This is a talk with some advice about life and my own journey and passion in artificial intelligence. The audience is a group of Drexel engineering students, friends and family in Philadelphia.

This talk was before the outbreak of the coronavirus pandemic. Stay safe everyone, stay strong, we'll beat this thing.

OUTLINE:
0:00 - Overview - The Voice poem
6:46 - Artificial intelligence
13:44 - Open problems in AI
14:10 - Problem 1: Learning to understand
17:15 - Problem 2: Learning to act
19:28 - Problem 3: Reasoning
20:44 - Problem 4: Connection between humans & AI systems
23:57 - Advice about life as an optimization problem
24:10 - Advice 1: Listen to your inner voice - ignore the gradient
25:12 - Advice 2: carve your own path
26:28 - Advice 2: Measure passion not progress
28:10 - Advice 4: work hard
29:05 - Advice 5: forever oscillate between gratitude and dissatisfaction
31:10 - Q&A: Meaning of life
33:11 - Q&A: Simulation hypothesis
36:15 - Q&A: How do you define greatness?

CONNECT:
- Subscribe to this YouTube channel
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/lexfridman
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman

It's wonderful to be here, wonderful to see so many faces that I've come to love over the years.

My advisor, my family's here, my mom, brother.

You know, I did ask security to make sure my dad doesn't, is not allowed in, but he somehow found his way in, so good job.

The topic of today's talk reminds me of something my dad once told me.

Fewer those who see with their own eyes and feel with their own hearts.

No wait, That actually was Albert Einstein, different Jew, similar haircut for those of you.

You know, there's a saying, there's an old saying that goes, give a man a fish and you feed him for a day, teach a man

to fish and you feed him for a day, teach

him how to fish, and you feed him for a lifetime.

A little known fact, it actually goes on to say, so that he may never discover how much he loves steak, or vegetarian lasagna for those of you who are vegetarian in the audience.

And the key there, the key idea is society tries to impose lessons to teach, to drive the human being, each of us.

But it's you discovering your own passion is the key.

And that's what the talk I'd like to talk about today.

And there'll be a lot of poems throughout.

And the central poem by Shel Silverstein called The Voice is 1 I think that will resonate throughout the talk.

There's a voice inside of you that whispers all day long, I feel that this is right for me, I know that this is wrong.

No teacher, preacher, parent, friend, or wise man can decide what's right for you.

Just listen to the voice that speaks inside.

This is precisely the voice I'd like to talk about on in this brief little time we have together over 2 small topics.

Now, from an optimization perspective, And 1 of my co-advisors has always told me when you show a plot, you have to describe the X-axis and the Y-axis as a good engineer.

The X-axis is competence, the Y-axis is confidence.

And there's something called the Dunning-Kruger effect, which is captured by this plot, and that is at the beginning of your journey of competence, when you're not very good at something, when you're first taking the first steps at learning something, as some of you here are in the engineering fields, you're overly confident.

It's the peak of confidence, and you're at the lowest stage of actually of your abilities, of your expertise.

And It's funny that I am speaking here before you today in a place of a complete sort of self-doubt and despair and not knowing what I'm doing at all.

And I feel like I have 0 expertise to impart on you.

And so in that sense, it's a funny position to be speaking with, especially some of the lessons, some of the advice I'll try to give.

And some of you sitting in the audience today may be at the very peak, especially if you're at the beginning of the college journey, university journey.

And I'd say to me, the biggest positive, the biggest impact of college and university education is the dismantling of the ego that's involved in going from that peak overconfidence to the valley of despair that I'm currently in.

Oh, And I should mention that this is also the time for me and perhaps for you where folks like Dostoevsky start making a lot of sense, talking about suffering and pain and how the really great men and women must, I think as he says, have great sadness on earth.

This resonates with everybody in their undergraduate years in engineering.

Now, the real thing I'd like to talk about is the broader optimization problem formed by the Dunning-Kruger effect, which is after the peak of confidence and the valley of despair, there's a gradient provided to you by your advisors, by your parents, by your friends, your loved ones, society in general.

The gradient over which you're optimized to achieve some definition of success.

What everybody else tells you you're supposed to do.

What everybody else at the small scale, on a daily scale, and on the weekly scale, monthly, yearly, and for the rest of your life, tells you what the definition of success is.

What I'd like to argue is some ideas of how to break out of that convention, of how to listen just enough to hear the lessons in society, advisors, friends, and parents, but for the rest of it, ignore their voices and only listen to your own voice.

And I'll tell you through my own story here.

So I was introduced as a research scientist at MIT.

And very recently I decided to step down from MIT to do my own startup.

I'm still affiliated there, but sort of give up the salary, give up everything, give up what I'm supposed to be, the definition under academic colleagues of what success is, of what the pursuit of the academic life is, because I'm listening to the voice inside.

And so I'm speaking to you at the very beginning of this journey, again, full of self-doubt.

And so take with it with a grain of salt, but perhaps it's interesting to speak from this position Because I would argue it's the most beautiful position to be in in life.

The opportunity, the freedom in the struggles that I'm undergoing now is really a gift that comes at the end of this journey of college.

Now who am I and what is the dream that I mentioned there at the end?

For me that's understanding the human mind and engineering artificial intelligence systems.

Visualized on the left here is just 3% of the neurons in the human brain.

It's easy to forget how little we know about this mystery that's just between our 2 ears.

And engineering machines that can reason, that can think, that can perceive the world is 1 of the ways we can understand this mysterious, beautiful thing that brings to life everything around us.

And the dream of creating intelligent systems, companions, ones that you can have a deep connection with.

That's what my entrepreneurship work, that's my research work is focused on.

Most of the work at MIT and before that has been on robotics and autonomous vehicles, but now the dream is to create a system that you can love and it can love you back.

A brief history of artificial intelligence to give you a sense, to give you a quick review if this is a totally new field.

Again, if you're undergraduate, perhaps this is a field that you want to, that you want to take on as your journey.

So it started on the theoretical end with Alan Turing and many of the ideas from philosophy to mathematics that he presented and from whom the field was born.

And on the engineering side, Frank Crosonblot in building the Perceptron, the first machine, so engineering machines that can do some aspect of learning, some aspect of search that we associate with artificial intelligence.

And then there's been accomplishments throughout, none greater at least to me than in this, at least for now, in a span of games.

There's been 2 branches of artificial intelligence that have dominated the field.

The early days have been, you can think of as search, as brute force search.

It's not quite as captivating to our imagination.

It doesn't quite feel like intelligence because it's brute force searching through possible answers until you find 1 that's optimal.

It's converting every single problem to a search problem and then bringing computational power to it to try to solve it.

But nevertheless, the peak of that, especially for those who play chess, especially for those who might be a Russian, is when IBM D-Blue defeated Garry Kasparov in 1997.

This is a seminal moment in artificial intelligence where the game that was associated with thought, with intelligence, with reason, was overcome, was the greatest champion in human champion was defeated by a machine.

And the seminal moment on the second branch of artificial intelligence, which is learning systems, systems that learn from scratch, knowing nothing, with 0 human assistance, was able to defeat the greatest player in the world.

Little side note, the first moment did have human assistance in the AlphaGo system from DeepMind and Google DeepMind and then the follow on a few months later the system called AlphaZero was able to learn from scratch by playing itself.

This is the to me the greatest accomplishment of artificial intelligence.

I'll mention when I discuss it about open problems in the field.

Then in a real-world application, like I said, I worked a lot in autonomous vehicles.

This is 1 of the most exciting applications with autonomous and semi-autonomous vehicles.

There's been deployments, lessons, explorations, a lot of different debates.

This is the most exciting space of artificial intelligence.

If you want to have an impact as an engineer, autonomous vehicles is the space you will do so in the next in the 2020s.

And the quick whirlwind overview of key ideas in artificial intelligence that were key breakthroughs.

So neural networks and perceptron like I said was born in the 40s 50s and 60s With the algorithms that dominate today's world of deep learning and machine learning have been invented in many, many decades ago.

In the 70s and 80s with convolutional for the computer vision aspect of things in the 80s and 90s with LSTM, recurrent neural networks, they work with language, work with sequence of data, were developed in the 90s and proven out in the aughts.

And then the deep learning quote unquote revolution, The term and the ideas of large scale machine learning using neural networks was reborn in 2006 in the early aughts and then proven out in the seminal ImageNet moment when computer vision systems were able to, in the challenge of object recognition, image recognition, and the image net data set, and the image net challenge, neural networks were able to far outperform the competition and do so easily from just learning from data.

There's a lot of unsupervised learning, self-supervised learning ideas that were born in the 14, 15, 16, just a few years ago, and a lot of exciting ideas in the past 2 years.

The past 2 years have been dominated by ideas in natural language processing with ideas of transformers.

Anyway, this might be outside the scope of what you're familiar with.

Transformers in particular with natural languages, some of the most beautiful and exciting ideas that without any human supervision, you can learn to model language sufficiently well to outperform anything we've done previously.

To do things like machine translation to a level that's unprecedented is really exciting.

And especially exciting is that bigger is better, meaning that as long as we can scale compute, we can perform better and better and better.

And it's a totally open question how, what the ceiling of that is.

And finally, the most exciting thing in artificial intelligence is the idea, you know, there's a concept of Big Bang for the start of the universe, a silly name for 1 of the most incredible mysteries of our human existence.

Same way self-play is 1 of the silliest names for 1 of the most powerful ideas in artificial intelligence.

It's a system playing against itself to improve continuously without any human supervision.

That is the most exciting aspect, the most exciting area that I'm excited and I recommend if you love learning that you explore.

So the open problems in artificial intelligence and possible solutions.

And 1 of the things, and I'll focus on number 4, which is something that is my dream, that is sort of my life aspiration, but I'll give a whirlwind introduction.

Learning to understand, learning to act, reason, and a deep connection between humans and AI systems.

So learning to understand, there's a lot of exciting possibilities here.

This is a lot of the breakthroughs in machine learning have been in something called supervised learning, where you have a set of data and you have a neural network or a model that's able to learn from that data in order to generalize sufficiently to infer on cases it hasn't seen before.

In the case of domain, in the domains of like autonomous driving, you can recognize lane markings.

You could recognize other vehicles, pedestrians, all the different subtasks involved in solving a particular problem.

Now that's all good, but to solve real world problems, you have to actually, you have to deal with endless edge cases that we human beings effortlessly take care of, that our ability to do reasoning and common sense reasoning effortlessly takes care of.

So be able to learn over those edge case, you have to do much larger scale learning.

And for that, you have to be much more selective and clever about which data you annotate with human beings.

Same way with, as children, we explore the world, we interact with the world to pick up the lessons from it.

The same way you can interact with a data set to select only small parts of it to learn from.

And I'll take Tesla, which is a car company that's using autonomous driving and its system autopilot that uses deep learning to learn how to solve all these different problems.

What they're doing is quite interesting in the space of active learning.

They're creating a pipeline for each individual task.

They take the task of driving and break it apart into now over a hundred different subtasks.

Each subtask gets its own pipeline, its own data set.

And there's a machine learning system that learns from that data set and is then deployed back into the vehicles.

And when the vehicle fails in a particular case, that's an edge case that's marked for the system and is brought back to the pipeline to annotate.

So there is ongoing pipeline that continuously goes on.

The system is not very good in the beginning but the whole purpose of it is to discover edge cases in the same way that human us humans learn something and you can think of our actually existence in the world as an edge edge case discovery mechanism.

So you learn something you construct a mental model of the world and you move about the world until you run up against the case, a situation that you totally didn't expect.

And we do that thousands of times a day still and we learn from those.

And that pipeline of active learning is a really exciting area that very few people are working on, especially in the space of research.

To me, that's the most exciting in terms of scale impact area in the next few years.

Learning to act, the second set of open problems in artificial intelligence.

This is where the idea of self-play comes in.

Is learning to build systems, whether through a reinforcement learning mechanism or otherwise, that are actually acting in the world.

In the case of self-play, the idea is that you have a really dumb system in the beginning that knows nothing, again, no human supervision.

And through randomization, you have other systems that also know nothing but know a different set of nothing and they compete against each other so you can you formulate the problem as a competitive setting and when you have 2 dumb systems that compete against each other a magical thing happens The 1 that's slightly less dumb starts winning.

And this little incremental step can be repeated arbitrarily and without any constraints on human supervision, annotation costs, without any constraints on having to sort of bring the human in the loop or bring the physical world in the loop.

It can all be done in computation in a distributed sense.

So you can, in a matter of hours on a distributed compute setting, create a system that beats the world champion at go.

And in fact, with DeepMind and all the games that have they've defeated the world champion in chess, not just the world champion, is the best chess playing program Stockfish in a matter of hours of training.

This is both the exciting and the scary thing about self-play is very few times is the ceiling ever reached.

What we hit is the limits of our computational power, which is computation power, especially the kind of mechanisms that are happening now, developments happening now.

The Moore's law is continuing in many ways.

So computation, if you just wait a few years, computation is increasing.

So we were yet to see the ceiling of the capabilities that these approaches are able to achieve.

This should be both exciting and terrifying.

Okay, The total biggest open problem that nobody even knows how to do.

This is an example of a state-of-the-art dog intelligence system solving a particular problem.

So we know nothing how to do about, how to do reasoning systems in artificial intelligence.

This is the, actually not very often talked about area because nobody knows what to do about it.

There's been subsets called program synthesis, communities that kind of try to formulate a subset of the reasoning problem and try to solve it but we don't know much to do.

Particularly Common sense reasoning, how to formulate enough about the world to be able to reason about the physics of the world, about the basic, especially with human beings, human to human, human to physical world dynamics.

Just there's millions of facts seemingly that are intricately connected that we learn and we accumulate in a knowledge base.

This process is a really exciting area of research that nobody knows what to do with.

The things I've described previously don't really have anything to do with humans necessarily.

The bypassion in my interest is that space between machine and human.

The community broadly could be called human-robot interaction, but there's a lot of different areas in which there's a deep connection between the human and machine that you all experience every day.

So recommender systems from Netflix to, much more importantly, social networks, The recommendation engines behind social networks, recommending what you see next in terms of both advertisement and about the content of your friends that you see, which friends you get to see more from.

The personalization of IoT of smart systems, semi-autonomous systems, like Tesla autopilot and different semi-autonomous vehicles like the Cadillac super cruise systems.

Whenever you have AI systems between you and a machine.

So there's a machine that does, that automates some particular task.

There's you human that are tasked with sitting there and supervising the machine.

And there is an AI system in the middle that manages that.

It manages the tension, the dance, the uncertainty, the human, all the T word, the trust, all the mess of human beings, it manages that.

That's a really exciting space that is in the very early days.

What I show there is where my sense is, where we stand.

In 1998, there was a lot of search engines.

Some of you may even be old enough to have used them.

Alta Vista, Excite, AskG is Lycos and so on.

Then Google came along, the Google search engine and blew them all out of the water.

They were all working on a very interesting, very important problem, but the approach and the fundamental ideas behind their approach was flawed.

I believe that personal assistance and a personal deep meaningful connection between an AI system and a human being, that's exactly where we're at.

Many people have in their home an Alexa device, a Google Home device, but most people don't use it for almost anything except to play music or check the weather.

Many of you use Twitter and social networks, but artificial intelligence plays a minimal role and understands almost nothing about you in recommending how you interact with the platform or the advertisements you see.

And autonomous vehicles, robotics platforms know almost nothing about you.

It knows almost nothing about you, except whether your hands are on the steering wheel or not.

It's, I believe it'll be obvious in retrospect how much opportunity there is to learn about human beings from the devices and from that to form a deep meaningful connection.

So now to return to my valley of despair to give some words of advice and again take them with a grain of salt.

So in this context, in this optimization context, My first piece of advice is to listen to your inner voice.

I think a lot of people, including a lot of very smart professors, advisors, parents, friends, significant others, have in them a kind of mutually agreed upon gradient along which they push you.

It's so difficult for me to articulate this in a clear way, but early on, I heard within myself a silly sounding, crazy voice that told me to do things.

1 of which was to try to put a robot in every home.

There's dreams that are difficult for me to articulate, but if you allow your mind to be quiet enough, you'll hear such voices, you'll hear such dreams, and it's important to really listen and to pursue them.

Advice number 2 is carve your own path, And if that means taking a few detours, take the detours.

Again, this is coming from the Valley of Despair.

In music, I was in a band, I had long hair.

I gave a lot of myself to the practice of martial arts.

And both music and martial arts have given me, again, very difficult to put into words, but it have given me something quite profound.

It gave flavor and color to the pursuit of that dream that's hard to articulate.

It's because I listened to my instinct, listened to my heart in pursuing these detours from poetry to excessive reading.

Like I mentioned, I took a James Joyce course here, so pursuing these avenues of knowledge through philosophy and history that seemingly have nothing to do with the main pursuit and starting the silliest of pursuits, starting a podcast.

Advice number 3 is to measure passion, not progress.

So most of us get an average of about 27,000 days of life.

I think a good metric by which you should live is to maximize the number of those days that are filled with a passionate pursuit of something, not by how much you've progressed towards a particular goal.

Because goals are grounded in your comparison to other human beings, to something that's already been done before.

Passionate pursuit of something is the way you achieve something totally new.

Again, I'm a little bit of a Russian, so maybe I romanticize this whole suffering and passion thing.

But the people who love you, the people who care for you, like I mentioned, your friends, your family, should not be trusted.

Parents and significant others will tell you to find a secure job, because passion looks dangerous.

Advisors, colleagues will tell you to be pragmatic because passion looks like a distraction from the main effort that you should be focusing on.

And society will tell you to be, to find balance, work-life balance in your life, because passion looks unhealthy.

Advice number 4, continuing on the unhealthy part, is work hard.

Make a habit of working hard every day, putting in the hours.

There's a lot of books and a lot of advice that have been written on working smart and not working hard.

I'm yet to meet anyone who has not truly worked hard for thousands of hours in order to accomplish something great.

In order to work smart, you first have to put in those few tens of thousands of hours of really dumb, brute force, hard work of all-nighters.

The key there is to minimize stress, not to minimize the amount of hours of work.

And to do that is, you have to love what you do.

And the final piece of advice, I love that picture.

Okay, it's to look up to the stars and appreciate every single moment you're alive, at the mystery of this world, at the beauty of this world.

Again, this is my perspective, take it with a grain of salt, but I advise to forever oscillate between deep, profound doubt and self-dissatisfaction and the deep gratitude for the moment, for just being alive, for all the people around you that give you their love, with whom you get to share those moments and share the love.

A poem by Stephen Crane that I especially like in the desert.

In the desert I saw a creature, a naked bestial, who squatting upon the ground held his heart in his hands and ate of it.

So I would say the bitter is the self-dissatisfaction.

That's the restless energy that drives us forward.

And then enjoying that bitterness and enjoying the moment and enjoying the sweetness that comes from eating your own heart in this poem is a thing that makes life worthwhile.

So with those silly few pieces of advice, I'd like to continue on the gratitude and say thank you.

Thank you to this university for giving me a helping hand.

And thank you to my family and all the friends that I've had along the way.

I've never been introduced with this much energy.

See all Lex Fridman transcripts on Youtube

Escaping the Local Optimum of Low Expectation