She's a 6-year old, even if she might be highly intelligent. There's no good reason to assume she can do that kind of mental arithmetic, or even that she has the knowledge to process the concept of multiplying two simple multi-digit numbers.
The blonde teenager made me laugh out loud. He was just holding the question about robots taking over the world in, blurted it's out then seemed a bit bummed that it would be a very long time. The youth are here to say that this is not the dystopian hellscape they were promised in the terminator movies.
Robots may not be able to take over the world for a very long time but artificial intelligence is ever increasingly replacing people's jobs and influencing people's behavior through advertising and content recommendations.
I love that she said (the kid) it took her several days to learn how to stack the cups. Its true. It took her weeks and months to be able to hold the cup too. Considering she was a baby and she learned all these really easy looking tasks in years. I mean all of us humans
I was wondering if she was maybe referring to professional cup stacking where you do it in special patterns as fast as you can and not just the act of stacking one cup on top of another
The game of chess is a great way to explain this concept that things easy for us are difficult for computers/robots: for humans, (without disabilities), moving the pieces is easy, but finding the best move is difficult. For a computer, moving the pieces is difficult, but finding the best move is easy.
Not really, chess is incredibly difficult for AIs and it took many decades of research for a computer to beat a human master at chess. There is so much complexity in the many moves you can make and the setting up of advance moves that most chess engines are just 'good enough' and not a challenge for a good human player. After we had a computer able to defeat a human master at chess the next challenge was defeating a Go master which is a more complicated game than chess and is a much harder challenge for AI.
@@_r1nky thank you for your comment. It made me realize I oversimplified. Please accept my apologies. I added an edit to recognize that for some people with disabilities, moving the pieces may be very difficult. I think maybe that is what you are pointing out, yes?
@Adam Sure, having better hardware helps, but the main issue is that the most commonly used algorithm to find the best move has a pretty terrible time complexity (exponential, like 30^n where n is the depth, or something like that). This gets too big, too fast even for today's hardware. Researched techniques that improved search tree pruning, heuristics, faster/more efficient data structures etc. were equally, if not more important, than improvements in hardware.
It occurs to me as I watch that this isn't so much a computer problem with learning as it is a human problem with teaching. The computer uses the language and problem solving that it has been given by humans. Because the movements/tasks described here were learned fairly naturally in life, humans have not had the reason to break down all the separate pieces until they've needed to teach it to a robot. To flip it, how would a computer teach a human how to solve the large number multiplication problem mentioned early on, or chess, mentioned later? I'm guessing that it would be unlikely to do that, as it has no concept of our possible mental steps that are required to learn it. Fascinating video, really enjoyed. #5 Hit it on the nose - and it's telling that he considers himself at heart a developmental psychologist.
The 6-year-old said the same thing, but in fewer words. "It took me days to learn how to stack cups at first." Think about it; she's had six years of her family just teaching her how to be a human being, and a "simple" task such as stacking cups still takes a few days to learn! A computer knows nothing until we program it, just as a newborn knows nothing until we teach them.
@@tk50000 Ok and computers are made with electrical wires that took us humans millions of years to figure out how to make after the existence of fire. 🙄 Also, a baby is learning even in the womb.
It's very easy to teach humans how to multiply huge numbers (you learned in grade school) and play chess perfectly (just try all possible moves and pick the one where you win). The steps aren't hard at all, but they are very repetitive and take your body a long time to carry out, and you'll be forgetting things and your attention will drift long before you've completed all the calculations. The advantage of computers is that they can do these simple but repetitive calculations very fast. But anything that doesn't involve simple and repetitive calculations becomes very hard for them, because they weren't designed/evolved to do those things in the first place.
@@eroshiyda Of course everything is learned, but you have to consider the relative difficulty. She took a few days to learn how to stack cups. How long do you think she needs to learn to play chess well? Which one is more difficult for humans? On the other hand, computer playing chess is already a solved problem. Stacking cups is not. Which one is more difficult for robots?
When we say "pick up a cup", we're describing an outcome that actually has very complicated steps. Like we move our muscles, we apply the right pressure, we use our past experiences, we observe the object visually etc. It is simple for humans of course, but if we describe all of these steps in the greatest possible detail, it would be an insanely long description. It would also be full of complicated rules, like what if the cup is very heavy, very big, slippery etc. When programming, you are required to break these steps down. But it's practically impossible to do that while accounting for every imaginable variation. So instead we create systems that learn on their own, which is also complicated. On the other hand, a computer's language is inherently tied to numbers (or rather things that can be represented as numbers). So operations involving numbers, such as multiplication, can be described in, relatively speaking, very simple steps with almost no variations or exceptions. It doesn't matter if it's 7*7 or 7777 * 7777, the instructions are still the same.
I like how once she gets to the grad she just stops trying to explain and let’s them explain because she knows that asking them to teach provides it’s own kind of learning, but also she has nothing left to add.
Hmm explanation about robot to a kid, must be a hard thing. Me when i see she's asian, and i meet asian kids almost everyday: well ok. Engineering is definietly her future Edit: before everyone cancel me, i'm asian ok? So relax
Didn’t expect to follow very far on this one, but actually understood a decent amount. Interesting parallels between human instruction ie prior knowledge, scaffolding, etc, and robotic instruction. Well done!
The idea that we link low level sensory models to higher level abstract models just seems to fit so well into the model of the human mind, with it's different areas of focus (at least how we currently see it). The real challenge is actually understanding what is happening in a neural model, whether human or machine, I've barely seen any insight into that.
If you look at 2:17 she said it took her a couple of days before she could do it properly so can it be said that Robots are still at their learning age or the maker of that robot is still unable to programme it at that level where it can hold that cup or place the cup just like it can calculate sums easily?
Pretty much. Our intuition is a result of billions of years of evolution which robots don't have. While AI is made by simulating evolution by natural selection, we don't have the computing power to train them to the extent that we've learned. At least, not yet :)
Since real world scenarios are susceptible to random changes, programming a robot for each case will be difficult. For instance, the cup can be placed in multiple ways, have different designs, or may be deformed (the list goes on, leading to many possibilities.) This adds to the complexity of machine learning algorithms, since each case will have to be separately programmed. Although the robots can be (theoretically) trained for all the cases, but it would be _very_ time and resource consuming. Thus, we use real-time analytics and try to maximize on that efficiency.
Wouldn't it be awesome if the sequence go backwards (like to professional goes first, and the kid goes last) and we (viewers) try to understand it. It becomes a self assessment as to our level of understanding.
I'm glad that she didn't just laugh at the teenager asking about robots taking over the world and gave an intelligent answer to calm any, rational or irrational, fears and concerns.
@@asdilia693 do you think people will become more like robots, without emotions, before robots become more like human, trying to be nice and such? Super interested in your insights on this.
Moravec's paradox is not a paradox. Our brains evolved to solve complex survival-critical problems like visual object recognition quickly and without conscious effort. It wasn't until we tried to engineer those skills that we realised how extremely complex they were. There is nothing self-contradictory about that.
no paradox is a paradox in the 'oh, look at this genuinely amazing fundamentally true self-contradiction I found!' every time someone presents a paradox it means they're being an idiot. they're thinking about something in a very, very wrong way. so if you're not perceiving a paradox here... good. you shouldn't. because it doesn't actually exist. no paradox does.
You don't really get to the point that actually IS paradox. That should be the idea of comparing a digital machine with an analog organism. The analog logic-unit is working with unit less amounts and relations that fade over time-periods. Comparing this with bits that occur in cycles is actually the paradox thing. A good brain is intelligent. That means that it is capable of making accurate predictions for chaotic situations over time. A good computer is capable of doing the most simplistic operations zillions of times per cycle.
They should have had the man at the end be the computer scientist explaining to everyone else. He seems like a better communicator. She didn’t do bad, but he kinda made her look rudimentary
Here's the answer to your Question Professor 11:47, in order to get the robot to accelerate in learning, we have to think outside the box of reality or maybe think like a robot lol. The answer to your question 11:09; By teaching the robot to pick up something like tofu for an example, we first have to calibrate angles on one hand of the robot that hold a plate with a fine edge that levels the surface it places its self on and the object Infront of it the other arms pushes the tofu by the center of its dimension the up the slope on the angled plate and then leveling its base plate to a center mass of its weight.
it's always amazing to me to see great teachers trying to explain complex ideas at various levels of difficulty, which is WAY harder than they make it seem. Thanks, wired, for these great videos, and thanks to the teachers for some of the most interesting explanations I've ever seen..
Seems like a key missing element is desire; a human typically learns to do something because they want to achieve a goal. A robot has no goals; it has instructions that it carries out meticulously, but it doesn't care if it succeeds, and doesn't have any understanding of why those instructions are important, so it is typically very bad at innovating or adapting to new situations. The underlying level of abstraction that humans will always apply is "I want X" and the cognitive effort is toward what effectively amounts to innovation; we make up a possible way to get X, we try it, and we adjust our approach based on feedback we receive. The robot doesn't want X, it wants the specific version of X that it has been programmed to want, and the circumstances to which it can adapt are therefore limited by the imagination of the programmer.
Well sort of. The ai field is currently working on ways for robots to recognize goals or set goals. Right now to train a robot in one task you set a goal which will reward the ai with feel good points. To make it easier for the robot to succeed we also set up smaller in-between goals and give it smaller amounts of feel good points. The robot has a probability matrix - a table in which there are small parts of actions or inputs it can do. In the beginning it tries out many random combination of actions. But as it finds more combinations that are more rewarding it can figure out a combination and sequence of actions that lead it to get the most feel good points. It does this by increasing the probability of choosing the same combination and sequence.
it appears to me, what's lacking in robots is the ability to think in forms (assigning a degree of correlation between a new form and known forms). while programs run with solely first principals, we can memorize first principles AND apply prior knowledge of similar concepts to aid in understanding the new
4:37 i am a Creationist. She goes, “ billions of years of evolution created humans”😂😅😂😅 Love it! I love it do hard every single time I hear an “evilutionist” say that. I don’t think she realized, in the moment, the silliness of such a phrase, if indeed she is a true evolutionist who does not believe in any religion of any sort.
I don't see the paradox in the statement, I find it very logical. Stacking two cups are for humans such an easy task, we don't really think about ist. We don't plan our actions, we just do it. The simplier the task, the less we plan because something other comes into play and that is intuition, a thing a computer (and therefore a robot) simply do not have. The more complex the task, the more we have to plan, maybe even break down the task to different steps and every step has it's own structure. Ask the kid not to stack the cups but to write a plan for this task. I think that would be the bigger challenge. Everything the kid forget to mention is probably when the intuition kicks in. Or am I the next candidate for reddit's "confidently wrong" here?
Made it to level 5 with a decent understanding. I suppose the easy and hard tasks for humans are just defined by the amount of data we've gathered and processed. Picking up an object has been done so often by us, but calculating big numbers hasn't, same with playing chess (especially since the tasks weren't a factor in evolution (until relatively recently)). In the same view, picking up objects needs tons of coordination and input; sound, hearing, movement, and sensation. In these challenging tasks like chess, AI excels because there is so little input and so much (automatically generated) training data, although it is still funny to see that even current huge general language models still struggle with arithmetic, even though brute force computers don't. Although specialized and tuned models have greatly helped in mathematics, helping to prove and make theorems in graph theory and representation theory. (If you wanna know more about current AI in math conjectures: th-cam.com/channels/9bkKi8Us7yevvP1KIBQHog.html, a research scientist at Deepmind) In my laymen's interpretation, I expect that generating "better" low-level latent spaces before higher-level latent spaces will be the key to improving AI generability and correct output. These relatively low-level latent spaces could maybe also be substituted with AI trained on specific data, which in turn could improve the generality of AI. (See facebook's Blenderbot 3 which uses the internet for aggregate output. In my eyes, the internet search can be seen as giving relatively high-level output as a low-level subsystem, which the aggregate uses to produce an even higher-level output.) They were also talking about purposive actions. In terms of AI, this is certainly a field that philosophers would love. Because the line between programmed and intent is starting to fade with AI systems, and although AIs are currently not at the point of sentience or purposivity. Looking at the exponential advancement of AI, I feel like this could change in the next 30 years. Exciting times!
As a normal human being, and how she's talking about picking up a cup, I am just mad thinking this is what I was programmed to be. A person, watching TH-cam videos about robots, using a robot interface, while drinking out of a cup.
Philosophy is the answer, the brain does 2 things, remember and predict. Don't worry about them learning faster, you can never 'learn enough', focus on deciding faster while avoiding terminal outcomes
As a CS student myself.. I think that Robots really hit a wall where there is as less as possible calculation involved. Like a robot can calculate throwing a ball precisely using projectile equation in a defined say,. plate however it will always struggle to retrieve it out of the plate.. even though it identified that the ball has reached its destination it cannot simply pick it up because the ball is anywhere in this plate this does not require excessive calculations only excessive scripts that break the backs of programmers because that aspect is very human and a human is needed to input that a kind of ability. Even in chess a Robot can calculate using the superior computing ability what is the next best move.... however it cannot calculate what is uncalculatable like picking the chess piece up physically and putting it in the next UNSPECIFIED area.
I found the clip of the robot hand playing chess particularly funny. I’ve often thought about the difficulty of using a robot arm and cameras to manipulate physical chess pieces in addition to the usual algorithms for playing chess. I had actually forgotten that it was called the Moravec Paradox. I recognized Hans Moravec’s name though and I probably learned about it ages ago in his book Mind Children.
It looks funny but then I tried to figure out how you could tell the robot it was picking it up wrong. It's doing everything correct based on a check-list (find piece, hold piece, pick up piece, move piece to desired location, place piece, let go of piece) but how do you tell it to look less like a 3 year old stabbing at its meal with a fork? Which then got me thinking, what is wrong with how it moved the piece? Is it solely etticate? Does it have to do with giving our eyes a better view of where we are placing it?
First Principle Observation of WYSIWYG Actuality is where Euler's e-Pi-i 1-0-infinity instantaneous Reciproction-recirculation fact of Newtonian Fluxion-Integral pivotal connection, nothing at No-thing Centre of Time Duration Timing is a perfect pivot for the Universe, ONE-INFINITY, which is the existence of i-reflection containment in wave-particle reciprocal connection and, as Dirac surmised, is the functional framework for the composition of probabilistic correlations of wave-particle inversions of 137 Partitioning nodal-vibrational formatting of all-ways all-at-once Timing-spacing phenomena, and "There's your problem", re-evolution circularity quantization cause-effect in which we are embedded bio-logical sequences of cycling histories, DNA and RNA nodal-vibrational emitter-receiver log-antilog interference positioning-location coordination of e-Pi-i infinitesimal shaping communication of AM-FM time-timing sync-duration chemistry. The appearance of a human body is hard enough to analyse, Totality of Mind-Body manifestation in this Universal Mirror Test at the Centre, I-reflection awareness-consciousness.., is ridiculously complex.
I love this concept. I was reading that Amazon could not figure out how to replace people loading boxes onto conveyors. Didn’t make them value the people though, just spending even more money to replace them all.
@@sarangtamirisa5090 maybe ‘saving time’ and our idea of efficiency neglects humanity and has as its goal maximizing profit. Imagine if efficiency was defined as maximizing human potential and implemented by having an ai watching human behavior
@@skypickle29 Well, of course efficiency is based on maximizing profits. We need a goal to progress towards and profit is a very driving goal. I think having an AI monitor your behavior and trying to maximize potential is less profitable and in some ways more inhumane than replacing the task itself by some form of automation.
I just experienced the software called Descript first hand. I am now convinced, way before robots take over the word, there is an indescribable danger in something a lot more discrete...
these people are basically trying to replicate millions of years of human evolution and condense it down as small as possible. it took millions of years for humans to develops eyes and limbs and then to learn how to use the two in sync. For a robot to go through the same process in a much much shorter amount of time is incredible.
I don't get it - if I was asked "Have you ever heard of Moravec's Paradox?" - I would truthfully answer: "YES; Just now." 🤷🏻♂ And no surprise, little Juliette was the quickest to note the non-traditional options of stacking cups! 👏🏻👏🏻👏🏻
These tasks aren't "easy". Try relearning how to walk as an adult. The little things we don't think about that allows us to actually walk becomes more pronounced when we actually have to put it together. These are the little things that you would have to program into the robot. Cognition is not easily programmed because the programmer has to be meticulously detailed oriented about things we don't cognizant of in out daily life.
Sure it takes a lot to learn how to walk, but compare that to learning to play chess at the grandmaster level. No matter how meticulous the details of walking are, most humans manage to learn it in a few years. Most humans can't master chess even if they devoted their lifetime to it. The level of difficulty is not even close. On the other hand, programming a computer to play chess better than a grandmaster is already a solved problem. Walking and stacking cups are not. Which one is easier now?
I have a hard time understanding why this isn't obvious to people. Robots are specialized tools that are good at what we created them to be good at. If you create a robot with the specific task of balancing something it will perform that task better than a human ever could because you give them sensors to collect all the necessary data and there is no real limit of how much data it could process at the same time. The same can't be done (yet) for humans, instead we are limited with the sensors we were born with, which gradually decay by age. We have even seen recently that robots can produce some degree of creative art, pretty much in the same way humans does, by copy others. It might be difficult to program a robot to create something completely creative and unique, but on the other hand, if it did. Would we recognize it as something creative or as an error in performance? Since we many times judge creative art based on who have done it rather than the end result.
the more complex answer, which I'm still going to state very briefly, can be articulated through my own lack of understanding, as a psycholinguistics nerd, of exactly how a technological device knows how to interpret and act upon lines of code. I certainly understand, at a gross/macro level, how meat-brains go about interpreting and acting upon instructions, but if we unfold that understanding and wrap it around silicon chips and metal-and-plastic shells and stuff, I don't really know how an object can physically understand and act upon a programming language. the psychologist at the end opens up with characterizing this problem in terms of the possibility of recognizing objects, but whichever way we want to approach explaining it, it's reducible to 'a problem of breadth and depth of understanding'.
I'm not sure how to pose this question because this is outside of my realm but here goes. Is it possible to scan our human brain activity as we do the tasks and as we look at videos of other people doing the task, can an algorithm meanwhile interpret the relationship of the brain activity in real time to the images/videos and actions. Would there be any value to such an experiment for the ai to be able to then create even better neural networks based on the data accumulated? Thoughts?
In my opinion, since the lower level tasks are so intuitive to us, we didn't need to devise algorithms and methods to do it. While games like Chess and multiplication were very technical tasks for us that needed us to devise algorithms and techniques, which is why one was easy to teach to computers and one is difficult.
It would appear both categories of tasks are taxing but one was not as useful out in the plains, so only one of them, arguably the more complex one, became "easier" for us. Multiplying two large number seems, in essence, seems more straight forward, but just tedious. Identifying stuff to pick up, generalizing, and then applying that to a future scenario seems like it is more computationally demanding. Thus the former easier to implement in computers. Am I missing something or is there not much to the paradox?
The last conversation just made me realize that a robot could not yet win a chess game if it actually needed to move the pieces on his own - let's take into account the time per play. I must also say that I have no idea what I am talking about; just thinking out loud.
We developed with the genesis of our design as sensory... then progressed to language. Robots were developed with language and are progressing to sensory. Some kind of mirrored effect like we're looking into our past. Consciousness seems to have decided it wants to exist a different way too. Balancing out its' understanding of the world.
Hey just a quick question that I was curious about. Don't we already have robots that do chess? Like the chess app? The only thing is that there isn't a physical computer, but dont we already have the 'brain' process of robotic chess game?
@@somerando6399 i think so lol. But isnt all robots just build on algorithms? Not trying to be combative but if there is 2 parts of a robot, perception and then actually doing the task, dont we already have that? I.e. chess playing against you based off algorithms.
That's a computer, a computer is not a robot. What they mean is actually carrying out the physical movements via a robot in a human way is actually pretty hard
@@guser436 ok, ok. Its starting to make sense. Just curious how is a robot now a computer, I mean it has a cpu, code to execute, how isnt it a computer?Also, I see what they mean by the physical aspect of carrying it out, but also in the video she said it was 2 parts, the perception, then actually doing it. A chess app already does the 'thinking' part of it right? like when i play chess on the app, its a computer generated opponent that is playing against me right?
As a software engineer, I'm a little disappointed that this was the topic. I'm not even entirely sure why it is even named other than to present the world of computing at an introductory level. They got to the answer that shows it isn't actually a paradox at all once she was talking to an expert. The human system is such a immense number of input receivers that have had millions of years to have selective pressures guide it to being really good at processing those inputs in certain ways. If you could build a robot with the same number of input receivers that are geared towards using and internally manipulating those inputs, it would, in theory, be able to produce a very similar to exact output - maybe in a more efficient manner since it is designed rather than being bogged down with our system's history and baggage of emergence. For the most part, it isn't that it is any harder for a machine to do something than a human. It's that we take for granted what the human body, as a natural computer, can do.
Chess has very minimal data even if it can be made very complicated in terms of posibilities. Chess can be described precisely and acurately with very few numbers. A picture of a chess game needs more info that will never be perfecly precice to the real world. We evolved to deal with increadably messy data, each species paying attention to information usefull to it.
I bet they'd be really fun to have at a party ;p But seriously, I'd just pull them away, grab a whiteboard, and just ask questions for hours on end. So fascinating....
“i can multiply big numbers” “okay four thousand one hundred fi-“ “no”
She's a 6-year old, even if she might be highly intelligent.
There's no good reason to assume she can do that kind of mental arithmetic, or even that she has the knowledge to process the concept of multiplying two simple multi-digit numbers.
Now that is intellegant
Ya, I think when she heard big numbers she was thinking more like 10 and 12. 😂
@@peterknutsen3070 ok
Imagine if she actually managed to solve that.
I love how Juliette's brain immediately went to different scenarios when asked how easy it is to stack cups. Smart kid!
She's a smart kid!
I was surprised. Did not think she would think of the difference of difficulty in stacking the cups on different sides
Smart kids are so adorable
She's adorable
This was the best part! I was blown away! The instant curiosity of children / ability to think outside the box.
The blonde teenager made me laugh out loud. He was just holding the question about robots taking over the world in, blurted it's out then seemed a bit bummed that it would be a very long time. The youth are here to say that this is not the dystopian hellscape they were promised in the terminator movies.
I, as a youth and an aspiring engineering, I plan to bring that future to reality soon enough
bro u're just imagining things he's just asking
Robots may not be able to take over the world for a very long time but artificial intelligence is ever increasingly replacing people's jobs and influencing people's behavior through advertising and content recommendations.
Du
@@user-dy7bv3qs4j it doesn't help that as AI is improving human intelligence is devolving. ( opinion)
let's go chelsea!! we interned together many years back haha
Ariana what are you doing here??
Hi jarvis !!
@@lallawmkimi8192He used to be a software developer.
I'd love to see a discussion between an AI expert, developmental psych expert, AND a neurologist.
Check out the latest Lex Fridmans podcast!
Host them on the Joe Rogan podcast at let them hit it too
I don't think you do 😂
you want a cognitive scientist
Lex fridman, you'll love them
I love that she said (the kid) it took her several days to learn how to stack the cups. Its true. It took her weeks and months to be able to hold the cup too. Considering she was a baby and she learned all these really easy looking tasks in years. I mean all of us humans
I was wondering if she was maybe referring to professional cup stacking where you do it in special patterns as fast as you can and not just the act of stacking one cup on top of another
😂 hi! That's me!!!!!
@@JulietteLeongoh wow it's actually you!
@@liliu5250 ❤️❤️❤️❤️❤️❤️❤️
Girl: “it’s easy for me to multiply big numbers”
Scientist: *I am going to end that girl’s whole career*
The game of chess is a great way to explain this concept that things easy for us are difficult for computers/robots: for humans, (without disabilities), moving the pieces is easy, but finding the best move is difficult. For a computer, moving the pieces is difficult, but finding the best move is easy.
Not really, chess is incredibly difficult for AIs and it took many decades of research for a computer to beat a human master at chess. There is so much complexity in the many moves you can make and the setting up of advance moves that most chess engines are just 'good enough' and not a challenge for a good human player. After we had a computer able to defeat a human master at chess the next challenge was defeating a Go master which is a more complicated game than chess and is a much harder challenge for AI.
Not for me, all those things are difficult for me 😤
@@Agret I don't think the time it took matters. Only the current stage. Like 20 years ago, AI sucked at drawing art, but now it can win competitions.
@@_r1nky thank you for your comment. It made me realize I oversimplified. Please accept my apologies. I added an edit to recognize that for some people with disabilities, moving the pieces may be very difficult. I think maybe that is what you are pointing out, yes?
@Adam Sure, having better hardware helps, but the main issue is that the most commonly used algorithm to find the best move has a pretty terrible time complexity (exponential, like 30^n where n is the depth, or something like that). This gets too big, too fast even for today's hardware. Researched techniques that improved search tree pruning, heuristics, faster/more efficient data structures etc. were equally, if not more important, than improvements in hardware.
Her ability to breakdown / simplify concepts is very impressive.
It occurs to me as I watch that this isn't so much a computer problem with learning as it is a human problem with teaching. The computer uses the language and problem solving that it has been given by humans. Because the movements/tasks described here were learned fairly naturally in life, humans have not had the reason to break down all the separate pieces until they've needed to teach it to a robot. To flip it, how would a computer teach a human how to solve the large number multiplication problem mentioned early on, or chess, mentioned later? I'm guessing that it would be unlikely to do that, as it has no concept of our possible mental steps that are required to learn it. Fascinating video, really enjoyed. #5 Hit it on the nose - and it's telling that he considers himself at heart a developmental psychologist.
The 6-year-old said the same thing, but in fewer words. "It took me days to learn how to stack cups at first."
Think about it; she's had six years of her family just teaching her how to be a human being, and a "simple" task such as stacking cups still takes a few days to learn! A computer knows nothing until we program it, just as a newborn knows nothing until we teach them.
@@tk50000 Ok and computers are made with electrical wires that took us humans millions of years to figure out how to make after the existence of fire. 🙄 Also, a baby is learning even in the womb.
@@eroshiyda how long ago do you think fire was discovered?
It's very easy to teach humans how to multiply huge numbers (you learned in grade school) and play chess perfectly (just try all possible moves and pick the one where you win). The steps aren't hard at all, but they are very repetitive and take your body a long time to carry out, and you'll be forgetting things and your attention will drift long before you've completed all the calculations. The advantage of computers is that they can do these simple but repetitive calculations very fast. But anything that doesn't involve simple and repetitive calculations becomes very hard for them, because they weren't designed/evolved to do those things in the first place.
@@eroshiyda Of course everything is learned, but you have to consider the relative difficulty. She took a few days to learn how to stack cups. How long do you think she needs to learn to play chess well? Which one is more difficult for humans?
On the other hand, computer playing chess is already a solved problem. Stacking cups is not. Which one is more difficult for robots?
Mike had the office next to mine in grad school. he's always been brilliant like this.
When we say "pick up a cup", we're describing an outcome that actually has very complicated steps. Like we move our muscles, we apply the right pressure, we use our past experiences, we observe the object visually etc. It is simple for humans of course, but if we describe all of these steps in the greatest possible detail, it would be an insanely long description. It would also be full of complicated rules, like what if the cup is very heavy, very big, slippery etc.
When programming, you are required to break these steps down. But it's practically impossible to do that while accounting for every imaginable variation. So instead we create systems that learn on their own, which is also complicated.
On the other hand, a computer's language is inherently tied to numbers (or rather things that can be represented as numbers). So operations involving numbers, such as multiplication, can be described in, relatively speaking, very simple steps with almost no variations or exceptions. It doesn't matter if it's 7*7 or 7777 * 7777, the instructions are still the same.
I like how once she gets to the grad she just stops trying to explain and let’s them explain because she knows that asking them to teach provides it’s own kind of learning, but also she has nothing left to add.
that child is so intelligent! she's got such a bright future.
based on what
@@flyingd5149 Based on the video. Probs. 🤡
hahah she was actually smarter than the 14 year old
Hmm explanation about robot to a kid, must be a hard thing. Me when i see she's asian, and i meet asian kids almost everyday: well ok. Engineering is definietly her future
Edit: before everyone cancel me, i'm asian ok? So relax
I think its scripted.
This escalated so quickly. Didn’t make it past the third explanation 😂
The first thing the expert said was, "people assume robots can do more because they did one thing" in 2000 words.
@@cevxj I prefer her explanation, I just had to pause and understand by sentence. Over all I think this was good job
Understandable. I found it all pretty straightforward. Though there have been some advances in the 20+ years since I dropped out of my AI PhD program.
Didn’t expect to follow very far on this one, but actually understood a decent amount. Interesting parallels between human instruction ie prior knowledge, scaffolding, etc, and robotic instruction. Well done!
As a Robotic Engineer I really liked that you can put word on the challenges we all face everyday.
The idea that we link low level sensory models to higher level abstract models just seems to fit so well into the model of the human mind, with it's different areas of focus (at least how we currently see it). The real challenge is actually understanding what is happening in a neural model, whether human or machine, I've barely seen any insight into that.
Wow, topic wise this is my favorite so far! Great expert and students as usual, too. Love this series.
If you look at 2:17 she said it took her a couple of days before she could do it properly so can it be said that Robots are still at their learning age or the maker of that robot is still unable to programme it at that level where it can hold that cup or place the cup just like it can calculate sums easily?
Pretty much. Our intuition is a result of billions of years of evolution which robots don't have. While AI is made by simulating evolution by natural selection, we don't have the computing power to train them to the extent that we've learned. At least, not yet :)
Since real world scenarios are susceptible to random changes, programming a robot for each case will be difficult. For instance, the cup can be placed in multiple ways, have different designs, or may be deformed (the list goes on, leading to many possibilities.) This adds to the complexity of machine learning algorithms, since each case will have to be separately programmed.
Although the robots can be (theoretically) trained for all the cases, but it would be _very_ time and resource consuming. Thus, we use real-time analytics and try to maximize on that efficiency.
6:26
Are these Mojo Jojo earrings 👀
I think so
Wouldn't it be awesome if the sequence go backwards (like to professional goes first, and the kid goes last) and we (viewers) try to understand it. It becomes a self assessment as to our level of understanding.
I think this works for us dummies coz we can learn easier
I've noticed the smarter the idea seems to get the more hand movement I see really intereesting
thank god the "expert" was a psychology prof. i would never have made it thru that conversation if i hadn't majored in psych
Why would you not have made it though if you haven’t majored in psych?
I am not a psych or a robotics major but i made through it. So you could've done it too
I'm glad that she didn't just laugh at the teenager asking about robots taking over the world and gave an intelligent answer to calm any, rational or irrational, fears and concerns.
the way that teenager said, "never heard of it" was hilarious. 2:46
It's the mark of a true genius when the genius in question can explain something so a kid can understand.
She was preparing for it. And also the issue is not that hard.
@@asdilia693 do you think people will become more like robots, without emotions, before robots become more like human, trying to be nice and such? Super interested in your insights on this.
@@dougb70 This idea is attractive, but personally I do not think people mostly would eventually be like rubots.
@@dougb70 I'm uncertain, sorry. Are you hinting that the first replier is kinda emotionless or are you genuinely asking a question?
I love the way these conversations progressed.
How would you describe machine learning in one sentence:
Giving data to a machine, and they start to learn based off that data.
Brilliant.
Moravec's paradox is not a paradox. Our brains evolved to solve complex survival-critical problems like visual object recognition quickly and without conscious effort. It wasn't until we tried to engineer those skills that we realised how extremely complex they were. There is nothing self-contradictory about that.
That’s what I was thinking too, but I was like maybe I’m wrong
you will find that a lot of paradoxes, if not all, are not paradoxes themselves but just problems with counter intuitive solutions
zeno, schrödinger's cat, ship of theseus, grandfather and the friendship paradox are all paradoxes that aren't really a paradox
What would be a better word for counter-intuitive phenomena?
no paradox is a paradox in the 'oh, look at this genuinely amazing fundamentally true self-contradiction I found!'
every time someone presents a paradox it means they're being an idiot. they're thinking about something in a very, very wrong way. so if you're not perceiving a paradox here... good. you shouldn't. because it doesn't actually exist. no paradox does.
Can we appreciate the Mojo Jojo earrings!
Totally
You don't really get to the point that actually IS paradox. That should be the idea of comparing a digital machine with an analog organism. The analog logic-unit is working with unit less amounts and relations that fade over time-periods. Comparing this with bits that occur in cycles is actually the paradox thing.
A good brain is intelligent. That means that it is capable of making accurate predictions for chaotic situations over time.
A good computer is capable of doing the most simplistic operations zillions of times per cycle.
They should have had the man at the end be the computer scientist explaining to everyone else. He seems like a better communicator. She didn’t do bad, but he kinda made her look rudimentary
Here's the answer to your Question Professor 11:47, in order to get the robot to accelerate in learning, we have to think outside the box of reality or maybe think like a robot lol. The answer to your question 11:09; By teaching the robot to pick up something like tofu for an example, we first have to calibrate angles on one hand of the robot that hold a plate with a fine edge that levels the surface it places its self on and the object Infront of it the other arms pushes the tofu by the center of its dimension the up the slope on the angled plate and then leveling its base plate to a center mass of its weight.
its the mojo jojo earings for me science is in good hands
Yeah
This is an amazing series... 🙂
The college student’s mojo jojo earrings are too cute!
it's always amazing to me to see great teachers trying to explain complex ideas at various levels of difficulty, which is WAY harder than they make it seem. Thanks, wired, for these great videos, and thanks to the teachers for some of the most interesting explanations I've ever seen..
The same difference between creative coding and mere data entry is sentience.
Brillant discussion most people don’t have the opportunity to hear, so thank you for that opportunity ❤
Dudes 14 and about to be a junior, no way. I was in 8th-9th grade at that age.
Seems like a key missing element is desire; a human typically learns to do something because they want to achieve a goal. A robot has no goals; it has instructions that it carries out meticulously, but it doesn't care if it succeeds, and doesn't have any understanding of why those instructions are important, so it is typically very bad at innovating or adapting to new situations. The underlying level of abstraction that humans will always apply is "I want X" and the cognitive effort is toward what effectively amounts to innovation; we make up a possible way to get X, we try it, and we adjust our approach based on feedback we receive. The robot doesn't want X, it wants the specific version of X that it has been programmed to want, and the circumstances to which it can adapt are therefore limited by the imagination of the programmer.
Well sort of. The ai field is currently working on ways for robots to recognize goals or set goals.
Right now to train a robot in one task you set a goal which will reward the ai with feel good points.
To make it easier for the robot to succeed we also set up smaller in-between goals and give it smaller amounts of feel good points.
The robot has a probability matrix - a table in which there are small parts of actions or inputs it can do. In the beginning it tries out many random combination of actions. But as it finds more combinations that are more rewarding it can figure out a combination and sequence of actions that lead it to get the most feel good points.
It does this by increasing the probability of choosing the same combination and sequence.
AI researchers consider progress in AI to be about making AI better at seeking goals
i hate the last guy for reminding me that electrical sockets look like faces. now i won’t be able to unsee that for a while 💀💀
it appears to me, what's lacking in robots is the ability to think in forms (assigning a degree of correlation between a new form and known forms).
while programs run with solely first principals, we can memorize first principles AND apply prior knowledge of similar concepts to aid in understanding the new
Very cool information! Thanks
Great job Maria! I love your earrings
Please do more episode like this, the content is amazinggg
4:37 i am a Creationist. She goes, “ billions of years of evolution created humans”😂😅😂😅 Love it! I love it do hard every single time I hear an “evilutionist” say that. I don’t think she realized, in the moment, the silliness of such a phrase, if indeed she is a true evolutionist who does not believe in any religion of any sort.
those mojojo earrings
Ikr?!
RIGHTT!!!
As someone with a 15 month old, eating broccoli is NOT easy for humans to learn.
For me, the first 15 minutes are not very useful. But the last 4:34 minutes give me a lot of information
I don't see the paradox in the statement, I find it very logical. Stacking two cups are for humans such an easy task, we don't really think about ist. We don't plan our actions, we just do it. The simplier the task, the less we plan because something other comes into play and that is intuition, a thing a computer (and therefore a robot) simply do not have. The more complex the task, the more we have to plan, maybe even break down the task to different steps and every step has it's own structure. Ask the kid not to stack the cups but to write a plan for this task. I think that would be the bigger challenge. Everything the kid forget to mention is probably when the intuition kicks in. Or am I the next candidate for reddit's "confidently wrong" here?
Made it to level 5 with a decent understanding.
I suppose the easy and hard tasks for humans are just defined by the amount of data we've gathered and processed. Picking up an object has been done so often by us, but calculating big numbers hasn't, same with playing chess (especially since the tasks weren't a factor in evolution (until relatively recently)). In the same view, picking up objects needs tons of coordination and input; sound, hearing, movement, and sensation. In these challenging tasks like chess, AI excels because there is so little input and so much (automatically generated) training data, although it is still funny to see that even current huge general language models still struggle with arithmetic, even though brute force computers don't. Although specialized and tuned models have greatly helped in mathematics, helping to prove and make theorems in graph theory and representation theory. (If you wanna know more about current AI in math conjectures: th-cam.com/channels/9bkKi8Us7yevvP1KIBQHog.html, a research scientist at Deepmind)
In my laymen's interpretation, I expect that generating "better" low-level latent spaces before higher-level latent spaces will be the key to improving AI generability and correct output. These relatively low-level latent spaces could maybe also be substituted with AI trained on specific data, which in turn could improve the generality of AI. (See facebook's Blenderbot 3 which uses the internet for aggregate output. In my eyes, the internet search can be seen as giving relatively high-level output as a low-level subsystem, which the aggregate uses to produce an even higher-level output.)
They were also talking about purposive actions. In terms of AI, this is certainly a field that philosophers would love. Because the line between programmed and intent is starting to fade with AI systems, and although AIs are currently not at the point of sentience or purposivity. Looking at the exponential advancement of AI, I feel like this could change in the next 30 years. Exciting times!
So data is the key?
Love how the first kid had a completely different idea of what ‘stocking cups’ means in the context here. I like her way better!
Robot: It ain't hard if you don't move the cup lady!
And a lot of people still think the human body, brain and universe doesn`t have a designer and creater
Why does this prove anything
That college student is a friend of mine!
Congrats! Your friend was made to look like an idiot!
As a normal human being, and how she's talking about picking up a cup, I am just mad thinking this is what I was programmed to be. A person, watching TH-cam videos about robots, using a robot interface, while drinking out of a cup.
Undergrad student wearing mojo jojo earrings. What a throwback
Yep
Philosophy is the answer, the brain does 2 things, remember and predict. Don't worry about them learning faster, you can never 'learn enough', focus on deciding faster while avoiding terminal outcomes
The so called "simple" stuff is not easy at all. It is just so complicated that our brains are hardwired to keep the whole process unconscious.
Yep. We already have a framework to integrate perception and action readily. Even a baby can look at something and reach out to grab it
As a CS student myself.. I think that Robots really hit a wall where there is as less as possible calculation involved. Like a robot can calculate throwing a ball precisely using projectile equation in a defined say,. plate however it will always struggle to retrieve it out of the plate.. even though it identified that the ball has reached its destination it cannot simply pick it up because the ball is anywhere in this plate this does not require excessive calculations only excessive scripts that break the backs of programmers because that aspect is very human and a human is needed to input that a kind of ability.
Even in chess a Robot can calculate using the superior computing ability what is the next best move.... however it cannot calculate what is uncalculatable like picking the chess piece up physically and putting it in the next UNSPECIFIED area.
I found the clip of the robot hand playing chess particularly funny. I’ve often thought about the difficulty of using a robot arm and cameras to manipulate physical chess pieces in addition to the usual algorithms for playing chess.
I had actually forgotten that it was called the Moravec Paradox. I recognized Hans Moravec’s name though and I probably learned about it ages ago in his book Mind Children.
It looks funny but then I tried to figure out how you could tell the robot it was picking it up wrong. It's doing everything correct based on a check-list (find piece, hold piece, pick up piece, move piece to desired location, place piece, let go of piece) but how do you tell it to look less like a 3 year old stabbing at its meal with a fork?
Which then got me thinking, what is wrong with how it moved the piece? Is it solely etticate? Does it have to do with giving our eyes a better view of where we are placing it?
college student 😮💨 proud that’s my future right there
I knew I’d find you here lurking
Hopefully you know more than that machine learning as a machine that learns by now.
First Principle Observation of WYSIWYG Actuality is where Euler's e-Pi-i 1-0-infinity instantaneous Reciproction-recirculation fact of Newtonian Fluxion-Integral pivotal connection, nothing at No-thing Centre of Time Duration Timing is a perfect pivot for the Universe, ONE-INFINITY, which is the existence of i-reflection containment in wave-particle reciprocal connection and, as Dirac surmised, is the functional framework for the composition of probabilistic correlations of wave-particle inversions of 137 Partitioning nodal-vibrational formatting of all-ways all-at-once Timing-spacing phenomena, and "There's your problem", re-evolution circularity quantization cause-effect in which we are embedded bio-logical sequences of cycling histories, DNA and RNA nodal-vibrational emitter-receiver log-antilog interference positioning-location coordination of e-Pi-i infinitesimal shaping communication of AM-FM time-timing sync-duration chemistry. The appearance of a human body is hard enough to analyse, Totality of Mind-Body manifestation in this Universal Mirror Test at the Centre, I-reflection awareness-consciousness.., is ridiculously complex.
6:38 LOVE HER MOJO JOJO EARRINGS!
same
ooh this is fun especially because today was my first day of applied computer science
It so interesting hearing that the little girl took a few days to learn how to stack cups. We don’t even think about learning how to do tasks anymore
Cool EARRINGS!!! She had some great questions!!!
THOSE MOJOJOJO EARRINGS OMGGGGG
Yay!
Robots be like: fetch image of stacked cups, great success sticker?
That getting real old
I love this concept. I was reading that Amazon could not figure out how to replace people loading boxes onto conveyors. Didn’t make them value the people though, just spending even more money to replace them all.
I dont think it's an issue of valuing people but more of finding the solution to automate the task doing which can save a lot of time and resources
@@sarangtamirisa5090 maybe ‘saving time’ and our idea of efficiency neglects humanity and has as its goal maximizing profit. Imagine if efficiency was defined as maximizing human potential and implemented by having an ai watching human behavior
@@skypickle29 Well, of course efficiency is based on maximizing profits. We need a goal to progress towards and profit is a very driving goal.
I think having an AI monitor your behavior and trying to maximize potential is less profitable and in some ways more inhumane than replacing the task itself by some form of automation.
I couldn't focus on the college student, because I was focused on MONKEY EARINGS MOJO JOJO LESSGO
WIRED should do genetic engineering next
This woman is so intelligent and able to talk about robots, because she herself is a robot.
Excellent job!
I just experienced the software called Descript first hand.
I am now convinced, way before robots take over the word, there is an indescribable danger in something a lot more discrete...
these people are basically trying to replicate millions of years of human evolution and condense it down as small as possible. it took millions of years for humans to develops eyes and limbs and then to learn how to use the two in sync. For a robot to go through the same process in a much much shorter amount of time is incredible.
I don't get it - if I was asked "Have you ever heard of Moravec's Paradox?" - I would truthfully answer: "YES; Just now." 🤷🏻♂
And no surprise, little Juliette was the quickest to note the non-traditional options of stacking cups! 👏🏻👏🏻👏🏻
"So that was pretty easy, right?"...after the guy struggled to put the penny in his left hand 😂
Those Mojo Jojo earings are neat
These tasks aren't "easy". Try relearning how to walk as an adult. The little things we don't think about that allows us to actually walk becomes more pronounced when we actually have to put it together.
These are the little things that you would have to program into the robot. Cognition is not easily programmed because the programmer has to be meticulously detailed oriented about things we don't cognizant of in out daily life.
She kind of said that at 4:24
Sure it takes a lot to learn how to walk, but compare that to learning to play chess at the grandmaster level. No matter how meticulous the details of walking are, most humans manage to learn it in a few years. Most humans can't master chess even if they devoted their lifetime to it. The level of difficulty is not even close.
On the other hand, programming a computer to play chess better than a grandmaster is already a solved problem. Walking and stacking cups are not. Which one is easier now?
I have a hard time understanding why this isn't obvious to people. Robots are specialized tools that are good at what we created them to be good at. If you create a robot with the specific task of balancing something it will perform that task better than a human ever could because you give them sensors to collect all the necessary data and there is no real limit of how much data it could process at the same time. The same can't be done (yet) for humans, instead we are limited with the sensors we were born with, which gradually decay by age.
We have even seen recently that robots can produce some degree of creative art, pretty much in the same way humans does, by copy others. It might be difficult to program a robot to create something completely creative and unique, but on the other hand, if it did. Would we recognize it as something creative or as an error in performance? Since we many times judge creative art based on who have done it rather than the end result.
the easy explanation to this is: people do not know what robots are or consist of.
the more complex answer, which I'm still going to state very briefly, can be articulated through my own lack of understanding, as a psycholinguistics nerd, of exactly how a technological device knows how to interpret and act upon lines of code. I certainly understand, at a gross/macro level, how meat-brains go about interpreting and acting upon instructions, but if we unfold that understanding and wrap it around silicon chips and metal-and-plastic shells and stuff, I don't really know how an object can physically understand and act upon a programming language. the psychologist at the end opens up with characterizing this problem in terms of the possibility of recognizing objects, but whichever way we want to approach explaining it, it's reducible to 'a problem of breadth and depth of understanding'.
LLMs could make Moravec’s paradox obsolete in some aspects. LLMs already overcame some of the challenges of sensorimotor and perception skills now.
Similar to the Morasex paradox: The more you want it, the harder it is to get.
I'm not sure how to pose this question because this is outside of my realm but here goes. Is it possible to scan our human brain activity as we do the tasks and as we look at videos of other people doing the task, can an algorithm meanwhile interpret the relationship of the brain activity in real time to the images/videos and actions. Would there be any value to such an experiment for the ai to be able to then create even better neural networks based on the data accumulated? Thoughts?
In my opinion, since the lower level tasks are so intuitive to us, we didn't need to devise algorithms and methods to do it. While games like Chess and multiplication were very technical tasks for us that needed us to devise algorithms and techniques, which is why one was easy to teach to computers and one is difficult.
Those Mojo Jojo earrings are dope!
ikr
This scientists are unbelievable, but i guess everyone needs a hobby or two
It would appear both categories of tasks are taxing but one was not as useful out in the plains, so only one of them, arguably the more complex one, became "easier" for us. Multiplying two large number seems, in essence, seems more straight forward, but just tedious. Identifying stuff to pick up, generalizing, and then applying that to a future scenario seems like it is more computationally demanding. Thus the former easier to implement in computers. Am I missing something or is there not much to the paradox?
The last conversation just made me realize that a robot could not yet win a chess game if it actually needed to move the pieces on his own - let's take into account the time per play. I must also say that I have no idea what I am talking about; just thinking out loud.
Depends heavily on the skill level of the human opponent and what chess format is being used. But an interesting observation nonetheless
I love how I always start these videos like oh nice, I understand the concept and as the video progresses, I am more and more lost xD
We developed with the genesis of our design as sensory... then progressed to language.
Robots were developed with language and are progressing to sensory.
Some kind of mirrored effect like we're looking into our past.
Consciousness seems to have decided it wants to exist a different way too.
Balancing out its' understanding of the world.
excellent video and pedagogy
Hey just a quick question that I was curious about. Don't we already have robots that do chess? Like the chess app? The only thing is that there isn't a physical computer, but dont we already have the 'brain' process of robotic chess game?
Isn't that based on algorithms?
@@somerando6399 i think so lol. But isnt all robots just build on algorithms? Not trying to be combative but if there is 2 parts of a robot, perception and then actually doing the task, dont we already have that? I.e. chess playing against you based off algorithms.
That's a computer, a computer is not a robot. What they mean is actually carrying out the physical movements via a robot in a human way is actually pretty hard
@@guser436 ok, ok. Its starting to make sense. Just curious how is a robot now a computer, I mean it has a cpu, code to execute, how isnt it a computer?Also, I see what they mean by the physical aspect of carrying it out, but also in the video she said it was 2 parts, the perception, then actually doing it. A chess app already does the 'thinking' part of it right? like when i play chess on the app, its a computer generated opponent that is playing against me right?
As a software engineer, I'm a little disappointed that this was the topic. I'm not even entirely sure why it is even named other than to present the world of computing at an introductory level. They got to the answer that shows it isn't actually a paradox at all once she was talking to an expert. The human system is such a immense number of input receivers that have had millions of years to have selective pressures guide it to being really good at processing those inputs in certain ways. If you could build a robot with the same number of input receivers that are geared towards using and internally manipulating those inputs, it would, in theory, be able to produce a very similar to exact output - maybe in a more efficient manner since it is designed rather than being bogged down with our system's history and baggage of emergence. For the most part, it isn't that it is any harder for a machine to do something than a human. It's that we take for granted what the human body, as a natural computer, can do.
The Mojo Jojo earrings are amazing
yep
YAYY MARIA!! so proud of u
Chess has very minimal data even if it can be made very complicated in terms of posibilities. Chess can be described precisely and acurately with very few numbers. A picture of a chess game needs more info that will never be perfecly precice to the real world. We evolved to deal with increadably messy data, each species paying attention to information usefull to it.
I so dig the Mojo Jojo earrings :)))))
Same
I bet they'd be really fun to have at a party ;p But seriously, I'd just pull them away, grab a whiteboard, and just ask questions for hours on end. So fascinating....
It would have been comedy gold if the little girl could calculate with the big numbers in her head xD