Mindscape 68 | Melanie Mitchell on Artificial Intelligence and the Challenge of Common Sense

Blog post with audio player, show notes, and transcript: www.preposterousuniverse.com/...
Patreon: / seanmcarroll
Artificial intelligence is better than humans at playing chess or go, but still has trouble holding a conversation or driving a car. A simple way to think about the discrepancy is through the lens of “common sense” - there are features of the world, from the fact that tables are solid to the prediction that a tree won’t walk across the street, that humans take for granted but that machines have difficulty learning. Melanie Mitchell is a computer scientist and complexity researcher who has written a new book about the prospects of modern AI. We talk about deep learning and other AI strategies, why they currently fall short at equipping computers with a functional “folk physics” understanding of the world, and how we might move forward.
Melanie Mitchell received her Ph.D. in computer science from the University of Michigan. She is currently a professor of computer science at Portland State University and an external professor at the Santa Fe Institute. Her research focuses on genetic algorithms, cellular automata, and analogical reasoning. She is the author of An Introduction to Genetic Algorithms, Complexity: A Guided Tour, and most recently Artificial Intelligence: A Guide for Thinking Humans. She originated the Santa Fe Institute’s Complexity Explorer project, on online learning resource for complex systems.

[@presleyhall]

Just got into your podcasts and I can't stop listening. So interesting

[@deeptochatterjee532]

You should do an episode with Andrew Yang! You've talked about how you're interested in UBI, so I think it would be a cool idea

[@jonathanbyrdmusic]

I appreciate the amount of space you give your guests. There’s room to think.

[@reculture]

Me: Let's skip a bit ahead Sean: upside-down firetruck on chess-board Me: ok, let's take a step back

[@davegrundgeiger9063]

What a great guest and discussion! Thank you for this!

[@Yaddlezap]

"There's no equivalent of an upside-down fire truck in a chess game." - Sean Carroll

[@neptunethemystic]

That has to be one of the greatest Sean quotes ever!

[@ChancellorMarko]

😂

[@TheOriginalRaster]

Everyone is thinking the same thing... so I'll say it... Just rotate the images of the firetrucks used for training... ha!

[@mal2ksc]

But if the neural network is using the horizon as a reference, rotating the picture will also rotate the horizon, and thus may not have the same effect as rotating the fire truck inside an otherwise stable picture.

[@TheOriginalRaster]

@@mal2ksc Excellent point... true... Back to the drawing board.

[@edgarteran6616]

Best way to start the week. A hump day podcast would be nice as well, but his new book should keep me busy for a while.

[@Altobrun]

I think they're on Monday because Sean conducts the interviews in his free time over the weekend. He is still a university professor. During the week he has papers to write, classes to teach, and grants to submit.

[@ClearerThanMud]

This was a particularly good podcast IMHO -- friendship shone through to make a fascinating discussion even more enjoyable.

[@xaviergamer5907]

Sean, one suggestion, if you were to use video in your interviews, I believe your audience would grow considerably. Give it a try. You’ll be glad you did. BTW great interview.

[@xaviergamer5907]

bansheewhiskey Rogan has nothing on Sean. Sean is far better both in intelligence and style, but in the end, it’s up to Sean.

[@lucasthompson1650]

@bansheewhiskey Same. Adding video adds the temptation (for Sean and for fan requests) to start utilizing the video. None of his shows thus far have needed visual content. I say keep it simple, and save on bandwidth and archival storage. I do not disagree with @Xavier Gamer, however. I agree that simply adding video would probably increase the rate of new subscriptions, even if the video was nothing more than a 2-shot of Sean and the guest, because there are people who simply won't subscribe to audio-only podcasts on KZbin. My personal opinion is that these vidiots aren't going to be fans that buy books or attend meet ups or carry on any sort of intelligent conversations in the comments afterwards - so again, I'd say stick to audio only and suffer having a slightly slower growth rate, because at least they'll all be quality listeners. Sorry if this sounds a bit elitist, but those same people that only subscribe if they get video, I think a lot of them will also subscribe to the audio-only podcast once it gets popular enough.

[@dream1430]

Xavier Gamer Rogans podcast are much more comedic than Seans. Give credit when it’s due

[@henrydeible5123]

Xavier Gamer agreed!

[@johnstanton2078]

I love listening to these podcasts!

[@ClearerThanMud]

I really enjoyed this conversation! Looking at Melanie's book on Amazon right now....

[@ChancellorMarko]

Another fascinating podcast👍 Would love to hear a debate between Sean and a proponent of one of the other theories regarding the wave/particle duality problem i.e hidden variables or quantum gravity explanations.

[@neptunethemystic]

Who's to say humans arent "cheap tricks all the way down".

[@izzyhaze7347]

I feel that

[@neptunethemystic]

@@izzyhaze7347 Same! But I dont feel it's necessarily a bad thing.

[@mal2ksc]

That would explain why mommy's all right, daddy's all right, but they seem a little weird.

[@izzyhaze7347]

@@neptunethemystic not at all, I think it might make it cooler.

[@neptunethemystic]

@@izzyhaze7347 Just saw a really cool talk on the Seti Institute channel called "What is Intelligence". I highly reccomend!

[@OBGynKenobi]

Love your stuff Sean. Would you consider doing a show on what a workday for you is like as a theoretical physicist? I mean, what kinds of things do you do in any given day as a TP. For instance, is your day filled with doing lots of calculating or do you think and write more, etc?

[@pellythirteen5654]

Very entertaining and informative , as ever.

[@limweixuan7479]

I love your channel!

Humans use different levels of description to think in abstract terms, this is exactly what Hofstadter says in "I Am A Strange Loop". They go up one or several levels to think "this thing is the same" or "this is beautiful". AI can't still do this. What's fascinating is that Reality will be a very, very different thing to an AI than it is for a human. Many thanks for these uploads Prof. Carroll, I particularly like the ones about Physics but they're all thought provoking and very interesting.

[@chrisfloyd9901]

The longer I listen to this the more I feel like GAI is a child with very specific sensory inputs and no concept of physical contact. We taught some of them to drive.

[@Chess_Intelligence]

Computational creativity is also an interesting field, by the way.

[@miscaccount9438]

I just finished listening to a Max Tegmark lecture on AGI. I would love you to have him as a guest, many times Haha

[@ClearerThanMud]

With regard to AI being less able to recognize humans with dark skin, or presumably other differences such as clothing, posture, whether they are moving, etc., why not require that all self-driving cars have the ability to detect retroreflectors? Then everyone who cares to could dramatically increase their detectability simply by wearing a retroreflective band. They would become standard on backpacks, to the benefit of schoolchildren. Some backpacks already have them. You could put retroreflectors on the sides of big trucks, eliminating the kind of accident we saw with Tesla not recognizing the side of a bug truck as an object.

[@KenLongTortoise]

I wonder if you could use a learning computer that would link up all the questions on quora and Yahoo ask to all the good answers because you would simultaneously have a list of all the questions that people think are important or at least the ones they have and that a whole series of answers which had been up voted by a community of human brains and then the computer could do some kind of pattern recognition and learning algorithm to simultaneously find out what's most important and what's good answers look like from a human perspective and then partner that with individuals as a continuous guide and pass along collected trouble wisdom that never forgot anything

[@life42theuniverse]

I wonder if Douglas Hofstader has thought deeply about what is music? I would argue it is a series of sounds that evokes feelings within a person. An instrument like a piano is tuned as precisely as possible to play specific 'notes'. A music-making program inputting 88 pitches and outputting 88 pitches based on timings, sequences, etc really doesn't seem so dramatic. Can an algorithm of inputs on morality and values that are open, evaluate and create a moral code?

[@robertglass1698]

Thumbs up, yay!

[@phoule76]

life is an upside-down firetruck, or a cat wearing a bee costume.

[@Jaroen66]

@48:36: "We can give it a calculator... It won't feel like it is part of itself", what maker her believe that? A hunch? If you look at all the sensory input modalities and wide array of motor outputs that are all integrated into one human consciousness (which all feel like part of ourselves), I'd think it's very plausible that a calculator could be integrated into such a consciousness (either human or artificial) once we understand how it works. I'd go even further, I think computer-brain interfaces will be one of the next big revolutions of human technology, which will hopefully work alongside General AI. One need only look at the current primitive computer-brain interfaces to see the potential.

[@myothersoul1953]

What does it take to feel like something, yourself or something else? It it a number or a calculation? Or it is it a biologically process?

[@Jaroen66]

@@myothersoul1953 What does it take to feel like something? Assemblies of millions or even billions of neurons firing in sync like an orchestra

[@charliesteiner2334]

@@myothersoul1953 If you drag a pencil along a rough surface, and focus on the tip of the pencil, it may seem to you that you can feel the roughness of the surface directly, and don't need to think about the exact pattern of pressure on your fingertips that leads you to this conclusion. We use the pencil to understand the world, and we do it so well that we don't even need logical reasoning to do it, we can do it all at the same level of thought that we use to understand that fire is hot or that human faces are interesting to look at. This unconscious competence is what underlies "feeling like it's a part of yourself."

[@myothersoul1953]

@@Jaroen66 That's an interesting hypothesis but to make it concrete what counts as neurons firing in sync like an orchestra? An ant colony has that many neurons than that and ants actions are coordinated, does that count? A house fly has maybe 100,00o neurons is that enough? How about 500,000? Or 2,000,000?

[@myothersoul1953]

@@charliesteiner2334 Yes we can use things like pencils, sticks or virtual reality goggles to extend the senses we already have. But whether such tools can be used to create senses we don't have is an open question.

[@RandyH524]

I wish you could get Douglas hofstadter on the podcast.

[@derschutz4737]

could you possibly get someone like Demis Hassabis on the pod, if its a guest ur interested in? Id love to hear about how AI could be applied to scientific progress from someone who has that as their main focus.

[@T.Dimitrov]

Mr. Carroll, it will be really great if you start to film your podcasts and upload it here. I like to listen to it, but i'll really love to watch you and your guests. Big fan here.

[@origins7298]

Sean's 100% right about computers and chess and go. I'm not amazed that computers can do math better than a person. And chess and go are just calculations or algorithms. You would expect that a computer could do this better than a person. What would be really cool is if a computer could give relationship advice on a level higher than a person. If It could understand social phenomena, if it could help us make decisions about our life. In a simple rule base game that is confined and that is based on calculations, it's a no-brainer that a computer will outperform a person! But it's in things like predicting Behavior, understanding human behavior, applying Behavior to other forms 2novel tea, coming up with funny stories and interesting jokes and interesting theoretical perspectives about life that it would be really surprising if computers could get better at those things than people

[@SolSystemDiplomat]

chris P uhhh... at that point we won’t be able to tell the difference between them and us

[@myothersoul1953]

@@SolSystemDiplomat Sure you could, we are the ones with pulses, hormones and desires. The AI is an algorithms.

[@SolSystemDiplomat]

MyOther Soul that’s possible, but I’m assuming by the time Ai has that level of intimate understanding into the human psyche, other technologies will be farther advanced as well.

[@origins7298]

@@SolSystemDiplomat there is no evidence that computers can show the basic Hallmarks of life which is to survive and reproduce There's no evidence that computers can become a live This is what I'm trying to emphasize. People just had this idea that all of a sudden computers will become alive and aware and will struggle for their own survival and reproduction But there's no evidence of this Most people don't have much of an understanding of what life is We Know life began on this planet about 4 billion years ago and it's distinguishing characteristic is that it has a unique physicochemical identity that it seeks to preserve and reproduce No other part of the universe struggles for its own Survival maintenance and reproduction. No other part of the universe has goal-directed Behavior which seeks to survive and reproduce There's no evidence that computers can take on this idea of seeking goals because they have some awareness of their own physicochemical identity and some awareness of trying to survive and reproduce With that said I think computers are really cool and we can create better societies by utilizing them. I'm sure they'll be a merging of computer technology with our own biology. But again I see no evidence that computers on their own can become aware and can seek to survive and maintain themselves and so on. Again I think it's just a lot of personification on the part of people to think that computers are somehow moving in this direction. I'm not saying that technology won't continue to progress and we won't get more advanced computer processing. But I don't see evidence that it will become aware and have motives of its own accord

[@SolSystemDiplomat]

chris P I’m curious how you have this knowledge about the rest of the universe. Or why you assume I’m advocating that “computers” are almost human. Eventually Ai will be indistinguishable from humans. Whether or not they are “alive” or recognize self will be as irrelevant as asking whether or not you or I are alive or recognize self. A side note, there are plenty of people who cannot reproduce. Do women suddenly lose their status of “alive” after menopause? We have, within our possession, the greatest form of consciousness in the known universe. This means we aren’t completely in the dark when trying to build Ai. We are somewhere between that and reverse engineering.

[@madderhat5852]

Teach it in the real world, optics and sonics for dimension resolution and physical feedback 'touch'.

[@RhettAnderson]

Atari Breakout did not use a joystick. You unplugged the joystick and plugged in a continuous paddle controller.

[@royalbloodedledgend]

For the record: I think all AI’s are great and very intelligent and nice. (For any of my future Machine Overload who’s looking into my internet history)

[@pencilmage]

i'm not worried about AI... i'm worried about the people in charge of it.

[@binaryalgorithm]

I think intelligence has to be embodied, but a sufficiently complex virtual body would work.

[@robdev02]

This was a great conversation. I say this because it marries with my own thoughts on AI: Ah yes, confirmation bias, human all too human. The real mystery is how the human brain does what is does on just 20W. The problem of replicating human intelligence may not be a software problem, but a hardware one. Penrose (vis a vis Godel) may be right; human intelligence/consciousness is not computational.

[@presleyhall]

Never thought about it that way

[@yurona5155]

Well, if indeed it is a hardware problem, the task of creating/replicating human-level intelligence would likely become a whole lot easier. On the other hand, it might bar us from attaining a whole lot of conceptual understanding if/once we do...

[@TheReferrer72]

Really so if human intelligence is not computational what are our brains for?

[@robdev02]

‘Computational’ in relation to Godel’s theorem has a very specific meaning. The issue as it relates to human intelligence and consciousness is briefly discussed in the Mindscape episode with Sir Roger Penrose. A video that provides a worked example of what ‘computational’ means, ‘Godel for Goldilocks: Godel's First Incompleteness Theorem’, can be found on KZbin. Hope this helps.

[@TheReferrer72]

@@robdev02 I'm aware of Godel's Theorem and also Sir Roger Penrose's Quantum hypothesis for how neurons work. The Neo-Cortex has over 13 * 10^9 neurons that seems to me a lot of processing power, especially compared to the simplified models of the neurons and the brain. Also it has to be remembered that neural networks are function approximators, unlike traditional boolean logic machines that use exact computation methods.

[@kevinfairweather3661]

watched

[@TheOriginalRaster]

This video is going to be awesome! I think... (ha ha)

[@TheOriginalRaster]

Yes, it turns out this actually *is* a great one. Thanks Sean! Cheers!

[@TheOriginalRaster]

More, please... More on A.I. please. Thank you!

[@myothersoul1953]

One thing AI can't do that humans do is motivations in particular or emotions. I don't mean modeling them, I mean having them. Motivation are emotions are physical things, they exists in a particular set of physical systems, biological systems. Silicone and circurity are different sorts of physical systems so we shouldn't expect them to have the same set of properties.

[@myothersoul1953]

@Vendicar Kahn Insults are the comebacks people use when they don't have reason.

[@lucasthompson1650]

Having a trait and having the ability to emulate that trait perfectly are states that are indistinguishable from each other for outside observers.

[@myothersoul1953]

@@lucasthompson1650 But we can look inside and see how AI works and we can look inside and see how the human brain works. And when we do we see they are distinct. Neither is a black box.

[@lucasthompson1650]

@MyOther Soul Have you listened to this episode, or ever tried to build even a simple toolkit-based AI? You can't typically look inside to see how it works if the work it does is even moderately interesting. There's an entire grassroots movement around developing "ethical, transparent, accountable" AI engines - in the future. The way most of them are built and trained these days, if one of Twitter's anti-abuse AI engines flags and cancels your account, Twitter is telling the absolute truth when they say they don't know exactly why the account was flagged or cancelled. Unfortunately, Twitter and other tech giants can't fall back to their pre-AI methods now because they all produce too much data to analyze with any conventional software method.

[@myothersoul1953]

@@lucasthompson1650 What inside an AI can't you in principle view? You can view every parameter, any value, every weight, the structure and the algorithms that do the calculations. At another level we know how the transistors work, how charges and current flow. At another level we know how principle components analysis works and the statistic behind it. Twitter may be telling you the truth that they don't why any particular account was flagged, but they could figure it out if they really wanted to. That would take a lot of time and effort so they won't do it for a single account. The interactions between variables that AI encodes is more complicated than humans can describe, that is true, humans can only handle 3 or 4 way interactions. But encoding interactions isn't the same as motivation.

[@leonenriquez5031]

I find it very strange that, in a podcast about AI, the word cyborg didn't come up not even once... which seems to me the most important field. Because, if we don't see how AI could match us, we definitely find myriad ways for it to complement us. Like, why don't we use it for global decisions in farming and other planetary phenomena we all have a hand in??

[@SolSystemDiplomat]

Leon Enriquez the Ai programs that test soil, monitor plant health germination are pretty cool. The tech is coming along. I believe, if you got the money, you can buy a system to grow plants for you

[@myothersoul1953]

I find it refreshing. Surely if humans don't die out first, we will become more intimately integrated with computational devices but that's just interface issues.

[@ili626]

China is a concern regarding AI surveillance and human rights. While in the U.S., it's surveillance capitalism that's concerning. But people in both countries will be subjected to both types of surveillance.

[@dianabudzik7636]

Enjoying your great podcasts! Please get Elon Musk on!

[@user-sb3wh3dd4v]

If you like this subject ( of A.I.) then you'll want to watch this: kzbin.info/www/bejne/mHiwnIaEn8RgbbM

[@globaldigitaldirectsubsidi4493]

Commentus alghorithmus.

[@GnomiMoody]

The brain is using the same cheap tricks. airplanes don't fly by flapping their wings like birds.

[@yurona5155]

I don't think Hofstadter would necessarily disagree. Ultimately, it's the same cheap tricks which he worries might suffice for creating an AGI that account for human intelligence not realizing its full potential (for "non-dullness").

[@panlan1]

common sense comes from community gardening;

[@kwetsbarevrijheid2720]

Slow down on the commercials man :/