This is a talk by Ray Kurzweil for course 6.S099: Artificial General Intelligence. For this entire recording, Ray did not use slides, so the video does not show any slides. This class is free and open to everyone. Our goal is to take an engineering approach to exploring possible paths toward building human-level intelligence for a better world.

INFO:
Course website: https://agi.mit.edu
AI podcast: https://lexfridman.com/ai

CONNECT:
- If you enjoyed this video, please subscribe to this channel.
- AI Podcast: https://lexfridman.com/ai/
- Show your support: https://www.patreon.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Twitter: https://twitter.com/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman

Welcome to MIT course 6S099, Artificial General Intelligence.

He is 1 of the world's leading inventors, thinkers, and futurists, with a 30-year track record of accurate predictions.

Called the restless genius by the Wall Street Journal and the ultimate thinking machine by Forbes Magazine.

He was selected as 1 of the top entrepreneurs by Inc.

Magazine, which described him as the rightful heir to Thomas Edison.

PBS selected him as 1 of the 16 revolutionaries who made America.

Ray was the principal investigator of the first CCD flatbed scanner, the first omni font optical character recognition, the first point-to-speech reading machines for the blind, the first text-to-speech synthesizer, the first music synthesizer capable of recreating the grand piano and other orchestral instruments, and the first commercially marketed large vocabulary speech recognition.

Among Ray's many honors, he received a Grammy Award for Outstanding Achievements in Music Technology.

He's the recipient of the National Medal of Technology, was inducted into the National Inventors Hall of Fame, holds 21 honorary doctorates and honors from 3 US presidents.

Ray has written 5 national best-selling books, including the New York Times bestsellers, The Singularity is Near from 2005, and How to Create a Mind from 2012.

He's co-founder and Chancellor of Singularity University, and a Director of Engineering at Google, heading up a team developing machine intelligence and natural language understanding.

I've been in this lecture hall many times and walked the infinite corridor.

Within a year of my being here, they started a new major called computer science.

Even biotechnology recently got its own course number.

Okay, how many of you do work in deep learning?

How many of you have heard of deep learning?

I became excited about artificial intelligence.

It had only gotten its name 6 years earlier, the 1956 Dartmouth Conference by Marvin Minsky and John McCarthy.

He spent all day with me as if he had nothing else to do.

The AI field had already bifurcated into 2 warring camps.

The symbolic school, which Minsky was associated with, and the Connectionist school was not widely known.

In fact, I think it's still not widely known that Minsky actually invented the neural net in

but he had become negative about it, largely because there was a lot of hype that these giant brains could solve any problem.

So the first popular neural nets, a perceptron, was being promulgated by Frank Rosenblatt at Cornell.

So Minsky said, oh, where are you going now?

And I went there and Rosenblatt was touting the Perceptron that it ultimately would be able to solve any problem.

So I brought some printed letters that had the camera and it did a perfect job of recognizing them as long as they were carrier

And he said, but don't worry, we can take the output of the perceptron and feed it as the input to another perceptron and take the output of that and feed it to a third layer and as we add more layers it'll get smarter and smarter and generalized.

And I said that's interesting, have you tried that?

Well, no, but it's high on our research agenda.

Things did not move quite as quickly back then as they do now, he died 9 years later.

Never having tried that idea, turns out to be remarkably prescient.

I mean, he never tried multi-layer neural nets and all the excitement we see now about deep learning comes from a combination of 2 things.

Many-layer neural nets and the law of accelerating returns, which I'll get to a little bit later, which is basically the exponential growth of computing so that we can run these massive nets and handle massive amounts of data.

It would be decades before that idea was tried.

Several decades later, 3 level neural nets were tried.

They could deal with multiple type styles.

Now it's not hard to add other layers, it's a very straightforward concept.

There was a math problem, the disappearing gradient or the exploding gradient, which I'm sure many of you are familiar with.

Basically you need to take maximum advantage of the range of values in the gradients, not let them explode or disappear and lose the resolution.

That's a fairly straightforward mathematical transformation.

With that insight, we could now go to 100-layer neural nets.

And that's behind sort of all the fantastic gains that we've seen recently.

AlphaGo trained on every online game and then became a fair Go player.

It then trained itself by playing itself and sword passed the best human.

AlphaGo 0 started with no human input at all.

Within hours of iteration, SwordPass AlphaGo also SwordPass the best chess programs.

Basically, you need to evaluate the quality of the board at each point.

They used another 100 layer neural net to do that evaluation.

So, there's still a problem in the field, which is, there's a motto that life begins at a billion examples.

1 of the reasons I'm at Google is we have a billion examples.

For example, there's pictures of dogs and cats that are labeled.

So you got a picture of a cat and a cat and then you can learn from it.

AlphaGo trained on a million online moves.

And that only created a sort of fair Go player.

So they worked around that in the case of Go by basically generating an infinite amount of data by having the system play itself.

Had a chat with Demis Hassabis, what kind of situations can you do that with?

You have to have some way of simulating the world.

So Go or chess are, even though Go is considered a difficult game, it's a, you know, the definition of it can exist on 1 page, so you can simulate it.

I mean math axioms are, can be contained on a page or 2.

It gets more difficult when you have real life situations like biology.

So we have biological simulators, but the simulators aren't perfect.

So learning from the simulators will only be as good as the simulators.

That's actually the key to being able to do deep learning on biology.

Autonomous vehicles, you need real life data.

So the Waymo systems have gone 3 and a half million miles.

That's enough data to then create a very good simulator.

So the simulator is really quite realistic because they had a lot of real world experience and the they've gone a billion miles in the simulator.

But we don't always have that opportunity to either create the data or have the data around.

Humans can learn from a small number of examples.

Your significant other, your professor, your boss, your investor can tell you something once or twice and you might actually learn from that.

Some humans have been reported to do that.

And that's kind of the remaining advantage of humans.

Now, there's actually no back propagation in the human brain.

That same year, in 1962, I wrote a paper how I thought the human brain worked.

There was actually very little neuroscience to go on.

There was 1 neuroscientist, Vernon Mountcastle, that had something relevant to say, which is he did I mean, there was the common wisdom at the time, and there's still a lot of neuroscientists that say this, that we have all these different regions of the brain, they do different things, they must be different.

There's V1 in the back of the head where the optic nerve spills into, that can tell that that's a curved line, that's a straight line.

Does he simple feature extractions on visual images?

It's actually a large part of the neocortex.

Does the fusiform gyrus up here, which can recognize faces?

We know that because if it gets knocked out through injury or stroke, people can't recognize faces.

They will learn it again with a different region of the neocortex.

There's a famous frontal cortex, which does language and poetry and music.

So these must work on different principles.

He did autopsies on the neocortex and all these different regions and found they all looked the same.

They had the same repeating pattern, same interconnections.

Otherwise, I could actually observe human brains in action, which I did from time to time, and there's a lot of hints that you can get that way.

For example, if I ask you to recite the alphabet, you actually don't do it from A to Z, you do it as a sequence of sequences, ABCDEFGHIJK.

So we learn things as forward sequences of sequences.

Forward because if I ask you to recite the alphabet backwards, You can't do it unless you learn that as a new sequence.

I wrote a paper that the neocortex is organized as a hierarchy of modules, and each module can learn a simple pattern.

And that's how I got to meet President Johnson, and that initiated a half century of thinking about this issue.

I came to MIT to study with Marvin Minsky.

1, that Minsky became my mentor, which was a mentorship that lasted for over 50 years.

The fact that MIT was so advanced it actually had a computer, which the other colleges I considered didn't have.

32k of 36-bit words, so it's 150k of core storage, 2 microsecond cycle time, 2 cycles for instruction, so a quarter of a MIP, and thousands of students and professors shared that 1 machine.

In 2012, I wrote a book about this thesis.

It's now actually an explosion of neuroscience evidence to support it.

The European Brain Reverse Engineering Project has identified a repeating module of about a hundred neurons.

It's repeated 300 million times, So it's about 30 billion neurons in the neocortex.

The neocortex is the outer layer of the brain.

And they can see in each module axons coming in from another module.

And then the output axis, the single output axis of that module goes as the input to another module.

So we can see it organized as a hierarchy.

The hierarchy comes from these connections.

There's 6 layers of neurons, but it constitutes 1 module.

And we can see that it learns a simple pattern.

And for various reasons, I cite in the book, the pattern recognition model it's using is basically a hidden Markov model.

How many of you have worked with Markov models?

That's usually no hands go up when I ask that question.

But a Markov model is not, it is learned, but it's not back propagation.

So it's very good for speech recognition and the speech recognition work I did in the 80s used these Markov models that became the standard approach because it can deal with local variations.

So the fact that a vowel is stretched, You can learn that in a Markov model.

It doesn't learn long distance relationships.

And something we don't fully understand yet is exactly how the neocortex creates that hierarchy.

But we have figured out how it can connect this module to this module.

I mean, there's no virtual communication or wireless communication.

So does it grow an axon from 1 place to another, which could be inches apart.

Actually, they all all these connections are there from birth, like the streets and avenues of Manhattan, this vertical and horizontal connection.

So If it decides and how it makes that decision is still not fully understood, that it wants to connect this module to this module, there's already a vertical, horizontal and a vertical connection that just activates them.

We can actually see that now and it can see that happening in real time on non-invasive brain scans.

So there's a tremendous amount of evidence that, in fact, the neocortex is a hierarchy of modules that can learn, each module learns a simple sequential pattern.

And even though the patterns we perceive don't seem like sequences, they may seem three-dimensional or even more complicated, they are in fact represented as sequences.

But the complexity comes in with the hierarchy.

So the neocortex emerged 200 million years ago with mammals.

It's 1 of the distinguishing features of mammals.

They were rodents, but they were capable of a new type of thinking.

Other non-mammalian animals had fixed behaviors, but those fixed behaviors were very well adapted for their ecological niche.

But these new mammals could invent a new behavior.

So creativity and innovation was 1 feature of the neocortex.

So a mouse is escaping a predator, its usual escape path is blocked, it will invent a new behavior to deal with it.

Probably wouldn't work, but if it did work it would remember it and would have a new behavior and that behavior could spread virally through the community.

Another mouse watching this would say to itself that was really clever going around that rock I'm going to remember to do that and it would have a new behavior.

Didn't help these early mammals that much because as I say the non-mammalian animals were very well adapted to their niches And nothing much happened for 135 million years.

But then 65 million years ago, something did happen.

There was a sudden violent change to the environment.

We now call it the Cretaceous extinction event.

There's been debate as to whether it was a meteor or a volcanic eruption.

The asteroid or meteor hypothesis is in the ascendancy.

But If you dig down to an area of rock reflecting 65 million years ago, the geologists will explain that it shows a very violent, sudden change to the environment.

The reason we call it an extinction event is that's when the dinosaurs went extinct.

That's when 75% of all the animal and plant species went extinct.

And that's when mammals overtook their ecological niche.

So to anthropomorphize biological evolution said to itself, this neocortex is pretty good stuff and it began to grow it.

So, now mammals got bigger, their brains got bigger at an even faster pace, taking up a larger fraction of their body.

The neocortex got bigger even faster than that and developed these curvatures that are distinctive of a primate brain, basically to increase its surface area.

But if you stretched it out, the human neocortex is still a flat structure.

It's about the size of a table napkin, just as thin.

And it basically created primates which became dominant in their ecological niche.

Then something else happened 2 million years ago.

Biological evolution decided to increase the neocortex further and increase the size of the enclosure and basically filled up the frontal cortex with our big skulls with more neocortex.

And up until recently it was felt, as I said, that this was, the frontal cortex was different because it does these qualitatively different things.

But We now realize that it's really just additional neocortex.

So remember what we did with it, we're already doing a very good job of being primates.

So we put it at the top of the neocortical hierarchy, and we increased the size of the hierarchy.

It was maybe 20% more neocortex, but it doubled or tripled the number of levels, because as you go up the hierarchy, it's kind of like a pyramid.

And that was the enabling factor for us to invent language and art, music.

Every human culture we've ever discovered has music.

There's debate about that, but it's really true.

Technology required another evolutionary adaptation, which is this humble appendage here.

If you look at a chimpanzee, it looks like they have a similar hand, but the thumb is actually down here.

It doesn't work very well if you watch them trying to grab a stick.

Yeah, I could take that branch and strip off the leaves and put a point on it.

We could actually carry out these ideas and create tools and then use tools to create new tools and started a whole other evolutionary process of tool making.

So Larry Page read my book in 2012 and liked it, so I met with him and asked him for an investment in a company I'd started actually a couple weeks earlier to develop those ideas commercially because that's how I went about things as a serial entrepreneur.

And he said, well, we'll invest, but let me give you a better idea why don't you do it here at Google.

We have a billion pictures of dogs and cats and we've got a lot of other data and lots of computers and lots of talent, all of which is true.

He says, well, I don't know, I just started this company to develop this.

And I said, how are you going to value a company that hasn't done anything?

So I took my first job 5 years ago and I've been basically applying this model, this hierarchical model to understanding language, which I think really is the holy grail of AI.

I think Turing was correct in designating basically text communication as what we now call a Turing-complete problem that requires...

There's no simple NLP tricks that you can apply to pass a valid Turing test with an emphasis on the word valid.

Mitch Caper and I had a 6 month debate on what the rules should be because if you read Turing's 1950 paper, he describes this in a few paragraphs and doesn't really describe how to go about it, but if it's a valid Turing test, meaning it's really convincing you through interrogation and dialogue that it's a human, that requires a full range of human intelligence.

And I think that test has stood the test of time.

I mean, just last week you may have read that 2 systems passed a paragraph comprehension test.

When I came to Google, we were trying to pass these paragraph comprehension tests.

Second grade test, we kind of got average performance.

And the third grade test had too much inference.

Already you had to know some common sense knowledge, as it's called, and make implications of things that were in different parts of the paragraph, and there's too much inference, and it really didn't work.

So this is now adult level, just slightly surpassed average human performance.

But we've seen that once something, an AI does something at average human levels, it doesn't take long for it to surpass average human levels.

I think it'll take longer in language than it did in simple games like Go, but it's actually very impressive that it surpasses now average human performance.

It's used in LSTM, long short temporal memory, but if you look at the adult test, in order to answer these questions, It has to put together inferences and implications of several different things in the paragraph with some common sense knowledge that's not explicitly stated.

So that's, I think, a pretty impressive milestone.

So I've been developing, I've got a team of about 45 people, and we've been developing this hierarchical model.

We don't use Markov models because we can use deep learning for each module.

And so we create an embedding for each word and we create an embedding for each sentence.

This we have a, I can talk about it because we have a published paper on it.

If you use smart reply on Gmail, if you use Gmail on your phone, you'll see it gives you 3 suggestions for responses.

They're simple suggestions, but it has to actually understand perhaps a complicated email.

And the quality of the suggestions is really quite good, quite on point.

That's for my team using this kind of hierarchical model.

So instead of Markov models, it uses embeddings Because we can use back propagation, we might as well use it.

But I think what's missing from deep learning is this hierarchical aspect of understanding.

That's why evolution developed a hierarchical brain structure to understand the natural hierarchy in the world.

And there are several problems with big, deep neural nets.

1 is the fact that you really do need a billion examples.

And we don't, sometimes we can generate them as in the case of Go, or if we have a really good simulator, as in the case of autonomous vehicles, not quite the case yet in biology.

Very often you don't have a billion examples.

See all Lex Fridman transcripts on Youtube

Ray Kurzweil: Future of Intelligence | MIT 6.S099: Artificial General Intelligence (AGI)