Moravec's Paradox
en.wikipedia.org
en.wikipedia.org
Simply put, robots today do not have the combination of speed, accuracy, gentleness, strength and 'hand-eye' coordination needed to do simple tasks that a 6 year old could learn in minutes.
You want a 4 ton pallet lifted from spot A to spot B? There's a robot arm that can do that same task over and over flawlessly. You move that pallet 12 inches to the left and rotate it 10 degrees? Best case, it realizes it can't do it and stops trying. Worst case, you have 4 tons of probably-broken goods thrown all over the place.
There's always examples of specific subsets being done well- a robot that can grab and align parts quickly (so long as they don't overlap and they're the expected parts); a robot that can carefully reach into a cluttered container and grab just one item (but only one item per minute vs a human's 20); a robot that can carefully lift an egg without damaging it (provided the egg is within acceptable size and weight bounds).
You want the whole package of skills? For now, you need a human.
Maybe it only has to be 90% accurate. A remote-control human operator can help those who get stumped. Record the problem scenarios for later analysis to improve the next version. The bot will gradually improve, and human intervention will gradually shrink.
This applies to cars also. Auto-cars can contact humans if stumped, such as lane closures or officers directing traffic. This includes both the passenger (if qualified) and remote-control operators.
The bot will be conservative such "when it doubt, ask for help". As it learns to solve problems via experienced-based tuning (above), there will be fewer times it has to ask a human.
Is this even remotely close to being true? I'm not aware of any AI projects that can match an average human score on a standardised IQ test.
In fact I'm not aware of any AI projects that can understand how an IQ test paper works without prior coaching.
Most humans won't just attempt the problems but will be able to work out what the questions are asking, even if they have never seen those kinds of questions before.
So this seems wholly wrong to me. I don't disagree that basic motor skills and perceptions are an important and difficult part of the puzzle, but I absolutely disagree that abstract thinking is a comparatively trivial solved problem - or that board game solvers are anything other than a warm-up exercise.
Even forming responses is very automatic. You yell “Stop” you don’t think what’s the best phrase to convey they are about to hit a car.
Though like breathing you can pay attention and do something else.
High vs low level tasks are defined in terms of what you can do while doing the task. You can be thinking about a math problem for example while talking with someone.
With regard to your ranking test, is it talking with someone or thinking about a math problem that gets ranked as low-level? It would seem to work either way, in which case, I guess it would be both.
[1] https://www.thefreedictionary.com/higher+cognitive+process
understanding the meaning of language
Understanding is vague in these terms. You can notice someone calling your name while distracted with other activities. That’s an automatic process and assigning meaning for each word in English is a difficult task. How do you teach a computer what cells are in the abstract?
There is also some high level language decoding where people extract meaning from poetry etc. But, we need to get the low level task done first.
PS: As to ranking, their are a lot of high level tasks you can’t do at the same time. Often described as things that take concentration, but few language tasks make this list.
Well, maybe, but is it conventional usage in any of those fields to call the cognitive processes involved in producing and understanding linguistic communication low-level? Regardless, my ranking it high-level is not just my private usage, as you claimed. In fact, I would guess that it is a more conventional ranking than the one that seems to be implied by your method.
> How do you teach a computer what cells are in the abstract?
Well, precisely, it's at a higher level of abstraction than just picking out the word (I am pretty sure that's a conventional usage of 'higher'.)
>There is also some high level language decoding where people extract meaning from poetry, etc.
Poetry may be particularly obtuse, but that does not make all other meaning-extraction low-level. A lot of it, IMHO, belongs in that 'etc.'
PS to your PS: There are also concentration-demanding motor skills. And the point about your ranking method is that it seems to put thinking about a math problem at the same level as talking with someone, which you are claiming is low-level.
Plenty of people can hold a simple conversation while doing differential equations. As the conversation’s complexity increases that stops being the case.
There is clearly a difference when talking to a distracted person. But, decoding “Do you want a soda?” and responding appropriately seems to require little if any attention.
Neuroscience gives support of this in terms of what parts of the brain are involved in different aspects of language.
Being able to do differential equations and simple conversations but not complex conversations means simple conversations are lower level.
Thus, it’s not language that makes complex conversations a high level task.
And your argument is also beside the point, now that you have restricted it to simple conversations.
But, you don’t lose language skills during these activities.
Thus it’s (A, B, C, ..., or Z) and language. Thus language is clearly the lesser activity.
PS: People can often count or do memorized addition etc, but doing complex math tasks is off the table. It’s odd that some things even prevent you from recognizing symbols.
What a coincidence it is that out of all A..Z , you should have picked math, the one that doesn't work, as your first example!
How so? You can recall facts not do math. Try adding say 2437 + 5627 while learning to juggle. It’s fine if you have already learned to juggle or have memorized the answer. But not if you are both doing the computation and learning to juggle at the same time.
On the other hand people will hold a simple conversation while practicing juggling.
If the math example works, why expand the scope to A..Z? Simply having math as an exception renders it pointless.
As I said you can’t do math. You can recall facts, like say what the capital of France is or what’s 3 * 7, but you can’t do computation requiring working memory. So, at most your making a semantic argument based on misrepresenting what doing math is.
This is not a semantic argument, we can argue about the difference between practicing juggling vs learning juggling. But that’s all worked out in the literature and people doing decades of experiments.
You claimed that my categorization of language (and, specifically, the cognitive processes involved in understanding linguistic communication, which is what I was discussing in the post you replied to) as high-level, was contrary to professional usage, where it is, you claim, regarded as a low-level capability. In support of that claim, instead of doing the obvious -- offering some citations (as I had done) -- you present, as the accepted way to make the distinction, a test: language is lower-level than mathematics because you can have a conversation while doing mathematics in your head.
Or should that be that you can do mathematics in your head while having a conversation? The first problem with this test is its symmetry: in itself, it does not rank the one thing over the other, as you can switch them around (for that matter, it cannot even establish that they are at different levels.) To try and get around that problem, you insist that mathematics is a high-level one, and so language must be low-level.
Implicit in this move is the corollary that, by definition, one can do at most one high-level task at a time. I suppose that when people refer to a cognitive capability as high-level, that is explicitly what they mean (though I doubt it), but if so, it is an assumption that should be possible to verify empirically (and you should do so, if you want to make use of it.)
Therefore, your claim that mathematics is the high-level capability is begging the question, as, through the corollary of the previous paragraph, it is tantamount to an a priori claim that language comprehension is the lower-level one. You have not shown that language comprehension is lower level; you have asserted it.
To try to avoid the problem that mathematics is not feasible concurrently with all conversations, you next modified the claim to include only simple conversations. Not only does this render the test moot, as it excludes language in general (the observation that a simple conversation is easier than calculus is trivial and uninformative in the general case), it is also a tendentious move; for one thing, you have not made a similar bifurcation of mathematics. Also, as you acknowledge at one point, it implies that, in your way of looking at the issue, simple language comprehension (or, for that matter, simple mathematics) is on a different cognitive level than difficult language comprehension or mathematics (note that there is a four-way comparison to be made there.) Again, maybe most people think that is how the concept of cognitive levels are to be understood, but I doubt they do -- and if they do, then it is sufficient for the point I was making in my original post.
Now you have come to say that complicated conversations are so cognitively demanding that it is not possible to do any sort of math while engaged in one, while math is not so cognitively demanding that you can't have a simple conversation concurrently. From the corollary of your position (as explained three paragraphs up), that would make mathematics lower-level than language comprehension! This follows from your own position because the latter is capable of excluding the former, but not vice-versa.
This just goes to show that your test is not achieving what you think it does. It is possible that you are coming from some actual science, but if so, it is getting lost in your insistence on defending this test. More relevantly, it is possible that people working in the field do, in fact, regard language comprehension as a low-level cognitive capability, but this is not the way to show they do.
In-vehicle activities, such as listening to the radio or an audio book, were associated with a low level of cognitive workload. https://journals.sagepub.com/doi/full/10.1177/00187208155751...
This stuff is not going to show up in a single study, but you can start looking if you really want to gain understanding.
Again I am talking the minimum threshold, complexity of speech is hard to control for. A book can easily get rather complex after all. But, the baseline does appear to be low.
I think you’re getting hung up on the mechanics of language as a communication medium vs the message it conveys. Decoding “Two roads diverged in a yellow wood.” is much harder than “Do you want a coke?” not because English got harder but rather the message was more complex. Wood being a collection of trees vs dead plant matter, yellow referring to the leaves at a time of the year etc.
PS: Related literature does say using spoken commons for common tasks is distracting. However, it's not clear how much of this is due to the task and the poor implementation of voice recognition features.
from "Again I am talking the minimum threshold..." to the PS, you appear to be trying to perpetuate the tendentious tactic of only comparing simple linguistic tasks with arbitrarily complex tasks of other types. As I pointed out previously, poetry may be complicated, but that does not mean that every other language comprehension task is simple.
The PS merely seems to be backtracking from the certainty with which you previously presented your case.
I have an additional question about this post [1], where you say "being able to do differential equations and simple conversations but not complex conversations means simple conversations are lower level", which at least strongly suggests that complex conversations are on a higher level than simple conversations. Moravec's paradox, applied to this case, would imply that understanding higher-level (complex) conversations would be computationally simpler than understanding lower-level (simple) conversations. Would you agree with this, or do you think it is an exception to the paradox?
The fact that they we still can't generally solve crap like the lost robot problem which insects are able to handle should also strike you as astounding.
That's what the Moravec paradox means. Stuff humans find difficult due to logic, abstraction and theory are often easy for computers, and stuff that bugs can do naturally are still impossible for computers.
If we recall that proof search is (Turing-)undecidable, for the systems that we care about, then it becomes unsurprising that we do not put AI to the task, other than as a light at the end of the tunnel, since proof search is not only AI-hard, but as impossible for AIs as for humans. That said, there have been constant improvements to the automated generation of human-readable proofs, and the papers have the same shape as AI research papers from previous generations of AI, with expert systems and inference engines [2][3].
Your focus on IQ tests is a little unhealthy [1].
[0] http://us.metamath.org/mpeuni/mmset.html
[1] https://en.wikipedia.org/wiki/Intelligence_quotient#History
[2] https://www.ijcai.org/Proceedings/15/Papers/172.pdf
[3] http://poincare.matf.bg.ac.rs/~janicic/papers/2014-amai-afta...
[1] https://en.m.wikipedia.org/wiki/Raven%27s_Progressive_Matric...
Or, to put it another way, the hardest part about teaching a computer to do a classical IQ test is how to hold the pencil, not how to find the right answers.
I think the explanation of it is pretty simple. To ourselves, the unconscious things like vision and bodily movement tend to feel effortless. Whereas the consciously controlled things like reasoning involve conscious effort, and they can feel difficult to us. The "paradox" comes from assuming that the amount of effort or difficulty that we consciously experience corresponds to the actual degree of sophistication of the processing involved.
> A veridical paradox produces a result that appears absurd but is demonstrated to be true nonetheless. Thus the paradox of Frederic's birthday in The Pirates of Penzance establishes the surprising fact that a twenty-one-year-old would have had only five birthdays if he had been born on a leap day. Likewise, Arrow's impossibility theorem demonstrates difficulties in mapping voting results to the will of the people. The Monty Hall paradox demonstrates that a decision which has an intuitive 50–50 chance is in fact heavily biased towards making a decision which, given the intuitive conclusion, the player would be unlikely to make. In 20th-century science, Hilbert's paradox of the Grand Hotel and Schrödinger's cat are famously vivid examples of a theory being taken to a logical but paradoxical end.
> A falsidical paradox establishes a result that not only appears false but actually is false, due to a fallacy in the demonstration. The various invalid mathematical proofs (e.g., that 1 = 2) are classic examples, generally relying on a hidden division by zero. Another example is the inductive form of the horse paradox, which falsely generalises from true specific statements. Zeno's paradoxes are 'falsidical', concluding, for example, that a flying arrow never reaches its target or that a speedy runner cannot catch up to a tortoise with a small head-start.
> A paradox that is in neither class may be an antinomy, which reaches a self-contradictory result by properly applying accepted ways of reasoning. For example, the Grelling–Nelson paradox points out genuine problems in our understanding of the ideas of truth and description.
> A fourth kind, which may be alternatively interpreted as a special case of the third kind, has sometimes been described since Quine's work.
> A paradox that is both true and false at the same time and in the same sense is called a dialetheia. In Western logics it is often assumed, following Aristotle, that no dialetheia exist, but they are sometimes accepted in Eastern traditions (e.g. in the Mohists,[12] the Gongsun Longzi,[13] and in Zen[14]) and in paraconsistent logics. It would be mere equivocation or a matter of degree, for example, to both affirm and deny that "John is here" when John is halfway through the door but it is self-contradictory simultaneously to affirm and deny the event.
Is there some established knowledge that the relative difficulties of these AI tasks appear to run contra to? I believe it is just intuition and assumptions they run contra to.
Hmm, I'd say to be a paradox, it needs to be so unexpected that it's an apparent contradiction. Merely slightly unexpected is not sufficient.
"I did not expect you to arrive so early" is not a paradox.
"The left facing arrow on your website takes me to the next page." whilst both unexpected and unintuitive, is not a paradox.
Some edge cases:
"I am speaking to you in person whilst I am watching you on a live broadcast on television" is maybe a paradox.
A statement like "Fruit juice is as unhealthy as sugary soft drinks." could be a paradox for many people.
Thinking about the topic of paradoxes some more, I would say part of my objection is in the following.
You could distinguish between two sorts of paradoxes
1) where the phenomenon P itself is genuinely paradoxical
2) where the phenomenon P itself is not itself paradoxical, but where our knowledge of the world is such that our picture of P appears to be paradoxical.
In cases of 2), I object to saying that P is paradoxical. The reason for my objection is that doing so would be confusing the map for the territory. What is apparently paradoxical really has nothing to do with P itself. It is all a matter of a particular person's knowledge relating to P.
In those cases, P is in no way paradoxical, and it is misleading to call it paradoxical. It would be less misleading to say that some group of people have flawed knowledge relating to P such that there is an apparent contradiction in their knowledge.
In short, in cases of 2), P is not paradoxical, rather, certain people's beliefs about P contain illusory contradictions.
Calling something a "contradiction", on the other hand, is a statement about the territory.
> some group of people have flawed knowledge relating to P such that there is an apparent contradiction in their knowledge.
> certain people's beliefs about P contain illusory contradictions.
These are both (specific forms of) paradoxes.
Actual contradiction is also (a different, specific form of) paradox.
This is a common usage of the word paradox, and is not incorrect, but "apparent contradiction" is also in common use , also correct, and is the way in which the word is used in the term "Moravec's Paradox".
EDIT: Both usages of the term are compatible, in the same way that "dog" and "Labrador" are both correct terms with which to refer to a Labrador (but only one is correct when referring to, e.g., a Husky).
It seems that most of this type of paradoxes result either from having an accurate-seeming model that nevertheless fails to predict an experimental result, or from making an intuitive but incorrect assumption that a model predicts something which, in fact, it does not.
Things like the stellar parallax and Olbers' paradox are examples of the former class, whereas Simpson's paradox and the Monty Hall problem are instances of the latter.
[1] https://en.wikipedia.org/wiki/Stellar_parallax#Early_theory_...
Yes, the stars thing may have felt paradoxical but it isn't itself a paradox. The thing is that if all it takes for something to be considered to be a paradox is for it to feel paradoxical to a sufficient number of people the a lot of well-established bits of scientific knowledge -- that are in no way paradoxes to people who have the relevant undergraduate degrees -- are paradoxes because they feel that way to the population at large.
Paradoxes of this type are definitely paradoxes if they contradict something that appears to be common sense and a consensus position of some sort, irrespective of the level of rigor. The Monty Hall problem confuses even career statisticians, but that's not because the common but wrong answer is somehow based on sound statistical knowledge; it's simply because answer seems so intuitive and self-evident that we don't bother to "shut up and calculate".
It seems to me that to count as a paradox, a result has to be both contradictory—not just different—from what is expected; and, the result has to be especially surprising; in Bayesian terms, the expected outcome must have had a high prior probability.
And it's possible to argue about whether a particular map is flawed or not, such as it being flawed because it contains certain flawed assumptions. If we can show that a map in which it seems paradoxical is flawed, and show that a different map, in which it doesn't seem paradoxical, is more accurate, then this can be grounds for arguing that the phenomenon is no deserving of being called paradoxical.
This result is highly unintuitive to humans, and is hence a paradox.
There's a distinction I've made in another comment between appearing to be contradictory to reasonably-well established knowledge vs appearing to be contradictory to intuitions. I don't think the latter are paradoxes, and I think the case in question is one of these.
In addition an object can be "in freefall" whilst travelling upwards. A ball thrown upwards begins "falling" the moment it leaves your hand.
(And this is not just an arbitrary definition of the word "fall" - if you are catapulted out of a slingshot, you feel weightless from the moment you leave the slingshot, not just when you begin to come back down)
These direct consequences of Newton's laws of motion, and are unintuitive to most modern people, even those who learned Newton's Laws in high-school science.
They are apparent contradictions, and hence I would classify them as paradoxes.
However! Maybe the reasoning feels difficult for an evolutionary reason, because it actually consumes a lot of energy. That would contradict the claim that the processing involved is simple.
The role of consciousness is only to act as a sort of "global workspace" to reconcile any discrepancies that can't easily be evaluated and explained by the pre-conscious modules of the brain. As such the vast majority of sensory input never even makes it to the conscious level.
More body mass means more nerves that need to send and receive information from the central nervous system.
So even assuming a constant amount nerve endings in cm^3 of flesh between whales and humans a whale will have significantly more.
Also some people have argued that fish can't feel pain or are less sensitive to due to their lower nerve density from my understanding. A whale is a mammal so I would assume it's going to have similar nerve density. Which means the brain has to be able to take input from a larger volume of body so that brain in large animals would make sense to be larger.
The intelligence comparison is a little bit too simplistic though. How do you define intelligence? Are we humans really "more intelligent" than whales? There are a lot of human behaviors that are not intelligent at all, and which seem to be a lot less intelligent than what whales do (and pretty much any other animal on the planet for that matter).
I think the answer to most (though not all) of these questions is no. Just because a body is bigger doesn't necessarily make it more complex to control surely? For the most part large mammals have the same number of muscles and bones as a small one, and past a certain body size eyes don't get particularly larger.
I seem to recall that even very large dinosaurs could have comparatively tiny brains compared to mammals - did they control their bodies clumsily and observe the world much less acutely than mammals?
So I have always found this explanation a bit suspect - it is a "just so" story because we observe that whales and elephants don't seem all that smart yet they have bigger brains and bodies but that doesn't make the explanation true by itself. Has there been any actual research that supports the hypothesis that large bodies require bigger brains to control?
How the mammalian brain divies up this process depends on enviromental factors and the body plan. Dolphins and whales have specific parts of the brain used for ecolocation and the water pressure affects the volume of the brain. Elephants have larger areas devoted to using their trunk. Aye-Ayes have large visual cortices as they hunt inscets in the dark for food.
Fascinating, that could in a way be interpreted to imply that our conscious attention/perception is focused on the things that our minds cannot predict/control and everything else is "automated" (subconscious).
That could also be related to our perception of uncertainty and the way we deal with it. Our attention tends to be focused on what we are the most uncertain of.
- "Grass is always greener on the other side"
- "You only fully appreciate great things when you lose them"
- All the people getting anxious or even depressed because social media constantly barrages us with the perfect pictures of perfect peoples' perfect lives, all the while forgetting that on many objective scales the fact that you're even able to get depressed by something like that puts you in the jackpot-winners category in the lottery of life compared to a huge percent of the world's population.
The main lesson of thirty-five years of AI research is that the hard problems are easy and the easy problems are hard. The mental abilities of a four-year-old that we take for granted – recognizing a face, lifting a pencil, walking across a room, answering a question – in fact solve some of the hardest engineering problems ever conceived... As the new generation of intelligent devices appears, it will be the stock analysts and petrochemical engineers and parole board members who are in danger of being replaced by machines. The gardeners, receptionists, and cooks are secure in their jobs for decades to come.[7]
Automation has already been heavily cutting at the so-called low-skill jobs starting from Industrial Revolution. Or by some definition, from the Agricultural one somewhat earlier.
But the examples of the gardener and cook that were in your citation are actually quite good, I admit.
Gardening is hard to automate, but harvesting of somethings (grains, mostly) is heavily automated. Cooking is more amenable to augmentation than full automation, but there are exceptions mostly around processed foods.
we're least aware of what our minds do best
If you ask me:
How much is 2x3?
That's easy! I just have to imagine 3 stones.
Then I imagine two piles of 3 stones.
Then I count all those stones. 6!
It could be a nice approach to AI to try and breed (via AI/Genetic Algorithms/Whatever) a system that can solve math questions, independent on how they are phrased.
It is certainly possible to teach a machine how to reason about algebra from the first principles, allowing it to solve much, much more complex problems than multiplying small integers. The program doesn't know how it is executed, either, but it does have a much better understanding of what multiplying actually is than most humans.
Every concept that we can come up with is articulated in terms of concepts that have been stored in the minds memory (inference) the funny thing is that we are born with almost empty minds as we grow older it starts to feel up to the point where we cant even think without using language which is also not inborn but is acquired here on earth through the mind.
The problem is to see behind this layer. If we observe the mind through the mind we only see that it is a mechanism for storage, processing and inference of the stored data. Learning is simply finding new inferences between data or acquiring new data from the senses that are connected to the mind.
But as humans we are biologically limited, our senses can not even pick up everything this already implies that the what we know through the mind will always be limited (unless if you believe in evolution and that our senses will evolve).
Meditation does this with a sort of denial of the knowledge that we acquired through the mind (closely related to the nature of the ego) that is why highly spiritual people often say perceived reality is an illusion, and to some a deity is a way to see outside of the box.
AI is trying to simulate the mind with what we know about it through it.
I also feel like this is what the incompleteness theorem is about, we have some sort of axioms of Intelligence that we got through the mind but these finite axioms can not proof all the dynamics of Intelligence.
A cleaning lady is going to have her job longer than a radiologist.