Why general artificial intelligence will not be realized (2020)
nature.com
nature.com
Businesses and institutions don't want to compete, because competition is hard and they might lose. So we get shit like this, and any good work that makes it through is a lucky side effect. Doing a good job is often not a competitive edge.
In just about any field or area, if someone is trying to tell you it's meritocracy, they're just pulling the wool over your (or their own) eyes.
The estabilishment/incumbent does not like competetion in any area. And the default mantra is claiming it's actually a meritocracy. Which usually translates to "we like the existing structure because it keeps us and our friends in power".
Now we're receding back to something more resembling the mid-late 1800s, where aristocracy was a big deal and you had situations like the Imperial British military, where the enlisted soldiers were impeccable, but the blue-blood officers did things like carry pianos into combat zones, and caring about things like self-improvement was a sign of weakness socially.
No. Intelligence is not «identical to human intelligence». This already reveals a very sloppy approach.
> cannot be realized because computers are not in the world
This just looks like a bad joke.
Reading the abstract made me very annoyed, there’s like 4 separate completely false claims.
If that's the case, can we just flag it for bullshit?
We don't do that for computers. Part of it is cost. Imagine attaching millions of thermometers, barometers, wind gauges, etc all over a computer/robot. They also have very limited interaction with the physical world. Also partly by design.
They exist solely as the hypothetical "brain in a jar".
Now, whether or not you need that to have human-level intelligence/consciousness/sentience/whatever, that's a different story. But that's the gist I get. "The constant feed of stimulus is somehow integral to achieving intelligence."
It seems to me that would be rather easy these days. Each of those devices can be online. Each can publish its measurements onto a message queue. Any "intelligence" can subscribe to that and make of it what it will.
Basically the internet can be the extended nervous system of a computer-based intelligence.
https://news.ycombinator.com/item?id=32310799
The idea of this project is to build an "embodied" AGI that has some ability to experience the physical world in a way that's at least analogous to the way a human does. So far I have been working on a lot of low level "plumbing", but I just finished getting the cellular modem / GPS adapter working, so now the bot can "sense" its position on the surface of the earth, and its velocity (if moving). Next steps include a 9 DOF accelerometer board for more localized sensing of changes of orientation in space, a camera, a microphone, and "other" sensors - possibly to include the aforementioned barometric pressure sensor, temperature sensor, etc.
Will this lead anywhere? I have no idea. But it seems like a possibly fertile avenue to explore.
By "computers are not in the world" he means that a computer is based upon an abstract interpretation of the world, not the world itself.
That's the difference between a lemon and a story about a lemon.
That's a UNIVERSE of difference.
We do not get directly to the lemon and we work on stories about the lemon.
And - again -, the idea was to demonstrate that something was impossible. Not to contemplate ideas like in a bad pulp fiction about the Sixties - less productively than Java without garbage collection, because in Java, that which is now garbage has formerly been actually used.
First of all, what makes you think that this was not clear.
Secondly - why, the human mind does not work through representations? Kant and the noumenon came here to be ignored? Schopenhauer and the Vorstellung? Is it not evident anyway?
Third: that idea, all those there written ideas, should be there to /entail/ that "something is impossible". That is their task. I would like to see those Key Performance Indicators met.
2) you say it's representations all the way down? Well it doesn't look that way to me. Or the author apparently. So that's two of us.
3) go with extremely unlikely then. Don't trip over it.
2) And your opinion should be postulate in discussion? Especially when it was reminded that the representational nature of the mind and knowledge was recognized since the dawn of thought then called a main problem (Kant) and the starting point (Schopenhauer) at relative maturity? You think that just saying that "you see otherwise" suffices?
3) And the critical difference should be one of "degrees"?! It has to be _proven_, whether it is 'impossible' or 'almost so'. And if that rhetoric that sounds like sophism just meant "'difficult' in front of obstacles", well that should be more obvious to the architect than to the commenting passer-by looking from the border of the construction site - it is a very gratuitous note which wastes everyone's time; and it if it meant so, it should have expressed so instead of literally different statements (and morally "bait").
I know I've beat this dead horse numerous times, but I do think it's critical to make the distinction between "human level" intelligence and "human like" intelligence. And I think we can have intelligence that is "human level", even in its generality, without necessarily being exactly "human like". Or perhaps without even particularly "human like" at all, much less exactly so.
In what ways do you think it would be human-level, but without sharing characteristics ("human-like")?
Our reasoning is guided (for better and worse) by a constant stream of subconscious biases, rising out of our desires and fears. No one is immune. Note that this is not necessarily a bad thing as in many (most?) cases, it allows us to short-circuit much mental effort. Nevertheless, those same biases can mislead us too.
A human-like artificial intelligence would need a sense of value, or at least self-worth and aspiration, to mimic those biases to be fully human-like.
Agreed. That said, I think the degree to which a given AI is "human like" is a value drawn from a continuum, not a binary "yes/no" dichotomy. Some AI's could be more or less "human like" than others, depending on how they are built / trained / etc. At least that's my position at the moment. I freely admit I can't prove it currently, and I'm not totally convinced anyone ever will. But maybe.
Anyway, if you take that as a given for the sake of argument, I consider an interesting question to be "just how 'human like' do we actually want/need our AI's to be?"
1. We already have AI that is "human level" or even better than human level in the "narrow AI" sense. This part is trivial to see. AI's can beat humans at chess, dividing large numbers, and any number of tasks. But this is "narrow" and by and large the program that can play chess can't also drive a car or anything else.
2. With that established, if we took AI's from (1) above and did nothing but make them more generally applicable, so that the same program can play chess, drive a car, or calculate complex integrals, you'd arguably have something that could at least come close to being called "human level artificial general intelligence".
3. But nothing about establishing (2) above entails, so far as I can see, the necessity of the AGI having human like behaviors, emotions, desires, intentions, attitudes, etc. The easiest way to explain what I mean here is to think in terms of embodiment. It seems clear to me that some portions of our intelligence and behavior are tied up in the way we experience the world through our bodies. Even down to the analogies we construct and use and how we understand them. For example, we talk about "the stock market falling" or a person "falling down" (both in the literal sense or like the movie of the same name), or we say the level of a lake "fell", and I'd argue that some of the understanding of the abstract notion of "fallingness" is tied up in our visceral experience of falling. We know what it means to "fall" in a certain sense, because we have all fallen. We've experienced stumbling, being off balance, beginning to tip over, seeing the floor rush towards our noses, extending our hands to break our fall, getting carpet burn on our hands as a result, etc. And I posit that we can't really divorce those experiences from our subsequent learning and behaviors. And a (2) AGI, unless specifically built to be embodied, and to have that ability to experience the physical world, probably won't have the same attitudes, beliefs, behaviors, etc.
That's not to say that you can't teach an AGI some notion of "fallingness" without it being embodied. My argument is simply that there is a distinction between "human level" and "human like" and that that distinction is largely related to our physical manifestation via our bodies.
No formal proofs of any of this, but this general view informs my overall position towards AI / AGI in many ways. Others may or may not find it useful.
Do you also think that some goal is necessary for an intelligence to make sense? If so, what do you think that goal of an AGI would be?
For "intelligence" in the abstract? No. But going back to the "human like" distinction... humans generally seem to have goals, so arguably an AGI that would register high on the "human like index" would have goals as well. But would an AGI necessarily have the same kinds of goals as a human? I can see plenty of cases where the answer would be "no".
Take hunger for example... humans often have a goal to "find food". An AGI running on a rack of servers in a data-center doesn't exactly have a perfect analogue for that. I guess you could set up some kind of metaphor based on maintaining a quality electrical feed, but those two things strike me as fairly different.
For instance, would its prime goal be to continue up leveling its intelligence?
My only contribution to that question at the moment would be to say that I think it might be over-simplifying to think in terms of the AGI having just one goal, or even having one clear "prime" goal. If there's no constraint that it be "human like" then I suppose its goal(s) can be whatever its designer programs in.
In the research I'm doing (which is very incipient and not particularly advanced at the moment) I'm exploring the idea of having the "AI" have a variety of internal metrics (I suppose you could consider them loosely analogous to human emotions) such as "boredom", "curiosity", "confusion", and so on. And the core processing loop may be driven by the "AI" trying to balance those various metrics. For example, if "boredom" is high, it may seek to engage someone in conversation, or go off and start reading Wikipedia pages.
If "confusion" is high it might start running internal simulations or using abductive inference to try and generate explanations for things it doesn't currently understand. Or if there is a current interlocutor engaged in conversation with it, maybe it starts asking "who, what, when, where, how, why" type questions.
A big hole in this scheme is that I don't currently know, or claim to know, exactly how to train it to use interrogatives to ask questions. I'm planning to start a deep dive into the current BOK on human childhood language acquisition to see if there is something to be borrowed from that world, but that's pure speculation on my part right now.
"On the Impossibility of Supersized Machines
In recent years, a number of prominent computer scientists, along with academics in fields such as philosophy and physics, have lent credence to the notion that machines may one day become as large as humans. Many have further argued that machines could even come to exceed human size by a significant margin. However, there are at least seven distinct arguments that preclude this outcome. We show that it is not only implausible that machines will ever exceed human size, but in fact impossible."
This is in fact more than likely. Depending on how one defines “general inteligence”.
There are problems which require superhuman reactions. (Think of plotting course to intercept a missile barage with multiple interceptors.) There are problems where our brains can’t fit the variables required. (Imagine a 5 dimensional Rubik’s cube with a billion segments along each edge for example.) There are very specific circumstances under which our inteligence can even function. (Imagine as simple a challenge as tying your shoelaces while exposed to the vacuum of space, or a room temp atmosphere of pure chlorine gas, or enchased in molten metal.)
Heck these supposed general inteligences can’t even reason themselves out from their own climate change experiment, or prevent viruses from harming their bodies.
> One of the leading critics was the philosopher Hubert Dreyfus, who argued that computers, who have no body, no childhood and no cultural practice, could not acquire intelligence at all.
So, I don't understand how the argument is at all credible. We have computers that do this. They are called robots. It may be the case that the best route to AGI is to put a robot through childhood. It would need an advanced enough learning framework first. All ML research is currently hashing out what that might be.
Surely you don’t think changing size has much to do with what makes humans intelligent.
Now, I'm not going to claim this will entirely enable AGI. But the base argument of "computers can not possibly have bodies" is nonsense.
This is an essential realization. The author's examples (swimming, bicycle-riding) are instructive. Also consider that each of us learned to understand a human language - an extremely complex task - with little explicit teaching. Our minds somehow learn to decode the complex patterns in language. Also languages often often refer to a very complex natural environment we must also perceive and comprehend. That amazing capability was not constructed.
For a shorter intro to Polanyi's thoughts, try 1964 book The Tacit Dimension.
Not ruling out AGI as possible. The question is whether we will be able to communicate with it if its referents are very different from ours. And then there's the question: why would it cooperate rather than compete?
Sure. And of course, do not forget that the testers are known for a disposition of potentially being fooled - and that we are still, collectively, societally, civilization-wise, generally unable to do proper selection.
Which is not in the script, unfortunately. Had not to be, because it was just a construed example of a successful Turing test. If not for the turn of events that a human may be made upset by the assumptions, but the mockery test was stated as intended to «provoke an emotional response», and juicily the invented machine, having failed all "common sense" reactions, responds exactly like a proper human to that.
Citation needed. This sounds exactly like something a human would claim!
Our intelligence is very much tailored to the environment we find ourselves in. The specific/general intelligence dichotomy is false.
wait until the author discovers robots
The other question would be: When I say the above, do the author of the article and the people who supported its publication see that as a sign that the AI field doesn’t want to hear their “truth”?
I'd be astonished if that statement were correct. For it to be a viable hypothesis it must at least begin to explain how the memories of my school graduation and the joy I felt are stored in the brain. But scientists like John Searle are dismissive of "wild ideas" and say consciousness is just a biological process in the brain, as he does at the beginning of this lecture [1]. Ok, I guess that settles it.
I'm going to assume AGI means human-like general intelligence, but emotional intelligence plays a large role in human intelligence. I'm surprised this isn't talked about more since we were all raised on Star Trek's Spock emotionless self and especially Data, trying to find the emotion chip so he can finally be human.
[1] Consciousness as a Problem in Philosophy & Neurobiology (John Searle)
What is the alternative to that statement?
Our consciousness could be "in the cloud"
On several occasions I've had people say they can prove consciousness is in the brain by saying that removing such and such part of the brain changes a person's emotions and thoughts.
My reply is always the same "If I removed parts from a radio receiver it wouldn't be able to pick up the signal as well"
The notion that consciousness is in the cloud is at least as plausible that consciousness resides wholly in the brain
Well, no. Because there is no evidence of a "central consciousness cloud" and no evidence of a communication mechanism to that cloud.
Like in Shakespeare's Hamlet: «To sleep, perchance to dream - ay, there's the rub, for in this sleep of death what dreams may come...». Afterlife as "plausible" - a critical suspect.
Note that in the history of 'plausible', it was also used for "nonsense, sophism, manipulation that rubes would applaud" (as opposed to "not only apparently" 'true') - the adoption was extensive (as opposed to, say, "what a sensible one would applaud") in its legitimate coverage. It has not been restricted de facto, as it may not de jure.
Who reviewed this disasterpiece?
Secondly, this would be more of a technical problem, not really "philosophical".
Thirdly, there exist disciplines which suffer from a tendency towards sophism - "persuasion", not "heuristic" -, as well noted in their history.
Edit: Fourth, sophism gives an illusion of "understanding" that has real world practice as a cruel challenger and cure.
Lol, what?
> Finally, I will argue that the belief that AGI can be realized is harmful. If the power of technology is overestimated and human skills are underestimated, the result will in many cases be that we replace something that works well with something that is inferior
That "«something that works well»" would be the mis-estimating thing.
(I am checking if I have missed some more sophisticated justification for the statement - or a development out of such non trivial paradox -, but for now I am not seeing it.)
However, most advocates of AGI (and advocates of strong AI) will today exclude Descartes’ immaterial soul, and follow the arguments of Yuval Harari."
What are we doing here? If we bring "immaterial soul" to the argument as a requirement than the author is right, AGI is unrealizable.
ps. I flagged the article, not up to HN's standard IMHO.
Don't waste your time, this is blogspam.
Westworld