Computer Scientist Explains One Concept in 5 Levels of Difficulty | WIRED 

“i can multiply big numbers” “okay four thousand one hundred fi-“ “no”

I love how Juliette's brain immediately went to different scenarios when asked how easy it is to stack cups. Smart kid!

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'd love to see a discussion between an AI expert, developmental psych expert, AND a neurologist.

Girl: “it’s easy for me to multiply big numbers” Scientist: *I am going to end that girl’s whole career*

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

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.

let's go chelsea!! we interned together many years back haha

Her ability to breakdown / simplify concepts is very impressive.

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.

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.

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.

This escalated so quickly. Didn’t make it past the third explanation 😂

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.

6:26 Are these Mojo Jojo earrings 👀

Wow, topic wise this is my favorite so far! Great expert and students as usual, too. Love this series.

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.

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.

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.

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?

Those Mojo Jojo earrings are 🔥🔥🔥

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.

Please do more episode like this, the content is amazinggg

Brillant discussion most people don’t have the opportunity to hear, so thank you for that opportunity ❤

thank god the "expert" was a psychology prof. i would never have made it thru that conversation if i hadn't majored in psych

I love the way these conversations progressed.

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.

I've noticed the smarter the idea seems to get the more hand movement I see really intereesting

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..

Can we appreciate the Mojo Jojo earrings!

This is an amazing series... 🙂

Dudes 14 and about to be a junior, no way. I was in 8th-9th grade at that age.

How would you describe machine learning in one sentence: Giving data to a machine, and they start to learn based off that data. Brilliant.

The same difference between creative coding and mere data entry is sentience.

It's the mark of a true genius when the genius in question can explain something so a kid can understand.

As someone with a 15 month old, eating broccoli is NOT easy for humans to learn.

its the mojo jojo earings for me science is in good hands

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

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.

The college student’s mojo jojo earrings are too cute!

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.

the way that teenager said, "never heard of it" was hilarious. 2:46

those mojojo earrings

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 RU-vid videos about robots, using a robot interface, while drinking out of a cup.

ooh this is fun especially because today was my first day of applied computer science

That college student is a friend of mine!

I couldn't focus on the college student, because I was focused on MONKEY EARINGS MOJO JOJO LESSGO

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?

excellent video and pedagogy

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?

Love how the first kid had a completely different idea of what ‘stocking cups’ means in the context here. I like her way better!

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.

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...

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.

Cool EARRINGS!!! She had some great questions!!!

Very cool information! Thanks Great job Maria! I love your earrings

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

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: ru-vid.com/show-UC9bkKi8Us7yevvP1KIBQHog, 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!

WIRED should do genetic engineering next

Undergrad student wearing mojo jojo earrings. What a throwback

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 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

6:38 LOVE HER MOJO JOJO EARRINGS!

college student 😮‍💨 proud that’s my future right there

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.

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

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

Similar to the Morasex paradox: The more you want it, the harder it is to get.

Robot: It ain't hard if you don't move the cup lady!

THOSE MOJOJOJO EARRINGS OMGGGGG

Robots be like: fetch image of stacked cups, great success sticker?

I want to volunteer for this. How do I sign up for this

Those Mojo Jojo earings are neat

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 💀💀

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.

Make one on time dilation

LLMs could make Moravec’s paradox obsolete in some aspects. LLMs already overcame some of the challenges of sensorimotor and perception skills now.

Very very cool

This scientists are unbelievable, but i guess everyone needs a hobby or two

Nice

Could you apply generalized game physics to the crumbling tofu problem? Like, instead of simulating 100 specific instances of how that tofu may crumble, just teach it the concept of "crumbling" (like environmental destruction in a game) and how that might apply to any object, inclusive of tofu. Then you could set your robo-finger pressure sensors or robo-vision squeeze detection to a generalized stopping point to prevent "crumbling". With the goal of making directives implicit rather than explicit.

My favourite robot is ED-209.

they just explain simple things in convoluted ways

This woman is so intelligent and able to talk about robots, because she herself is a robot. Excellent job!

For me, the first 15 minutes are not very useful. But the last 4:34 minutes give me a lot of information

Do this again but with recursion

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.

He relaxed after she assured him the robot takeover would be well past his time.

the ‘paradox’ exists coz current robotics tech is way behind that of humans in both sensory and massively parallel computing

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.

"So that was pretty easy, right?"...after the guy struggled to put the penny in his left hand 😂

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?

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 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! 👏🏻👏🏻👏🏻

Go, Chelsea, Go!

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.

When we develop the deep learning required to solve this, that is the time to worry.