments go, since "artificial" isn't supposed to mean "synthetic." The standard dictionary definitions are "made by humans" and "false or pretending."

I don't think the author has enough knowledge about what is actually going on in the field of AI to provide reasonable arguments, all she/he spoke about was Watson and Clocks, i think she/he must understand there is more to AI than what comes on the news.

I'm happy to see an article like this on HN. It is much closer to my thoughts on AI then the general optimistic inevitability of AI one usually sees in the tech community.

Ha! Artificial intelligence us always whatever we can't work out what to do.

As a general principle, if someone claims X is not AI, then they need to define a test they will accept as a true test of AI.

"The agency and awareness of an amoeba" is achievable now[1].

[1] http://www.smithsonianmag.com/smart-news/weve-put-worms-mind...

I think the author should be using the term "agency" instead of "intelligence", and then I might agree. The amoeba or the most basic animal doesn't have much intelligence, but it has drive and desire to survive and reproduce. That's not true in artificially intelligent beings. And while you could hard-code such drive (see the Belief-Desire-Intention frameworks in the multi-agent systems literature), ultimately it will still always be "pseudo" by definition. The notion of survival in a machine will only ever be a trick designed by a human to get a machine to do something for us, not something innate.

Maybe one day we'll implement some Adam and Eve robots with an initial hard-coded survival DNA, that the robots will abide by and transmit to their children robots. Then those children robots will evolve and one day think of building thinking machines (maybe out of some organic material) to work for them and will think of faking some survival code (maybe using some sort of protein?) in them to create pseudo-agency... ;)

I'm surprised whenever I see what seems to be belief in dualism in ostensibly serious scientific discussion about humans. This author goes so far as to attribute dualism to amoebas.

That's okay, natural intelligence is really only pseudo-intelligence, too.

This article makes a number of very bold claims about poorly defined concepts, and provides zero evidence to back anything it says up.

I agree that Watson is nothing like a true artificial intelligence. But I do not believe that this means we are unable to create such a thing.

I think true AI will be a highly-parallel learning machine, possibly modeled off of something like the human brain. But I do not think that we need to fully understand the human brain to create an AI smarter than us.

I think one of the big challenges may be in providing the proper sensors / inputs to this learning device.

I think once we have a true AI, we won't really understand how it works, just as we can train a neural network but can't look at the network itself and make much sense from its parts.

I think it's possible that the road to the singularity may be gradual, and we might not even realize when it happens.

Edit: Today's xkcd is somewhat relevant: http://xkcd.com/1450/

Man that comment thread takes me right back to Freshman philosophy class. I particularly appreciated the notion of an Amoeba having 'agency'. Caveat the mechanical reconstruction and processing of the organics of DNA, systems have been built and discussed in Artificial Life conferences that not only have the 'agency' of an Amoeba but emergent behaviors in flocking, social adaptation, and mutual defense. And what does that say about intelligence? Pretty much nothing, just like all mass being attracted to all other mass clearly demonstrates the concept of gravity but says nothing at all about what gravity is.

But it is a lot of fun to speculate about.

"But it's striking that even the simplest forms of life — the amoeba, for example — exhibit an intelligence, an autonomy, an originality, that far outstrips even the most powerful computers. A single cell has a life story; it turns the medium in which it finds itself into an environment and it organizes that environment into a place of value."

As someone who does work in this field, we're not that far away from a fully atomistic simulation of a single cell (with a couple assumptions about reactive force fields, quantum cut-off, numerical stability, time scales, yada yada...)

My problem with the article is, it never explains the difference between "processing" information and "having" information. And I think that the entire argument hangs on this distinction. The author essentially claims, that a amoeba has some essence of live which a computer can not posess, but fails to show that the distinction is important beyond semantics. ( To clarify, I think that this is an entirely defendable position, it is just one of the questions where details of definitions really matter.)

The brain is clearly marvelous, but it's not magic. It's a physical object with rules which govern it just like anything else. To think that we won't be able to crack the AI nut just smacks of insecurity on the part of the OP. We are, whether anyone wants to accept it or not, just meat robots.

I think a wide-spread problem in this discussion is that people misinterpret the word 'just'. It clearly means "just meat, no magic". But what people mistakenly understand is "just meat, inferior to magic".

I'm starting to believe (hard) AI should get the same status as politics and religion on HN. Those comment threads debating the true nature of AI are getting repetitive.

Reply on news.ycombinator.com