10 differences between artificial intelligence and human intelligence

Support me on Patreon: / sabine
In this video I will explain what the main differences are between the current approaches to artificial intelligence and human intelligence.
For this I first explain how neural networks work and in which sense they mimic the human brain.
I then go through the ten most relevant differences that are: Form and function, size, connectivity, power consumption, architecture, activation potential, speed, learning technique, structure, and precision.
Finally I express my opinion that the benefit of research in artificial intelligence is not the reproduce human-like intelligence, but instead to produce types of intelligence unlike our own that complement our own abilities.

[@satanicmonkey666]

I like how you explain information. It is clear, concise, and to the point. I hope the channel gains the viewership the quality of the content demands in the near future. Thanks for making these videos.

[@georgehunter2813]

The German Scientist. Technical rationality spoken in concise articulate English. It's a German thing.

[@ixglocTV]

@@georgehunter2813I can tell you don't understand German because if you did you wouldn't have such a high opinion of German Scientists.

[@georgehunter2813]

Ich spreche nur ein bisschen Deutsch. How wonderful it would be if I were articulate in German as I have learned to be in English. German is a very articulate language. People can express things in German that other languages can not. Scientists in general are to be admired for what they have achieved, and for the torch light they provide to our pedestrian lives. I hold Scientists in high esteem. German Scientists are some of the best communicators. They are a good reflection on the society they come from. The Germans are a strong rational people with a genuine warm nature. That Hitler thing was just a recent blip in their history. The present day Germans are the same people who marched in 100, 000 strong at Waterloo to decisively tilt the balance in favor of Wellington and stopped Napoleon. These are my Germans.

[@ixglocTV]

@@georgehunter2813 "People can express things in German that other languages can not." This is true for every language. It's the reason why there are diff. languages.

[@ixglocTV]

@@georgehunter2813 "German scientists are some of the best communicators." The opposite is true. I can't imagine how you came to believe this. German scientists tend to write in an overcomplex and lofty language, which is partly due to the nature of the German grammar. But German scientists have given up on their mother tongue anyway and are switching to English like Frau Hossenfelder.

[@PiperPruiksma1986]

Most Underrated Channel. You are at the teaching level as PBS Spacetime

[@BillyViBritannia]

@D R I think she is a tad bit closer to "normal people" with her explanations. PBS is right there balancing on the edge of being too technical sometimes but I appreciate that too. If she could improve the editing and standardize some call to action scripts and community interaction I think she'd be right up there next to PBS, they complement each other really nicely.

[@rytr1995]

@@BillyViBritannia At a certain point in almost every Spacetime video, he gets too technical and my brain starts to glaze over everything that's being said.

[@PiperPruiksma1986]

@ozayevable you obviously don't watch the KZbin Channel SpaceTime.

[@rytr1995]

@ozayevable It's an education channel, you act like it's there to provide some ultimate scientific breakthroughs, or go incredibly deep into complex subjects?

[@Ben-xj6su]

PBS spacetime is overrated

[@miketacos9034]

When the creepy robot kid walks into school with a turtle

[@wishiwsthr]

Good one

[@NyscanRohid]

Hello, fellow Nazarene.

[@peterjansen7929]

That's not my main worry - the creepy robot cop is, who may shoot me for pointing a turtle at him!

[@Debilitator47]

@@peterjansen7929 This explains the boardroom scene in the original Robocop. (quick, somebody 'shop in a turtle into the guy's hand).

[@simateix6262]

lol

[@ataxias7]

I have a background in both physics and AI, and can say that Sabine's articles on physics are great, but this one focuses on secondary details rather then the actual main differences and is riddled with inaccuracies. The most serious one is that 1) modern neural networks consist of about 100 neurons; rather, make that approx. 100 billion neurons, 2) that activations (the output of neurons) are parameters of the networks; they are not - in the simplistic neural network shown, only connection weights (including biases, which can be seen as connections) are, and 3) that artificial neurons are not either "on" or "off", but use smooth, continuous functions; that was the case a few years ago, but nowadays very popular activation functions like ReLUs can turn neurons "off" and are not smooth. In general, what is called "artificial intelligence" and human intelligence don't have a lot in common. AI nowadays is based on statistical machine learning (neural networks are just one implementation of statistical machine learning), which is a process to optimize a function by mapping inputs to desired outputs in a way that holds as accurately as possible and for as many data points as possible, i.e., holds in a statistical sense. The only reason people call this "intelligence" is the following implicit argument: machines can learn to do tasks T1, T2, ...., Tn. Humans can also do tasks T1, T2, ..., Tn, and use intelligence to do them. Therefore, machines must be intelligent. However, the process, or underlying mechanics, matter A LOT - you can't just look at the outcome, or apparent behavior. We don't say that a calculator that can give you the square root of pi^7 is intelligent, just because a human who would be in a position to do that is intelligent. People get confused because AI, unlike calculators, is not explicitly programmed to solve these tasks. However, the fact that programs/machines can learn functions mapping inputs to outputs does not make them really "intelligent", in the same way that we don't consider fitting a curve on some data intelligent behavior. To sum up, we don't have a really good idea what "intelligence" really is, we simply realize that humans and other complex animals seem to be able to perform some tasks, while simpler animals cannot, and intuitively ascribe this difference to "intelligence". But just because some tool we made can reproduce the same apparent behavior, this does not mean that it qualitatively possesses the same attribute; in fact, we know it works very, very differently.

[@ninibupu]

she created the video 3 years ago, 2 years before your comment. AI just keeps improving rapidly, it's crazy how out of date your comment is.

[@ataxias7]

@@ninibupu Which parts are "crazy" out of date? Btw all my points of criticism were valid also for the time of the video, not only for when I wrote it.

[@johnbrowne8744]

Excellent. You are very good teacher.😊

[@1SpudderR]

John Browne...Hmm There are teachers...teaching, and/or learning from teachers? Your comment is ambiguous? Understanding the difference gives You the Communicating ability to comprehend what you have experienced..”Teaching or Learning?”...I know what I prefer! RDR

[@davidsmith5946]

HEY,,,,YOUR CALLED THE SAME AS ME BUT WITH AN E

[@ArtisticLogic]

To be honest I have had a better understanding of physics from Sabine than any other science channel out there. Her explanations are clear and easily understood. Good job on your channel

[@vzuzukin]

Just discovered your channel and happy I did. I loved your book “Beauty Leads Physics Astray” and found it a bold challenge of dogma. I find your contribution to round table discussions with celebrity physicists refreshing. You bring true Skepticism to the table - literally! Will be following your educational series 🙏

[@selocan469]

Another brilliant and informative video Sabine, Thank you very much. It would be also nice to mention the concepts such as Strong/General vs Narrow/Weak/specialized/ localized in AI, and supervised and unsupervised learning

[@dsnodgrass4843]

your last couple sentences really brought it together for me. I've always thought that our relationship with AI/Neural Networks would be like our primitive hunting ancestors' relationship with dogs; where their sensory strengths, combined with our discernment and toolmaking/tool use, helped us out-compete every other "apex predator" on the planet. Seems pointless (besides research) to ask AI to do what we already can do; when it can do so much else at a level impossible to us, with our particular and limited faculties.

[@I_dont_want_an_at]

we want AI to do what we already do, because we want to stop doing it. For example, we want to stop cleaning and make the AI do it.

[@Aziz0938]

Like what?

[@MegaRudeBoy69]

My base instinct brought me here and got suckered into learning something. Wunderbar!

[@NondescriptMammal]

You had no choice. Your instincts were predetermined. You had to watch this video.

[@ferretappreciator]

@@NondescriptMammal they didn't /have/ to but they were always going to, if we're going off determinism 😉

[@TheLovewhite]

yeah sooo hot!

[@rockyraccoon]

I just found your channel! Thanks so much for sharing these videos. I typically watch PBS is space-time, and love that channel as well. But so far, I haven’t found anything that lived up to the level of quality of that show. I hope you continue making these videos!

[@PierreVilleneuve88]

I like how you always have clothes with original and nice designs. Keep up the great content!

[@fancymcclean6210]

Why only a 1000 views for this excellent content? Sabine deserves more due to her lucidity of explanations. I've subscribed. Flaxen Saxon.

[@fancymcclean6210]

Just noticed that post was less than a day- might explain the low views. Sometimes I'm a silly old sausage. Flaxen Saxon.

[@gmeast]

I have noticed that way too many people spend way too much time on social network nonsense instead of on the things that will strengthen the human existence (such as gaining a clear understanding of those things helping to drive human intellectual and scientific advancement). I am concerned for our human race. Yes, Sabine seems like a grounding force.

[@dougoverhoff7568]

@@gmeast What? You mean Jerry Springer isnt mind expanding? Unfortunately, the average person isnt interested in gaining knowledge, improving their mind, or indulging in intellectual endeavors of any kind. TV, sports, movies, Face book, and other distractions are what fills their days. No wonder the human race has made so little progress at conquering hunger, war, and producing safe, comfortable environs for all mankind. In its lengthy existence, we have, however, discovered how to booty pop and to drink from a kegger upside down. Now that's progress!

[@lastchance8142]

One's interests are a function of nature + nurture. Unfortunately, the majority of people are either lacking higher IQ or were never mentored by their parents/teachers into productive pursuits. Be thankful if you are, and unselfish with your gifts.

[@Boris_Chang]

Confucious say: Man who spends too much time thinking about views and likes and subscriber counts spends not that time listening to wisdom and attempting to comprehend it.

[@horsefacehorse5702]

I didn't understand a word of that but you look awesome

[@ricardobialetti7929]

As a musician, I offer an axiom for life, "practice your mistakes and you'll get really good at them"

[@timkellermann7669]

Well there a lot of mistakes in this Video to practice on. I hope she does so, preveously i saw potential in this chanel. Not anymore

[@theastuteangler]

Deep. I like it.

[@theastuteangler]

@@timkellermann7669 practice your spelling and grammar mistakes before you criticize Sabine. And make your damn bed.

[@timkellermann7669]

@@theastuteangler Englisch is not my native language, and yes i realise that it is realy bad. That's why I don't try teaching other people english on youtube. Sabine Hossenfelder should try to do the same or at least factcheck her video (just had a problem with this one). Nobody would dare to copy my bad grammar, and that's why it is my problem if nobody understands me. Not the problem of thousands who believe everything I say on KZbin. All I am triing to do is protect people from taking wat she says in this video as Fact, because it is not. Of courese that I am a random person on the internet to you does not give my words a lot of wheight, so if you still don't think that this Video should be redone just take a lock at my other comment below and take a minute to confirm what I have written there and you will see what I mean. It was not ment as an insult to Sabine but as an atempt to stop the spread of false information about a subject I think every day about since 2016. And yes I should have made my bed a week ago but i was bussy developing a deep reinforcemnt learning algorithm to solve the sparse reward setting. Stop protecting people and start thinking for yourself!

[@timkellermann7669]

@@theastuteangler You should watch the 3 blue 1 brown video she recommended and the other videos on deep learning from 3 blue 1 brown. They are actually good and should help you see the misinformation in this video.

[@edansw]

Some mistakes: *Popular DNN have billions of neurons, and definitely not 300. *SOTA NNs are almost always not fully connected networks, and have predefined structures (e.g. Conv/LSTM nets) *SOTA NNs does separate the process into specialized networks with different purposes, e.g., pretrained image classification networks generates image embeddings.This also indicates that NNs does not start from scratch every time (i.e. transfer learning) *Real neurons have temporal activation, and also can aggregate potentials to mimc continuous activation (see Spiking neural networks )

[@HallerInfrarotGmbH]

Recorded a few ago, just a maybe. :)

[@cheponis]

1. Today's nets have tens of millions of simulated neurons in hundreds or thousands of layers or more. Agreed, nowhere near human numbers or complexity. But it was the huge number of simulated neurons that enabled the explosion in AI starting in about 2012. 2. A non-linear function such as RELU is used in artificial neurons, and, in fact, is essential; otherwise the net would be a linear function and could be modeled by one big equation. What sort of non-linearity is optimal is still unclear.

[@LKRaider]

Edmond Dantès are there NN that can loop on themselves? For example to reactivate processing after an initial prepossessing and combination with other results? Or are all NN directed graphs of connections from input to output?

[@brandonwithers5631]

@@LKRaider Recurrent neural networks can loop on themselves.

[@BlackJar72]

Well, the cortices found in the cerebrum and cerebellum do have layers, though not nearly so neat and rigid -- and the connection structure is very different. Generally a very good overview though.

[@SirCharles12357]

You are the modern day Hypatia! Love how you break down and organize you points clearly.

[@bosoerjadi2838]

You're already in my top ten of favourite KZbin channels. And I'm subscribed to several hundred.

[@xDosReis]

Bo Soerjadi same

[@PainfulRenegade]

Can you both count how many? Curious who has the biggest... (Subscribing-Counter)

[@vaclavkrpec2879]

(Sorry for very late comment.) Although I really admire Sabine's work and videos, which (to my best knowledge) are generally of very high quality, I feel compelled to correct her in a few points, here: 1/ AI doesn't equal to ANNs. ANNs are a collection of (various) models used in Machine Learning (ML). Although ML is one of the most important tools in the AI toolbox, AI is certainly not restricted to ML in general, nor ANNs in particular. In fact, conventional algorithms are used just as well (and pretty often) in cognitive technologies, e.g. in Computer Vision and decision-making systems. In short, it is written nowhere in "the textbook" that AI needs to be implemented using ANNs. 2/ There are many models of neurons (or neural layers) used in ANNs, not just one. Some of them may indeed produce continuous output, but others do produce binary outputs, i.e. "fire" or "not fire" state of the "axon". 3/ The neural network is often unidirectional, yes, but other topologies are also used/possible, including cyclic/recurrent ones (e.g. Hopfield network). 4/ Backpropagation (and its variants) is the most used learning algorithm in ANNs, but it's not the only one. The interesting thing here is that if you don't build-in the (pseudo-)Hebbian learning principle into the learning algorithm upfront, it typically "pops out" from the derivation anyway. It seems that (pseudo-)Hebbian learning is an inherent property of (A)NNs. Her general message---that AI is most unlike human/natural intelligence---is perfectly correct, though. We don't have a constructive definition of intelligence, we don't have any general AI yet, nor we're any close to creating one. Neither is AI bound to be conscious---in fact, we'd prefer it not to be (whatever it means, which we also don't know). The goal is to solve problems (as always in Informatics), not create any sort of "simulated (human) brain".

[@sdemosi]

I think she's provided an excellent introduction albeit with some statements that are useful generalisations but not strictly true. Everything you say is correct.

[@LookToWindward]

I wonder if you still share the opinion that we are very far from a general AI with stuff like Palm Pathways now coming out that can engage in long chains of reasoning, make jokes, and even explain why jokes are funny. Is it still just a "magic trick" or are we getting close to some real magic?

[@vaclavkrpec2879]

@@LookToWindward Yes, unfortunately. Of course, it's a matter of definition and we only have a behavioural definition of intelligence, not a constructive one---unless you simply choose one that works for you (like many companies in the business indeed do), but that's a bit of cheating. These examples are still ones of systems _dedicated_ to a purpose and yes, we have those and some of them work very well indeed. But general AI would have to be a system which would be able to "behave intelligently" or "show ability to intelligently decide" in _any_ situation, not just in instances of particular one. Well, I wrote "unfortunately", but there isn't anything unfortunate about it, of course. Again, we don't really need general AI... Dedicated solutions of problems in hand is what we need (and what we develop, indeed).

[@therealkellyiom]

@@vaclavkrpec2879 Are you in favour of the bootstrapping approach where we simulate the brains of 'simple' creatures and gradually increase complexity? Because we are human, we're imperfect so would we actually need to manage the risk of an AI developing depression for example? While we're nowhere near this problem, I do think our lawmakers should anticipate development of AGI and who ultimately takes responsibility. I could see it being used by some as a way to live forever in a virtual world. Definitely not me though :D

[@vaclavkrpec2879]

@@therealkellyiom I think it's an interesting _academic_ endeavour, but I'm not sure how practical it is. As noted before, what we use AI for is solving actual problems, not imitate nature (although we inspire ourselves in it a lot; after all, ANNs and e.g. generic algorithms are clear infringment of nature's copyright ;-)). Creation of autonomous "brain" immediately yields the problem of how to make it do what we need... This indeed raises the question of ethical issues---and you're right, if one day we will be able to create "artificial brain" capable of autonomous decision making and/or even exhibiting consciousness, regulation shall be necessary. The question is whether we ever will, not necessarily because we couldn't, but simply because we won't see any point in doing so. We don't do everything we can do, just because we can---that's silly....

[@ezrawilson6986]

Your videos are getting more views, Dr. Hossenfelder. I'm glad to see that.

[@theklaus7436]

You have become one of my favourites and I can see I’m not the only person who apparently seems to agree. Great explanation, and especially not afraid to say what issues the physics community has. I think it’s a wonderful time for everyone but especially young people

[@Macroleverinc]

This is a wonderful explanation. Seems that we humans have a special capacity to generate an abstraction. Its like the human brain at some point replaces or expands a particular neural network with an "abstract network" that is able to calculate an invariant far more faster than a computer. That could be why humans are able to identify the turtle even if it moves rapidly.

[@camerontaylor7471]

It’s not faster... it’s simultaneously... that’s what quantum computers do so efficiently....

[@projectmagnet]

Love your channel! When studying the human brain, I believe the concept of consciousness and awareness is also a factor, which if followed through, shows that even memory may not be local to the physical body at all. We have to consider out of body consciousness, and then begin to determine how the conscious field synchronizes with the physical body through the brain, rather then pretend that the consciousness is always trapped inside the brain as if it originated there. A person can be knocked out, things go black, but then after a time of synchronization the human becomes conscious again. This is not possible with a computer, if you struck it with a blow, it will either not miss a beat or it will be damaged for good and never recover. This indicates to me, that the soul or spirit links to the physical body awareness through the brain using a vibration path. Add to this data the fact that so many people have out of body experiences, and you can then begin to make more meaningful experiments and observations. The human brain is not like a digital calculator, it is more like a human to spirit interface tool. Whatever we truly are, is not limited to the physical body.

[@jonathanrobertson3406]

I'm starting to seriously like this channel. Well done.

[@mithamitha8457]

Artificial intelligence is a very big field with examples like the navigation computer or an old school chess computer. Within that field you find the machine learning area and within that field you find the neural network technology. In the beginning we used some theories from biological neurons to get started with software that can learn patterns by itself. As soon as we had fast computers we began by trial and error to see what could be done on a computer. How a task was done by a human brain was of no concern at that point any more. By now these two things have little in common because people did not intend to make them alike. This is important to understand the differences between the structure of artificial neural networks and the structure of biological neural networks. And yes, we have many applications that require tons of training material. But this is not the case for all artificial neural networks. There are examples for self learning systems that do not require such training material, like the systems that learned to play Atari games or Starcraft or Go. There are even first prototypes for object recognition that learn practically by showing them fruits by hand in front of a camera. This is at least an important point in understanding the differences between artificial intelligence and human intelligence. This new systems are a good deal more advanced than plain old feedforward neural networks (which are not "current approaches to artificial intelligence" as suggested in the video's description but are rather old concepts from the 1980s).

[@jeanmariedr]

Thank you for all your so clear explanations on such difficult subjects (I am electronic engineer , retired, and you help me to stay rather informed). On the other way I find your approach on the relation between physics and beauty very interesting and bringing some "fresh air": thank you again. And ... I like your songs !!!

[@susanne5803]

That was a very helpful basic explanation of artificial intelligence for someone who had to look up the meaning of "loss function". I think the brains evolutionary pressures made it to approximate danger and success as good as barely necessary while avoiding cost explosion. It was made to learn - but always only barely good enough for many - not for all. Stanislas Dehaene is a good read, I suggest his book about numbers and how our brains deal with them. Thank you very much!

[@Old299dfk]

Why has nobody mentioned how good that outfit looks?

[@mundusuys8739]

I like it too. The oufit matches the topic. I've noticed this in some of her other videos too and I appreciate it.

[@hooya27]

I like to think she has only that on, and nothing else. Also, check out the "is faster-than-light travel possible?" - omg!

[@ronaldthomas7523]

Intelligent sexiness.

[@dt7843]

Because we have been made to love etch other. I love here for ever. Sabine love me like crazy 🤷🏽‍♂️ Keep hoping dude . 🤷🏽‍♂️

[@BBQDad463]

Because---wait for it---she blinded us with Science! 😎💙💛

[@MythAdvocate]

You are so well spoken, brilliant and concise. I am an instant fan.

[@scudder991]

Dr. Hossenfelder, i confess to being smitten by your intellect, ease with communicating complex science, and elegance.

[@toddtrimble2555]

Very sensible suggestion at the end, that neural networks and AI are better directed toward tasks that humans are manifestly not good at, rather than trying to get them to simulate human intelligence (which could be a fool's errand). Great summary.

[@andygoris6667]

Awesome video! Two other things: the brain also has chemical mechanisms (larger scale) for communicating with itself, the body, and from the body. Also many areas of the brain have bi-directional pathways. Areas can control what each other does. Simple example: consciously paying attention to one conversation of 10 you are hearing in a crowded room.

[@bernhardjanssen9284]

You have open a new dimension i learn each time more thank you so much for this !

[@docsmithdc]

Once again a wonderful lecture video.Had my physics instructor been anything like this I may have gone into physics instead of medicine.Oh...and by the way..Sabine you look absolutely ravishing .

[@vdrlng]

Intelligent and deadly; she is a Bond villain with a PhD) 🍷🗞️

[@briankamras2913]

It feels good to hear someone describe ML and NN in a simple way that describes it's shortcomings. As an engineer working on an ML project I've been utterly disappointed at ML tooling, specifically tensorflow.

[@pwnmeisterage]

As a tinkerer working with NN, I was surprised at how much human supervision is needed to train/learn towards objectives. And at how dismally a clever AI performs when that learning is applied towards an unfamiliar domain.

[@jonasfelix7700]

The sentence „...no one really knows how the human brain learns, but that’s not how...“ is the key question to everything.

[@ogungou9]

kzbin.info/www/bejne/g3PFnJ2iiaiLqpo&feature=emb_logo

[@breearbor4275]

i had a philosophy professor in undergrad who did a lot of research on artificial intelligence and neural networks. these are some nuggets of wisdom he gave us: 1. people always say we're only "20 years away" from "true" AI (AI that is fully conscious or indistinguishable from humans). but the problem is, they've been saying that for more than 20 years. we're really nowhere close, and it probably isn't possible to replicate human intelligence on the kinds of computers we have now. 2. people often worry that rogue AI will turn against us in some sort of sci-fi scenario. the simplest reason why this isn't the case (aside from how far away we are from creating AI that can think for itself) is that all the people developing AI - the military, security institutions. etc. - are people that love to have absolute control over everyone working for them. they would never create something that wouldn't obey orders. (the same goes for private corporations imo). we should be worried about people in power, not machines themselves. 3. the example he would use to demonstrate the point Sabine is making in this video was just pointing out how humans and computers deal with simple computations. give a computer a complex math problem and it will pop out the answer instantly. computers are ALREADY better than humans at most things they do, hence why we use them. but a calculator obviously doesn't pass the Turing test. the way computers solve problems is so completely different from humans, it's wrong to assume that as computer technology improves they will somehow beat us at our own game.

[@stevennoel6724]

Rapidly becoming hooked..at least to the funky fashions, as well as to the illuminating lectures

[@hdgehog6]

Sabine is so refreshing to listen too... Puts things in proper perspective.....

[@rajarsi6438]

Not really, she only has the management level of thinking.

[@thk4711]

Good video about a topic which is discussed a lot at the moment often by people that have no knowledge about the topic. I think it will be the same with AI than with many other new technologies . At the beginning people overestimate what it can do but in the long run they underestimate how much it will change the way we are living.

[@GEOFERET]

How do you manage to explain things so simply, without oversimplifying, but sticking to the basic and important aspects of each phenomenon? Very few people have this ability. Also, very perceptive that about what we really want artificial intelligence to be. Thank you.

[@toddwasson3355]

Great video, Dr. Hossenfelder. Another (easier) way of training neural nets besides back propagation is to use a genetic algorithm for selecting weights. You start with a random population, "mate" the most successful ones with each other (meaning manipulating the link weights Sabine described), throw in a random mutation here and there, then generate a new population for the next generation. Rinse, wash, repeat. These are really neat to watch evolve because they're designed to mimic evolution. An even easier way is to use a particle swarm algorithm to perturb the link weights. I've used both successfully for creating AI driven cars for vehicle simulation. I've seen them "learn" to drive simulated cars that are set up so poorly a human can't drive them. For the AI it's no problem.

[@gregor-samsa]

In fact Genetic Algos are Another Concept of sub Symbolic Data Processing. Sone more Concepts e.g Rule Engines, Data from senses, thrown in and the result looks like ai and has a Complete different meaning in German Language. Sorry others are Great but this Talk is nonsene.

[@superfluousification]

More people need to watch your videos. I just discovered you tonight and it is so refreshing to hear facts without woo

[@ultrametric9317]

Thanks Sabine! Looking good!

[@scottk6659]

Youre one of those great teachers who can get students interested, and explain rhinks in a simple and clear way Thank you for sharing your knowledge . In that sense you remind me of Richard Feynman who could hold peoples attention for hours with his interesting lectures. Do you lecture or teach anywhere Sabine? If so, lucky students

[@karte2626]

Exactly!!! Machines do unpleasant or dangerous tasks that humans hate or avoid to do!!!

[@markrenton1093]

Sabine, I love your staccato voice, very clear, descissive .

[@vast634]

In the youtube edutainment sector you are supposed to have poppy background-music, wear a funny t-shirt and crack jokes.

[@tidyyy]

looking fantastic, sabine! :)

[@vdrlng]

Tidyy > Yes, Sabine is a sexy beast 👠 ; with a serious intellect) She's off the hook) []. .. .....📞

[@lastchance8142]

Calm down.She's married

[@bitcodexx]

@@lastchance8142 But she has no ring. My neural network needs to analyze this.

[@user-il5hy5tr5z]

bitcodexx y’all strange💀

[@KarlBunker]

@@lastchance8142 >Calm down.She's married 😢

[@davew4998]

Excellent video Sabine. I was turned off AI over 40 years ago when my over enthusiastic computer science lecturers kept banging on about computers 'thinking'. I was always afronted by this, which seemed arrogant beyond belief, totally ignoring any philosophical discussions on dualism. Your video has provided a stark contrast between the brain and AI computer algorithms. It's actually made me more interested in AI than my useless lecturers did.

[@davew4998]

@The Great Bodhisattva Hachiman That's very true. I can barely touch my knees now, let alone my toes.

[@pkr619]

This video makes it easy to understands the differences between the artificial and the natural NNs. It is a good primer on neural nets as well. Thank you for putting time and effort into creating such informative content.

[@ChrisBrengel]

Great video, thanks Sabine!

[@ambujkumar7737]

This channel has one of the best way to drliver knowledge👍

[@happygimp0]

As an electrical engineer, i have to correct you on some points: - To point 1,4,5: Neural networks does not have to implemented in software. It is possible to use an ASIC or FPGA for it or even discrete transistors. Which means the functions and structure. A ASIC would also be much more power efficient than software on a GPU or other processor that supports massive SIMD which is more power efficient than a normal CPU. A ASIC, FPGA or a Processor with SIMD can calculate neuronal networks in parallel, the same is true for multiple CPUs cores but less efficient. You wont get near the power efficiency of a brain with current transistor sizes. - To point 3: You do not need layers in a neural network, it just makes it much easier to implement and we can understand it better.

[@Feyd01]

Such an intelligent and attractive lady.

[@billyt8868]

just from the thumbnail i already like where this is going.

[@geoffreystuttle8080]

Perhaps because Sabine is so appealing, the technical information is easy comprehended..? I think so.

[@CaptainJeoy]

Yeah! It's kinda always titillating for me when I watch her videos and how she explains rather abstruse subjects!

[@QContinuuum]

Sabine, I really appreciate your videos a lot and I have learned and enjoyed most of them. However, this seems not to be your best one. I suggest you watch some of Robert Miles' tubes on AI safety, and if you want some pointers to enter into the youtube search field, here they are: 1) Computerphile deadly truth of general ai (--> robert miles stamp collector), 2) robert miles instrumental convergence, 3) robert miles orthogonality. But actually, all of them are really great!

[@LouisGedo]

2:31 - 2:33 This needs to be a poster in your online store!

[@yellowNred]

Louis Gedo I second that. Please make it available.

[@codehustler]

either this is utterly outdated, or was not well researched in 2019 already. a) We build MLPs with much larger Neuron counts than 300, there are nets with more than millions of neurons. Even when I wrote my on MLP back in 2008, I was running more than 300 virtual neurons. b) Virtual neurons do not have to fire continuously, they may, depending on the architecture, also implement an activation function. c) Even back in 2019, we already had deep neural nets which have a much more complex architecture than just input-hidden-output layer. Oversimplified, deep nets are a bunch of different net architectures trained for different purposes interconnected, similar to the different regions of the natural brain. Means, that deep nets come with a pre-defined architecture such as our brains do.

[@Car_Mo]

Thanks for another brilliant, eye opening learning experience! I've mostly thought of what's passed of as "Artificial Intelligence" today as "Virtual Intelligence". Proponents are so eager to present machines that appear and interact as "human like" as possible, as if the more it looks and talks like a human the more intelligent it is. There needs to be more focus of what Sabine talks about in the video, self learning machines that can eventually become "intelligent" enough to improve themselves. It's harder than making talking heads, but infinitely more useful.

[@pierretetreau7497]

Sabine your videos are truly with the best of the best on youtube. Thanks

[@jonn7291]

You look great in this one. You always look great. Really interesting topic as well.

[@Squbber]

1. Outfit on point. 2. Educational description of the mechanics of artificially intelligent neural networks on fleek.

[@FutureChaosTV]

Machine Intelligence is hyper specialized (due to the low amount of high performance neurons) whereas Animal/Human Intelligence is highly generalized (due to an enormous abundance of slow neurons)

[@NikolaosSkordilis]

Neuroscientists believe that the massive interconnectivity of the human brain, i.e. the number of synapses between the neurons, is probably more important than the raw number of neurons. These ~100 billion neurons are inteconnected by up to ~350 trillion synapses in an adult human brain, assuming an average 3,500 synapses per neuron (the neocortex, amazingly, has up to 7,000 synapses per neuron).

[@FutureChaosTV]

@@NikolaosSkordilis Yes, you are right. The "intelligence" lies in the interconnects. Computing + Storage + Network in one slowly pulsating "mesh".

[@myothersoul1953]

I think that is an important distinction but how generalizable intelligences are can be overstated. G factor intelligence might be such a case, it could certain broad mental abilities are independent of each other.

[@MrCmon113]

I don't see how that follows. Also a fast processor can simulate multiple parallel processors and when a task is parallelizable multiple processors can simulate a fast one.

[@tentative_flora2690]

You did simplify quite a bit. But I don't see this as oversimplification. You seem to have effectively communicated the key points. And I say good work.

[@Steve_V1066]

I am so glad I've found you! I find your style easy to comprehend and really gain a better understanding of what you are covering. Thank you for what you do!

[@latentorder]

I feel like I should mention that your estimate for number of neurons in current neural nets is probably going to give people the wrong impression. Tiny neural networks (for instance those running on our phones) may be in the several hundred neuron range, but the big modern nets which make more sense for comparison to the brain are vastly larger. There's no clear mapping between human neurons and neurons in an ANN, and more complicated nets (eg conv, transformers) makes comparing the number of neurons rather difficult, but for instance GPT-2 had 1.5B parameters when you released this video, GPT-3 has 175B parameters, and Google released a mixture of experts model with 1T parameters. This translates to neuron count at least in the millions.

[@tandemungwa4884]

Sabine, I'd love to get your thoughts on consciousness. I know this isn't your particular field of study (much further from it than AI, I imagine), but I think I and many people find it fascinating and mysterious, especially due to the implications on materialism.

[@SabineHossenfelder]

I'll be attending a consciousness workshop next month! Hope to have something useful to say after that :)

[@Filemonefly9]

@@SabineHossenfelder Have a feeling you might just decide to fly in on a Dragon to that workshop?.. 😅

[@keithprice475]

I find it fascinating that AI research tends to treat consciousness as a kind of fancy 'add-on' that is somehow not integral to intelligence! Whereas I regard it as the primary reality without which no real intelligence can exist.

[@dennismertens990]

@Ralph Glass Could you better define what you mean by "intuition of oneself"?

[@dennismertens990]

@Ralph Glass hello there. I am inclined to support your thought on how conciousness may have emerged as an evolutionary advantage. Perhaps not a direct effect, but an emergend property of intelligence. The only issue is that, for me at least, there is no clear meaning to neither "selfaware conciousness" nor "intuition of oneself". If there is no way to describe it punctually, then there is no way to work with it.

[@eriktempelman2097]

Stunningly good explanation of one of the key topics of today. Thanks.

[@donaldhysa4836]

That AI Car be like "Does this count as moving constantly at a high speed? Weeeeeeeeeeeeeeeeeeeeeeeee!"

[@juanausensi499]

If the computer does exactly what you told, any problem that arises is your fault.

[@funquay2219]

Maybe the AI car watched too much Top Gear and thought that you are supposed to drive high performance cars round and round in circles!

[@funquay2219]

Maybe the AI car watched too much Top Gear and thought that you are supposed to drive high performance cars round and round in circles!

[@djgenetic111]

@@juanausensi499 most of these AI approaches rely on random processes, so the idea is exactly that the computer does not follow any strict algorithm. Which also means that you cannot explain how exactly an AI system reaches a decision.

[@amihart9269]

It's a problem with a bad error function, but arguably humans have an error function as well. Your brain makes predictions, and based on the prediction, it either brings pleasure or displeasure, and the neural network can adjust accordingly. A person can be born with a bad error function, causing them to either lack any sort of emotional response to things, or even be able to interpret things like pain, and this causes them to also do things as ridiculous as driving around in a circle. Children who can't interpret pain can sometimes claw their own eyes out not understanding what they're doing. Human brains sort of have a built-in error function that we evolved. The programmer's error function they defined for the car probably allowed for driving in circles to produce very low error, so the car is just responding to that error function. It is learning correctly, but according to a wrong error function.

[@OL9245]

First of all, thank you for your channel. I have great admiration for the depth of your thoughts on physics and science and it is always a delight to watch your videos. Having said that, I find the differences you state between artificial intelligence and human intelligence to be superficial. Many are unimportant technicalities. For example, the fact that artificial intelligence is sequential and not parallel is clearly a technical limitation of little importance. Conceptually, artificial intelligence algorithms have always been parallel. Historically, their implementation in parallel machines was not possible. This impossibility is gradually being overcome without the algorithms being radically challenged. I hope I am not saying something stupid, but I even believe that one of the differences you mention is erroneous, or else I did not understand your explanations : You say that artificial neurons are activated gradually, but this is not true. Like natural neurons, artificial neurons discharge, or not, according to the crossing of a predetermined activation threshold, the value of which being one of the parameters of each neuron. There is therefore no continuity in the functioning of artificial neurons, at least in the sense that I understood in what you are saying. You also say that we like computers for what they are capable of doing non-human. Again, this quite debatable. The bulk of artificial intelligence research focuses instead on typically human capabilities; ideally, driving a car at least as well as a human can do: that's the grail. The ability to drive a car has been theorized as the ability to make good and stable decisions in environments from which the information we get is limited and changes in real time. The experimental model of real-time limited information that is currently being explored is video games, and that is why, after managing to beat humans at go, DeepMind is now looking to beat them at Minecraft. Today, they are capable of defeating good players, but not yet the best in the world. What incredible progress! On the other hand, I think you missed the big questions of the discipline about the differences between man and machine. Two of the great questions of today, which pose theoretical and even philosophical problems, are what we can call common sense on the one hand and morality on the other hand. Some people think that these abilities are mandatory to safely drive a car in the city, but the question is open. A clue is that today's artificial intelligence does not have these capacities of common sense and morals and yet is almost already able to drive cars in the city almost correctly. Will this ability be indispensable in the long run or not? And how can we get there ? If we ever can in the first place. We don't have the answer yet. Common sense is the ability to learn by oneself the most obvious things by interacting with one's environment. For example the intuitive sense of physics such as the laws of gravity, inertia, and that a falling object will necessarily stop when it hits the ground. Until now, we do not know how to equip an artificial intelligence with this capacity to learn common sense and this remains an open theoretical question: are current neural networks a sufficient architecture to acquire a common sense, and if so, should we simply wait until we can build bigger and more sophisticated neural networks? Or is it, as you say, an architecture problem? This question is not clear-cut. Another open question is that of moral sense, i.e. the ability of an intelligence to discern in a general way things that are good from those that are bad. Philosophy of morality is a very old branch of philosophy, but it is nowadays experiencing a revival of activity in the community of philosophers, precisely because this question has received a new light thanks to the problem of artificial intelligence. Their work is fascinating. To sum up, I find that you have emphasized unimportant differences, some of which are not even relevant anymore since the publication of your video, although you have missed the most fundamental differences, those that even question us the most about our human nature. I've learned a lot from all this by watching Robert Miles' videos. He has a channel named after him and he has also made a lot of videos on the computerphyle channel. He is a great specialist and everything he says on the subject is absolutely fascinating.

[@parsleyjoe1]

Artificial intelligence ≠ neural networks

[@user-mm8pm7ol3r]

It is though, in practice.

[@parsleyjoe1]

@@user-mm8pm7ol3r No. Neural networks can be used to create AI systems but they are just a component of the plethora of AI applications, related mostly to learning and perception of software agents. But intelligence is made also of rigorous logical and mathematical thinking, efficient communication, cooperation, knowledge representation and inference. These things cannot be done with ANNs.

[@user-mm8pm7ol3r]

@@parsleyjoe1 Do you actually work in the field or is your view based on your philosophy about what intelligence is? I don't work in the field but I have read a couple of actual textbooks written by actual practitioners, as well as several studies in application to my sphere of interest (finance). Everywhere I looked, AI seemed to be synonymous with ANNs. This may change in the future, but that is the state of the craft today.

[@parsleyjoe1]

@@user-mm8pm7ol3r I actually do study and work in the field. AI has been historically a lot of things (I named some at the end of my previous comment) and only recently ANN technology made some breakthroughs, thus the public opinion thinks that AI and ANNs are synonyms, but they are not. For example, in many cases of AI applications, software systems cannot afford the luxury of approximation and for this reason ANNs are not adequate as part of the solution

[@manit77]

@@user-mm8pm7ol3r current neural nets will never gain conscienceness. It's currently a sophisticated statistical algorithm for determining out.

[@att.6134]

@Sabine: Factually this video is still mostly correct, especially given this short video length. Then again, other parts didn’t age optimally. Good that you made new videos :-)

[@crisprtalk6963]

"Warm, wet and wiggly". I love science.

[@moriscnam]

Brief but informative. Thanks.

[@PhokenKuul]

Not to detract from the highly interesting and intelligent monologue here, nor to attempt to objectify you in any untoward or unsavory way, but you are rocking that outfit.

[@mqcapps]

She has two kids mit mann... PAY ATTENTION to the info

[@PhokenKuul]

@@mqcapps Not sure what that has to do with her looking great in that outfit. Are you saying women who have kids can't look great? Or are you saying women in relationships can't look great? Are you saying highly intelligent women can't look great? Are you saying women giving interesting, fact filled talks can't look great?

[@mqcapps]

@@PhokenKuul she is awesome and for the FIRST time the topic is explained such that i understand it ...if she wants to explain this in jeans or a lab smock fine with me...

[@PhokenKuul]

@@mqcapps No one would argue that what she is talking about is interesting and well explained. Nor would anyone argue that she is free to wear whatever she wants. Are you arguing that she doesn't look great in that outfit? Or are you arguing that I have no inherent right to let her know she looks great? Or are you arguing that there is no value in letting someone know when they look great?

[@mqcapps]

Her looks ...how does SHE feel ...

[@madscientistshusta]

So glad I found the channel, invaluable resources for those of us wanting to dip our toes in these fascinating subjects!

[@markspqr]

There is an old book I recommend everyone read "Mechanism of Mind" by Edward DeBono the inventor of lateral thinking. It explains the brain in an amazing way that has been constantly proven by Neuroscience since the 1960's he even predicted neurotransmitters

[@theastuteangler]

He invented lateral thinking as much as Newton invented gravity

[@altrag]

Missing one very important thing that ties into a lot of the bullet points: Parallelization. Yes, a neuron may fire millions of times slower than a computer can perform an operation, but real neurons can all be processing at the same time. Even our most parallel supercomputers only have a couple thousand nodes that can run simultaneously. So that's an obvious impact on things like speed and size, but it also has impacts on things like precision -- the fact that ANNs require precise calculations to propagate values (and back-propagate weights) means even in a highly parallel computing environment there will be a significant synchronization overhead. The real brain gets away without needing (as much) of that by simply being able to handle signalling errors in a way that (mostly) works out. Overall though, the most critical issue (I suspect) is "size." Having a million times more neurons simply makes the human brain "better" at everything that we consider humans to be good at, and its just way too early to know how an ANN with 100,000 or 100 million neurons will be able to evolve itself. Just to take the most basic example, if we take the few dozen neurons used by voice recognition and the few dozen from each of say 10,000 image classifiers and a few hundred from self-driving car designs and a few dozen from every video game that uses ANNs and so on and stick them all in the same box well.. we've just invented a "brain" that has specialized "regions" to do multiple tasks, and that's without putting any effort into the kind of organizational complexity that the human brain comes with. And to make matters worse, we don't really know all that much about how the human brain operates either. For all we know its simply a massively parallel function optimizer as well, and we just don't yet have sufficient understanding of whatever quadrillion-dimensional function its trying to optimize.

[@kimwelch4652]

The classic example of Neural Network failure is the "Tank in the Woods" problem. The military tried to teach a Neural Network to identify tanks hiding in the woods. After training the network they ran it through tests which it promptly failed. Eventually they discovered that all the training pictures with tanks were taken on an overcast day while all the training pictures with no tanks were taken on a sunny day. The test picture was a sunny day with tanks so the network failed assuming sunny days have no tanks. Humans sometimes exhibit this same problem. You never know what your kids are really learning.

[@marsrocket]

That’s an urban legend. Not to say it’s impossible, but you’d have to be a spectacularly bad scientist to make that kind of mistake. www.gwern.net/Tanks

[@MrCmon113]

That doesn't have anything to do with neural networks in particular, that's about statistics in general.

[@kimwelch4652]

@@MrCmon113 It's not about statistics; it's about how humans interpret statistics. The numbers are the number; the tank pictures are what they are. We interpret these things according to neural network pattern matching which doesn't always work in an easily predictable way.

[@MrCmon113]

@@kimwelch4652 How you interpret things is what statistics is all about. And to anticipate selection and observer effects as the above is part of the job of a statistician.

[@tomcan48]

*The MAIN difference between AI and human intelligence is simple ONE, conscience. A second consideration of the this is that a human brain is composed of TWO hemispheres that enable comparisons the two independent/dependent conscious perspectives that enables both artistic and analytical semantics.*

[@neptasur]

If "intelligence" means connecting data points . . . And "consciousness" is something else entirely.

[@bozo5632]

They ARE entirely different things. You could have a lot of either one without the other.

[@jumperpence]

Recently found your channel, and working my way through the back catalogue. Excellent content, and may I say you look great!

[@GnosisMan50]

She looks very attractive with her dress

[@rajarsi6438]

Animalistic lust, many are completely imprisoned by it. Very funny.

[@xenonchikmaxxx]

Thanx for the last section explaining that humanity really doesn't need mimicking of humans.

[@kdeuler]

interesting. i'd like you hear your thoughts on "quantum life", reviewing claims that quantum effects govern some life processes.

[@alinaqvi385]

Very good and simple explanation! Thank you.

[@catman8965]

I think Sabine is too intelligent for most KZbinrs.

[@catman8965]

faultroy Well, YES there are many intelligent KZbinrs, but there are many that can't think past a government conspiracy theory. Sabine presents her points with well thought out logic above what most people (including myself) would normally think.

[@rclrd1]

@faultroy I think he was just referring to all the dumb remarks in the comments section of _every_ KZbin video.

[@pboston6RR]

Your parting comment about using neural networks to do things we are not good at, is profound! I can drive a car and make good, smooth changes during traffic, but my Tesla Autopilot is MUCH better at keeping the driving lanes path centered on the lane, maintaining the speed within posted limits, seeing and responding to traffic lights and stop signs. That’s why I sue it. Thanks for another very enjoyable and thought provoking presentation.

[@georgekhumalo5283]

PhD level content as always, Love your work Sabine. Please cover more AI topics in future.

[@dff1286]

I love watching your videos. always refreshing and always informative.

[@StuMas]

Another huge difference is: AI is not Self-Originating. Unlike Humans, AI requires an Intelligent Mind to "create' it. ...alegedly.

[@yoooyoyooo]

For now

[@StuMas]

@@yoooyoyooo Well, unless they manage to 'create' themselves spontaneously, somewhere in isolation (away from Mankind)... I'd still say, No.

[@TheRainHarvester]

Hi Stu, and Yoyoyo, regarding the first "thing" to get it started: EXACTLY! I think you both may really like some videos I'm making. I'd like to invite you to my channel. Search my channel "heartbeats & blood flow". It's about emergent behaviour in a simulation. Stop by my bio-digital lab and say hi!

[@trinityrogue5745]

With total RESPECT I SAY " HER INTELLIGENCES , BEAUTY AND CONFIDENCE IS AWESOME

[@1MarkKeller]

First I thought she was Valeria Richard's, but now she's wearing an X-men uniform.

[@Roboprogs]

Or perhaps something a female guest would have worn on the original 60s Star Trek.

[@ernestmoney7252]

@@Roboprogs No, in Star Trek she would have shown more skin. Google: The Top 50 Babes in Star Trek Original Series

[@rickstone9092]

The brain is truly amazing. Let's me know that things in the Universe are 'by design' and not random. What people do with their intellect may be random, but most things in the universe are by design.

[@thearchitect5405]

While the architecture for AI is man-made, the neural network is not. It is largely self-taught, so in other words, it's random. If it works for simulated evolution, then it works for actual evolution as well.

[@soultrick7474]

...And on top of all she has a great style, love it!

[@sirWinningAtLife]

This is a very overlooked viewpoint in the artificial intelligence and technological singularity debate.

[@headcrack]

The typical 300 Neurons per artificial Networks are the past. A modern Pohoiki Springs Setup consisting of 768 Intel Loihi Chips already has 100M Neurons, 1/10 of the human brain. Even when these chips started to sell only to universities with the beginning of 2019, that numeric barrier between human and AI can be considered broken now. en.wikichip.org/wiki/intel/loihi#Pohoiki_Springs_.28768_chip.2C_100M_neurons.29 Second the human brain does not function as a common ghost like Sokrates described it. The human brain jumps every seconds between different parts of it. A German philosopher Richard David Precht had its breakthru years ago with his book "Wer Bin Ich Und Wenn Ja Wieviele". Daniel Kahneman wrote a book "Thinking, Fast and Slow", which had a great impact on Machine Learning, systems evolved, consisting of thousand networks of smaller networks, called Reinforcement Learning. An impressive Example is IBMs Project Debater, when a AI firstly outperformed a human national champion debater within this year. kzbin.info/www/bejne/o2TYXmSvqdmljNk

[@SabineHossenfelder]

I wasn't talking about specifically designed chips because at least so far they're not general purpose machines. But, yes, topic for another video :)

[@headcrack]

​@@SabineHossenfelder Personally, I consider Reinforcement Learning Systems as a footstep to general purpose machines, since our prefrontal cortex is no more - www.nature.com/articles/s41593-018-0147-8 Their difference is the (maybe still) missing ability for operant conditioning, not only to reward useful behaviour, but to punish unuseful ones. Still, the big challenge is to differentiate between useful and unuseful, Short-Long- And Mid-Term Memory would be nessecary for that. But there is another significant difference between human and AI NN: Signal Propagation Delay, nonexistent in artificial NN. And even our human brain relys on networks of maaaaany clusters of only hundreds of neurons: www.mpg.de/10655144/neurons-form-synapse-clusters. Each of us is not only one;-) Given, NN only do only work up to (20%+?) inner connections, for the human brain the average is 3000, the max is 10k, for my understanding we can predict the limit of (useful) cluster size with an very old formula n=n*n^n+1 : only some hundred... Please understand, I am obsessed by that topic, being an it senior heavily adhd affected by myself : Try to imagine a network of networks, when connections unpredictable break down. I totally like the way you talk.

[@owndoc]

The brain has 100 billion neurons. 100M is 1/1000th of that.

[@seancidy6008]

So enlightening and no-nonsense.