0:01
[Music] It's really really great to have you with us on Giant Ideas, Sebastian. Thanks for joining us. >> Yeah. Hi, Cameron. Hi, Tommy. >> Wow, lots to cover.
Search this video
0:15
Um, why don't we start with with autonomous vehicles?
0:17
Um, why do driverless cars matter, Sebastian?
0:22
What's the impact going to be on society?
0:25
>> Well, there's two fundamental things.
0:28
One is safety and one is cost.
0:28
On the safety side, we lose more than a million people every year in traffic accidents.
0:36
It's the leading cause of death for young people worldwide, and that is just not acceptable to me.
0:41
Uh self-driving cars have proven to be exponentially safer already at this point.
0:46
But the second thing is cost.
0:50
Um it turns out that a car is mostly not used.
0:52
Whoever owns a car on this show, please check.
0:59
Your car is probably being parked right now.
1:00
In fact, cars are being parked 96% of the time and driven maybe 3 to 4%.
1:06
And that means that you're wasting a ton of resources, okay, by virtue of not using your car.
1:11
So, the car is being shared among multiple people.
1:12
You get much higher utilization.
1:15
You don't get these inner cities full of parked cars.
1:20
You don't have to have a garage at home and all that stuff and never have to look for a parking spot again. >> Am I right?
1:25
thinking that for you the the safety issue and and this goal of trying to eradicate almost all of those million lives lost every year was a very personal thing based on on sort of tragedies that happened with you in the in Stanford with your colleagues.
1:40
I've now um lost a number of friends and co-workers to traffic accidents, but the most stingy thing was when I was 18 and my neighbor Harold went on a drive with a friend.
1:53
They found themselves on a sheet of ice, crashed into a truck, and both died instantaneously.
2:01
It took like less than a second.
2:03
I was thinking, "Wow, here's a promising life of an 18-year-old, my best friend at the time, gone for no reason."
2:12
Like, why do we do this to ourselves?
2:14
Uh why do we have teenagers drive on a sheet of ice in the first place, but then why do we use technology that's extremely dangerous as if it's just something completely, I know, casual?
2:27
Has has this therefore been a a a very very missiondriven thing for you all the way back to to that incident that you felt you simply had to solve this problem?
2:37
>> I mean that's basically the way I live my life that I see things that are just not okay. I want to fix them.
2:40
When it involves people's death, it's particularly painful.
2:45
And I'm sure many of your esteemed listeners have similar experiences with themselves or loved ones.
2:51
And yes, it's driven me to really think about as an AI researcher, as a roboticist.
2:56
How can I fix this big problem of more than a million deaths every year?
3:01
And the answer is make the car smart.
3:05
It's artificial intelligence.
3:06
Just make the car smarter than people are.
3:08
computer self-driving cars. They don't text. They don't drink. They're not distracted.
3:13
They're not on the phone. They're not fatigued.
3:14
They can look in all directions all the time.
3:19
Virtue of doing this, they can be safer than human drivers can be.
3:21
And now we know for a fact that they're safer than human drivers. >> Yeah.
3:26
And we love we love incredibly missiondriven entrepreneurs.
3:27
And sounds like you are exactly that.
3:29
Um, you've been at the center of the autonomy story really since the very beginning.
3:32
It'd be great if you could paint a picture for our audience of where it started, your role, uh maybe give us some tidbits and then where we are today. >> Yeah.
3:42
In 2001, 25 years ago, my god, the US government came with this idea of creating a competition for self-driving cars.
3:50
Their focus at the time was the war in Iraq.
3:55
They lost lots of soldiers to IEDs, improvised explosive devices, and they want to be able to move about without putting people's life, people's lives at risk.
4:04
Be as it is, they came with this thing called the DARPA Grand Challenge.
4:09
DARPA is a part of the US government that's responsible for the internet and stealth bombers and many other great innovations.
4:15
And they grand challenge was intended to be a challenge to the scientific community of professors around the world to build a car that could drive itself 140 miles roughly 200 km through a desert environment similar to Iraq.
4:29
They picked the Mojave Desert here in southwest United States and they said whoever can build a car that can drive itself without a person inside completely autonomous 200 km is going to make a million bucks.
4:45
And look, a million bucks is doesn't go far in terms of research budget.
4:47
When you run a university, you have you have billions of dollars uh that you use for research.
4:53
But as a individual professor, I was at Stanford at the time, it felt like a lot of money like my god, I can I can put a computer in a car. I can make it safe.
5:01
I can live up to my mission and figure out how to make cars safe and I can a million bucks. That's how it started.
5:09
>> And then you you won that challenge, right?
5:10
Am I right in thinking that Larry Paige, the Google co-founder, was sort of incognito in dark glasses watching who would win this and then and then came up to you?
5:18
It was the most surreal thing possible.
5:21
Like because normally I do lots of racing bicycles and marathons and the race is always the hard thing.
5:27
You sweat, but when you send your self-driving car off to a race, almost like sending a child off after training it for a year or so, and it races, but you're sitting there with sipping coffee and possibly champagne.
5:38
Um the race unfolded over 7 hours.
5:40
We had uh total of almost 2009 196 contestants applying for it.
5:47
20 23 finalists were allowed to race and of those uh 23 finalist uh five made it.
5:54
Um four within the lot time of 10 hours and I was at Stanford.
5:56
Our little Volkswag was the fastest. So we we kind of won. Yes, we did win.
6:01
But I always told people look it's not really important that we won at St.
6:04
important that the scientific community won because that was the birth moment at least in the United States for the self-driving car.
6:12
>> So, fast forward to today and getting into a Whimo in San Francisco uh is one of the most amazing hero product experiences I've ever had.
6:19
It just it just blows your mind.
6:21
My son uh when I was there with him was was uh very confused to see cars driving around without people in the front.
6:26
And now he he's he's 5 years old, but he talks regularly about the Whimo.
6:30
Um, but you you spent 25 years from, you know, from the time that Larry watched you in his dark glasses to today. 25 years has gone by.
6:38
Give us a bit of a color of how how we got to where we are today and where you think we are in the adoption curve.
6:44
>> Well, thanks Tommy and Cameron for asking that question and appreciate your son being enthusiastic.
6:48
Hopefully, he's going to be a future computer science student. Okay.
6:54
>> Right now, he's showing he's showing a lot of proclivity for art at the moment, but um >> Okay. Um, a little bit is art.
6:58
Mostly it's just straight math, physics, computer science and what we built.
7:02
Um we then uh founded a team and the very first team which actually a company that that Google acquired for me u built street view and street view showed the world a sliver of the self-driving car and it was able to record data but it wasn't really robotic.
7:18
It was mostly like human driven to record the world from the street level.
7:21
And then we moved into what we called project chaffur.
7:23
In 2009, we started a team at Google trying the impossible.
7:29
And we we built in the beginning eight cars that we drove on public streets in California uh with a safety driver.
7:37
They could take over, but still if you go like 100 km/h, it's pretty scary to be in a car that has been programmed overnight.
7:43
And we started driving um pretty much all streets in California.
7:47
streets in California. Uh there was city streets like San Francisco including the the Lombard Street if you ever been here or like surface streets all the way highway 1 the scenic highway to Los Angeles and around Lake Tahoe and
8:00
bridges places where there was just no GPS because you're in a tunnel everything and we had to get it to a point where the car was 100% correct and just to lay out how the hard this is when you um use chat GPT today or any large language model you have what's called hallucination. Hallucination is
8:18
Hallucination is the mo moment when the model makes something up that's not correct and it happens.
8:24
Okay, there's nothing wrong with it.
8:26
The technology is amazing, but it's a side effect of the way these things are being trained.
8:30
When you have a henation in a large language model, you you shake your head and you move on. And that's okay.
8:36
But if you're self-driving car hallucinates, it'll run a red light. It'll hit a person. It's not acceptable.
8:44
So we the bar in terms of reliability is so high that it took a much much longer time to train our artificial intelligence to be really really reliable.
8:55
And it's you get a lot of appreciation for people because people are good at this stuff.
9:00
They're really good at this stuff.
9:01
When we focus and we don't drink, we're not fatigued.
9:05
You're generally pretty safe. Okay.
9:05
But to get the same capability in the computer took about 15 years.
9:10
And maybe let's just talk about the the landscape of competitors out there.
9:15
Um, obviously Whimo is is very dominant in the US in in China.
9:20
BU's offering Apollo seems to be doing very very well.
9:23
Uh, strangely undervalued I think because BYU I think is is valued essentially at its cash reserves which seems odd given that Whimo some people think is worth almost $200 billion within Google um within Alphabet.
9:36
I I guess it' be great to hear from you just sort of how do you see the landscape?
9:40
There's also Wave here in London which was founded by a friend of ours Amar Sha. They're doing great.
9:43
Very different approach technically.
9:45
Maybe give us a sense of the competitive landscape.
9:48
Is there going to be one dominant winner?
9:50
Is it going to be you know a whole bunch of players with different technical approaches?
9:53
And also maybe unpick for us if you can just the importance to to Google to Alphabet of of Whimo.
10:00
How central to to Alphabet's future do you think Whimo is?
10:03
future do you think Whimo is? Look um in the US the and I'd say globally the the number one player right now is obviously way more in that way more has now driven on more than 100 million miles and never harmed a person which is something that
10:17
people just can't do okay in a 100 million miles of human driving which takes many many people in exploitation you kill more than one person um on its heels in the US is certainly Tesla which has a different approach Tesla has been trying to do the same capability based on camera only. And while the
10:34
And while the hypothesis is correct that eyes are sufficient to drive, in practice, it has been a bit more challenging than using lighters and more advanced sensors. But yes, you're correct.
10:46
Unbeknownst to most people in China, there's been massive progress in not just Bo, but Pony and Diddy and other companies who effectively started the Silicon Valley lab and hired a top dog from places like Whimo and then eventually moved over development to China.
11:01
And Chinese people does work twice as hard as Americans.
11:05
996 is the famous word from 9:00 in the morning to 9 in the evening, six days a week.
11:09
If you work on the math, it's more than the German 37. 5 hours. Okay?
11:10
and they've been um now launching systems and the services in China that are effectively on par getting very close to what VMO can do.
11:23
Um and similar in we see this in large language models between I don't know Grock and Alibaba and and Meta and and OpenAI and and Google that there's a race going on and um the the teams by and large are all like surpassing each other all the time.
11:40
There's nothing really secret about how you defend this.
11:42
Um the the basic rule is the more data you have, the more experience you have, the more reliable your system becomes.
11:48
And anybody who puts the money into generating the data and trading the systems eventually can increase the increase liability to the point where you can drive in public.
11:56
So big motion in China right now.
11:58
We all waiting for Germany to follow suit.
12:02
German companies of course are German car companies are the best in the world. There's no question.
12:06
They put a lot of emphasis on driver assist system where the driver is still the driver in charge.
12:11
Um and Germany has not yet made the move to a complete autonomous system.
12:15
But that's just a matter of time given the compounding data mode you you alluded to there.
12:20
Do you think it's a winner take all dynamic where there will be you know one leading autonomous driving car or will it break up by region?
12:27
How do you think about that?
12:29
>> That it depends on first local regulations and how the big players playing the game.
12:32
playing the game. um in the right hailing or right sharing world we find historically it's a winner take all position in United States it's effectively Uber who is the winner uh even though lift still exists in China we find Diddy is the winner in Singapore
12:48
it's Grab and when you look historically the smaller players that don't quite get the liquidity end up losing because there's a network effect and the network effect says the more people participate the more people participate the better the service to everybody, right? So
13:02
So early network effects for example is the fax machine.
13:06
The more people used to I mean this dates me used to have fax machines the better for everybody.
13:10
And then right sharing the more people use your ride sharing services the more you can afford putting cars all over the place and the shorter the wait time for the next car and the less the waste in terms of time wasted on the driver side.
13:24
So the cheaper you can make it.
13:24
So there is a possibility that a single winner emerges in this space but we are we are far away from this.
13:32
far away from this. we far away from even making the capital investment ability to build let's call it a million self-driving cars that is itself is already challenging but then also you find that local places different in regulations and they favor certain companies right so I would say Uber worked really hard in China um I'm I
13:52
wasn't part of the Uber or the China team at the time but it reported the time it went into headwinds because China is China and America is America and America likes American companies and China likes Chinese companies That might not be fair, but overall um local companies seem to have an advantage in these things. >> And if you were starting the Google
14:09
>> And if you were starting the Google self-driving car project today, how would your approach differ given there's been these huge advances in in foundational models and multimodal AI or would you take the exact same approach?
14:20
>> Well, when we started out first, uh we thought of this more as a like a 3D geometry game.
14:24
Like think about a video game where I use a lighter to really figure out how far exactly is this car wave up to millimeter precision.
14:31
How is it range changing?
14:33
Meaning what's its speed and so on and it became this big blocks worth of like blocks moving around.
14:40
Um in the beginning uh there was a good amount of machine learning involved ever since the DAR grand challenge but it wasn't the predominant uh solution.
14:48
as time moved on uh machine learning AI has massively improved especially when it comes to the understanding of images or video.
14:56
So for example going from a pixel cloud of like a 3D point cloud to saying this is a person this is a dog this is a trash bin that uh became much more feasible in the last call it 10 years or so.
15:11
So there has been a shift on the team to use much much more machine learning.
15:15
But there's one important caveat which is machine learning doesn't solve everything.
15:23
There's always going to be situations that are extremely rare but you have to handle correctly.
15:29
And if if you there's so rare that there no training set you can drive 100 million miles and you're never going to see the situation. Maybe it's I don't know.
15:35
Uh we had a case where a lady lost a stroller with this crying baby inside and the stroller zipped across the street.
15:43
And you can't train for this, right?
15:46
You still have to react correctly.
15:48
The machine learning model cannot say, "Oh, I never seen this. Just keep driving." That's not okay.
15:53
Uh it has to be able to react corrected even completely unforeseen situations.
15:57
And that's where to be quite frank a hybrid method seems to win today which is one that uses machine learning for understanding and prediction but also uses this kind of 3D logic the physics the point cloud to understand the physics of the space involved.
16:12
>> So Sebastian are we envisioning a world where close to 100% of cars on the road are going to be driverless or are we not approaching anywhere near that?
16:20
How long is it going to take?
16:21
is it going to take? try to paint a picture for our listeners if you can of when this world is going to arrive because it's arrived in San Francisco but in cities around the world you know there are almost no driverless cars on the road today I I really recommend everybody to come to San Francisco just
16:36
like the cable car in the Golden Gate Bridge the driverless car has become a tourist attraction and it's surreal to sit in a car without a driver inside and see it drive you it's something that takes a lot of courage in the beginning and after a few minutes a lot of appreciation and you feel how safe this video has become. Uh this technology can
16:53
Uh this technology can now be rolled out pretty much all across the world.
16:58
Maybe not in crazy places.
16:58
I recently visited India and India has different rules of traffic than the United States. Leave it at that.
17:05
Uh but by and large in European cities could be rolled out if the regulatory framework and the governments um play along.
17:12
Um so it's it's not a um if question, it's more like a when question.
17:18
Now I want to say to this uh and you're going to probably allude to this in a second.
17:22
Um I also worked on flying cars and I believe those are basically where the self-driving car was like 10 15 years ago.
17:30
So there will be more innovation and traffic.
17:32
It's not going to stop there.
17:34
Like there's going to be I mean the car itself is no older than 150 years.
17:39
The aircraft the airplane is no older than 150 years.
17:40
that that in the history of humanity which is like hundreds of thousands of years all these new things are relatively recent and don't expect innovation to stop.
17:51
>> So is autonomous air travel the future then? >> Yeah.
17:54
So if you look at the airbased stuff the work we did a company called Kittyhawk that I founded with Larry Page and we recently sold to Boeing.
17:59
Um what happens there is um it's it's not really a flying car.
18:04
It hasn't wheels, but it's something that takes off like a human drone uh electrically and then goes in a straight line.
18:10
And we were able to prove that these systems are about three times as energy efficient as green as a Tesla is, which at the time was the benchmark for most energy efficient vehicles on the ground.
18:22
And by virtue of being in the air, they're actually safer.
18:25
Turns out air travel is safer than ground travel.
18:28
It's because there's almost nothing to hit in the air, but there's lots of stuff to hit on the ground.
18:34
Um, so you could conceive we have a future where you don't really need roads anymore, right?
18:39
You can you can use roads, turn them into parks or whatever, right, into bicycle lanes, but then have these places where these things just take off like a drone and fly on a straight line.
18:47
They're also um not just greener, which is actually really remarkable.
18:52
People don't think of aviation as greener than ground transport, but but by and large it actually is greener.
18:57
Um this is something we can debate at length.
18:59
Um when you look at the actual energy consumed per traveler passenger mile, a plane wins or a train.
19:05
Turns out even though people hate when I say this, especially in Europe, um but it's also I mean it's also quieter.
19:11
It's these things are so incredibly quiet that you can't hear them anymore.
19:16
>> We had Reed Hoffman on the podcast a couple months ago and we he came up with an interesting analogy that you know intelligence is to the uh AI era what connectivity was to the internet era and everything is going to have intelligence.
19:27
everything is going to have autonomy.
19:28
You are a godfather of autonomy.
19:31
Um you know we we see aircraft, we see cars.
19:33
What are other big ideas or big sectors that will become autonomous and and really impact the world?
19:41
>> If you look at um companies, let's say Telos, complete different example, right?
19:46
So Telos employ tens of thousands of people. Um and what do they do?
19:48
I mean lots of stuff from sales to customer service to fixing the network and and cell phone towels and so on.
19:55
Um you can imagine that with intelligence you could build a telco with like 10 people. Um how would that work?
20:03
Well all these things that break all the time you have to fix you would automate.
20:09
Every time you interact with a customer you automate this and that's underway.
20:13
This is I'm not making this up.
20:15
This is definitely underway.
20:16
So you can take industries that are very labor intense and highly automate them.
20:21
Um that's happening in manufacturing.
20:23
So soft uh con's vision the the the company that employs what 1.
20:28
employs what 1.5 million people to manufacture I think things like your iPhone massive numbers of people tu the CEO recently said you like to guide to one employee okay then entire company is run by one person and the rest is robotic uh that sounds very much like
20:43
science fiction but bear with me in the following sense everything that you take for granted listener okay like your cell phone your car even even like very basic things, your food, your your your your clothing is a result of massive automation in the last 100 or so years and didn't exist before. And then even
21:01
And then even like your light switch is home, right?
21:05
Didn't exist 100 years, 50 years ago or your you almost certainly didn't have a flashing toilet at home like basic stuff like a warm shower.
21:10
Um so if you if you look a little bit zoom a little bit out and ask yourself what's happening in society you find that um things are moving really fast and there's really no reason why human labor is being used the way it's being used today.
21:23
The same way there's no reason why 150 years ago we all worked in farming and should have stayed this way right now less than 2% of us work in farming because of automation and in the future less of 2% of us will work into any company that we know today.
21:38
Yeah, I think that's really interesting when when you think about the debate of reshoring manufacturing jobs in the US and Europe.
21:43
You know, if you look at BYD, one of the the big the biggest uh EV producers, I think they make one car per day per two employees of the factory.
21:52
And I think the US is is closer to six, Europe, probably even more.
21:56
And so you you look at that and wonder where those jobs are going to be as we reshore them into into the US and Europe. >> Yeah.
22:03
Look, when I talk to my Silicon Valley friends about the mandate to rebuild manufacturing in the US, we talk about things like 3D printing, which would allow everybody to have effectively a factory in their basement.
22:15
Uh the best 3D printers today cost less than 2,000 bucks.
22:19
Best in the sense that there's a lot of um innovation happening at the grassroots.
22:24
The same way in the 70s and ' 80s there was a lot of innovation for PCs in the grassroots um for computing um with where all the innovation was back then and we all benefit from it today.
22:34
Um so there's there's thinking about how can we rechange manufacturing and maybe we can make it on demand as opposed to I know like take the clothing industry um the clothing industry is responsible for 10% of global emissions.
22:48
People don't know this and only two in three garments ever gets sold. Okay.
22:52
So a third of the stuff just got disposed, destroyed, sent to Africa, what have you, or donated it.
22:59
Um, if you could make clothing on demand, right, all of a sudden you could cut global emissions by, I don't know, 3%. It would be amazing.
23:05
Let's talk a little bit about what you describe and others describe as the singularity.
23:09
You said we are living through the singularity now in terms of super intelligence.
23:14
What does that mean for you in practice?
23:18
>> So what is a singularity?
23:18
A singularity is a moment where things accelerate so vastly that you won't be able to say what happens at the other end.
23:26
Okay, that's a singularity.
23:28
Uh and we can we can scribble about the definition.
23:30
But you look at um look at society today.
23:32
Um if you if you think about itself back say 100,000 years, you could have easily predicted the next 10,000 years because not much really changed to be honest.
23:43
Maybe a few little innovations like I don't know steel what have you, but that's pretty much it.
23:48
Stone Angels were boring.
23:50
Okay, if you lived 500 years ago, same thing.
23:52
Europe would be engulfed in war after war.
23:54
You lived 100 years ago, things a little bit faster. We have a steam engine.
23:58
You have uh the airplane innovation and stuff like this.
24:02
Um you have lived 50 years ago, 20 years ago.
24:04
When I was a student, I was able to predict the next seven years. Okay.
24:09
Okay, so in computer science, my own field, every seven years there was a major new programming language and a major new programming framework and things changed and past that point I couldn't predict.
24:18
Today I cannot even predict the next seven months because things are moving so insanely fast and that to me even the experts in the world can't predict what's happening that to me is a singularity.
24:29
We now see for the first time that for the vast majority of people, the machines are as good or better in big aspects of their daily work.
24:41
And I'm not scared of it.
24:41
I actually love it myself because I think we're wasting a ton of time ourselves doing stupid stuff.
24:46
But it's amazing to see how big machines now are so smart that they beat the world's best chess player and the world's best goal players, but also are as good as the world's best or maybe close as good as the world's best lawyers and doctors and all these things that we inspire our kids to do to be safe in life and have a great job.
25:04
>> And why is the pace the exponential pace of improvement and change? What's going on there? Why is it happening now?
25:09
Well, so the big innovation in the last few years, which I'm sure all your listeners have been talking about every night for many, many years now, is data.
25:17
Okay, so um OpenAI was probably the first to see this.
25:24
They got the world's text data together, which is hundreds of billions of documents and emails and pieces of software and they put in a big table in a big machine to predict more data. So how does this work?
25:37
like um you can always predict the next word in a simple way.
25:42
I say like the dog ate the dog food is a good word or maybe homework but not the propeller, right? That's not a good word.
25:50
So that idea of statistically figuring out what's a good next word um based on three, four, five words has been around for several decades in statistics, machine translation, linguistics.
26:02
But now um with the latest language technology, you can predict the next word based on hundreds of thousands of words in context.
26:12
And it turns out, strangely enough, uh predicting the next word, if you take enough context and you're smart enough about it, looks really intelligent.
26:21
So intelligent that large language models can write a rap song, they can write a haiku, they can translate languages.
26:28
In fact, they can do 90% of what a software engineer does today in terms of writing computer code, software engineering software.
26:38
They can now do this automatically just by looking at past data and extrapolating from it.
26:44
>> Sebastian, I think I've read an interview where you talked about in the next period of time because of those advancement, the pace of those advancement, we're going to enter a world where the basic needs of humans are going to be met in a very significantly different way.
26:57
We're going to have essentially close to free food, energy, even clothing, potentially housing.
27:04
And it's something that almost every one of the senior leaders in AI that I've spoken to and that we've had on the show uh on John Ideas, people like Mustafa Sullivan who founded Deep Mind, that seems to be a view that that people right at the center of AI's progress share.
27:19
And I don't it's never actually been particularly clear to me the leap between the pace of change we're talking about and all the amazing things that AI can do and it can be a great lawyer and so on.
27:29
But why why does that mean we're going to get almost free energy almost free food um these very physical basic needs that human beings need? >> Yeah.
27:38
I mean, if you look at even pre AAI and look at automation and and what we're doing in society, food is now so massively available that obesity has become a bigger problem than malnutrition.
27:52
There's still pockets in the world where food is very short right now, usually for political reasons, but if you were to distribute our food better, we could feed every human being to be honest.
28:01
Um, and that's new in history, right?
28:04
If you go back two, three, 400 years ago, all Europeans effectively lived in extreme poverty. Uh, take clothing.
28:14
Uh, 3 400 years ago, you would rip the clothing off dead people's body as you supply for clothing for many people.
28:21
Now, I mean, often the teacher is cheaper than the laundry to wash it.
28:28
It's amazing how disposable it has become. Or look at education.
28:30
Um, education is effectively available worldwide uh for pretty much everybody.
28:37
Maybe not the same quality, but online education I built a company for on that topic that really democratized access to high quality online education.
28:44
Um, K through 12 education for kids is available.
28:48
Vaccinations are available.
28:48
I think 95% of kids under the age of five are vaccinated. Some very large number.
28:56
Um so all these things have been improving even preai and they're all amazing um accomplishments because what they did to us people is they allowed us to escape the daily churn of farmwork and learn for example how to read or write and how to become scientists and how to really advance the world knowledge.
29:14
We've done amazing work in making argiculture more effective and and all these things. It's all preai.
29:22
Now AI will will superpower all this even faster right so we still live in a world where all of us effectively work most of us do and we work mostly in offices now and the western world is like 70% or so and what we do in offices is communicate with other people one another right so we write memos we write
29:41
orders we listen we go to meetings what have you we write email um all that stuff now can be automated very soon so free enormous creativity in society we shouldn't put us into an incredible position historically never happened before that we have a lot of free time to be creative. How amazing would that
29:57
How amazing would that be?
30:00
>> One last question on the topic of autonomy.
30:02
Uh you have designed some of these systems from the ground up.
30:05
You've been really at ground zero.
30:06
What are some of the potential failure modes that can be designed into these autonomous systems that aren't maybe being talked about enough >> especially in self-driving cars?
30:16
Um what you find is that there is um a long tale of situations that even the smartest engineers may never anticipate that might come to bite you.
30:29
Um so for example um let's take Whimo.
30:32
Whimo is just about now to go to autonomous highway driving.
30:37
Um and highway driving in many ways is the simplest or the easiest way of driving because if if if it goes well on a good day, nothing really moves.
30:45
In the United States, all the different lanes have the same speed limit.
30:48
So, we all basically run motor rolling around the same speed and relative to you, all the other traffic is effectively standing still.
30:55
But why did it take way more a long time to do autonomous driving, driverless driving on highways is because there's this very rare incident where like you have an exit in front of you or like a pedestrian or a cow or what have you, something that you haven't really thought about like a cow on a highway.
31:12
Okay, you think that is impossible unless you're in India, but in reality, yeah, there'll be a cow on a highway occasionally or a sheep or whatever, a dog, a coyote, a deer.
31:19
Deers are pretty common in highways.
31:22
So, how do you um deal with these very rare instances that could be conceivably fatal?
31:27
Like deer can actually kill people.
31:30
They're very hard to avoid.
31:31
Moose kill people regularly when when cars in highways hit a moose.
31:36
So, how do you deal with this incredibly rare instance and still make your car safe?
31:40
And that's challenging because as I said before, your tolerance for error, your tolerance for hallucinations in self-driving is effectively zero.
31:49
Maybe let's talk a little bit about education.
31:51
You obviously founded Udacity, this massive online learning startup.
31:55
Um maybe before we go into some of the questions about the future of education, just tell us if you can the story of how Udacity came about.
32:04
The company was started uh 2011 when I was teaching as a professor at Stanford University and I taught a a graduate level class on AI.
32:11
Today a very hot topic but trust me in 2011 it was an esoteric topic and I decided to um to put the class online and not just the lectures as videos but I I created a classroom environment where people had to take homework exams and quizzes and pass the the midterm exam and the final exam.
32:32
And I made it so that the online students would effectively have to pass the same exams as like the Stanford PhD doctoral students. Okay. Uh we we put it online.
32:43
We sent out one email to a couple of friends and I expected maybe 500 students would sign up.
32:47
I mean Stanford's campus was about 200 every year. So maybe a thousand.
32:51
That would be optimistic.
32:53
Uh the next morning, this was a Friday evening, we had 5,000 sign up.
32:59
Sunday morning 10,000 and Monday morning 14,000 people had already signed up which is when my dean at Stanford found out because I was being applauded on blogersphere for finally doing away with Stanford tuition which a message he didn't quite appreciate and then it went on to reach about 160,000 and I I realized my god my impact teaching 160,000 students is more than I can do in 10 years of life, 10 consecutive lifetimes at Stanford University.
33:34
So I really focused on how can I teach these 160,000 students effectively.
33:40
So we really built this new modality online interactive quiz-driven learning system overnight.
33:48
They would teach 160,000 students and and these were not just your typical SE students.
33:52
Some of them were like soldiers in Afghanistan.
33:53
Some were like was a single mother trying to raise their infant.
33:56
There was a person on the deathbed whose last wish really was to finish this class before he died.
34:00
People you would st would normally not admit, let's put it this way, okay?
34:04
But anywhere as smart as the best Stanford student.
34:07
In the end, 23,000 of these students finished and I got a chance to stackrank the Stanford students relative to these online students. And guess what?
34:18
The top 412 finishers were not at Stanford.
34:23
and the best Stanford doctoral student ranked number 413.
34:30
And that opened my eyes to I mean Stanford is an amazing institution.
34:32
I love Stanford and it does a fantastic job but it's very confined.
34:36
It only accepts a few thousand students every year, a few hundred students.
34:41
And the world has 8 billion people.
34:43
So if we just open the floodgates and let everybody study at Stanford, we would have so much more impact than having a small elite university here in California.
34:55
>> And is that the way that the world is going to go particularly now with AI?
34:59
And and how do you think AI is going to supercharge the efforts that you began really with Udacity to democratize education?
35:05
>> Yeah, first let's I mean Udacity became a global force.
35:07
We do a lot of Middle Eastern education in places like Egypt and and Saudi Arabia where people in Nigeria, Northern Africa where people really have no chance to ever get a great technical education.
35:17
Even moved the needle for probably 10 million or 20 million people at this point in terms of their life income and so on. It's quite amazing.
35:25
Um now AI is the next chapter and um there's a there's a very famous uh 1982 paper by a gentleman named Bloom who proved that if you give a normal kid a tutor to the entire like youth and lifetime when they grow up they will perform at the level of a highly exceptional gifted kid.
35:45
So you can turn a normal kid into a worldclass gifted child.
35:50
It's all relative of course just by tutoring it oneonone.
35:52
And what does a tutor do?
35:55
A tutor is much better to adapt to kids needs. Right?
35:57
So whereas a teacher might teach 30 kids at a time, a tutor teaches one kid at a time.
36:02
And it turns out in terms of math, English and so on, a tutor delivers better results.
36:09
Um now what AI can do is AI can now understand the learner deeper than ever before.
36:14
And rather than putting a single curriculum out for every kid identical like today when we teach kids, we have an identical curriculum for all the kids effectively.
36:21
It could adapt to the child and say this child is more of a visual learner or this child is needs more exercises or this cares more about history and less about math and then really evoke a a conversation with the child and help the child to become better and better. It's unproven. It's a hypothesis.
36:37
Um in child education hasn't been shown in adult in in uh professional training where I work it has been shown to work better.
36:45
Um, and I think they're going to go into a world where all of a sudden even learning gets superpowered and all the kids learn so much more than in the past.
36:55
>> So to maybe summarize, moving from democratization of education to mass personalization where everyone gets an individual tutor tailored to them. Amazing.
37:04
Um, you've had some huge successes uh like Whimo, but you've also had some challenges.
37:09
You had to wind down Kittyhawk.
37:10
What what heristics do you use when you're committing years of your life to an idea?
37:13
We know building a company is a challenge.
37:15
It's it's an adventure uh but always has ups and downs.
37:19
What are the heristics you use?
37:21
>> The majority of Kittyhog was sold to Boings. We did okay.
37:22
Still available as a company right now called Whisk. Arrow.
37:28
It wasn't exactly the path I anticipated in some cases.
37:31
Um I always think that there are some these amazing things where we can change the world and my goal is to make the world better for other people.
37:38
So for example, education.
37:40
What if you can truly democratize education, right?
37:42
That's the mountain you want to climb.
37:44
So you pick a mountain in life that will motivate you 10 years later. Right?
37:47
So self-driving cars motivates me 10 years later.
37:52
Transportation uh health motivates me 10 years later.
37:55
And that's the mountain you out to climb.
37:56
But then you if you actually this is a mountain never been climbed before, right?
38:00
So you pick something that hasn't been done before.
38:02
There's no book you can buy on Amazon says how to do it because it hasn't been done before.
38:05
And then you go and um and start climbing, right?
38:08
And as anyone knows, that's a complete different story.
38:11
story. you're going to put your first foot forward the second and then you I know you you run out of a false summit and you have to retract which happens all the time in innovation that you build something that doesn't work and you have to undo it possibly even fire people in this process um that that is painful for most people but it has to
38:28
keep motivating you because rather than you didn't make any progress maybe in terms of getting your goal done but you learned something interesting that other people don't know yet and you didn't know so you're not going to make the same mistake again so you're constantly making mistakes you go into as a climber let's Okay, take the mountain climbing algae. You go into bad weather, right?
38:42
You go into bad weather, right?
38:45
So all of a sudden there's a thunderstorm, right?
38:46
You can't see the summit anymore.
38:48
That doesn't mean the summit is gone.
38:49
You still have to believe it exists, but maybe your colleagues can't see it and they get frustrated and you have to keep climbing and you have to be very flexible.
38:56
Um my estimate is if you know what you're doing, if someone can be a little bird in your ear and tell you exactly every every moment exactly what to do, you're typically five times faster than if you're a true entrepreneur.
39:09
It takes five times longer to figure things out than if you could pick up a book at Amazon that will tell you how to do it.
39:17
So if you do something new like building the marketing education or building self-driving cars, four out of five days are effectively spent learning something interesting without making real progress towards the goal.
39:26
And then pick something that really is meaningful and to me the most meaningful stuff.
39:30
I always have this thing called grandmother test.
39:33
Like can I go to my fictitious my grandmother's passed away obviously the fictitious grandmother and tell her I'm working with something like this. Okay.
39:40
And if I tell her look I'm building a car that can get you anywhere that's self-driving she probably gets it.
39:46
But I say I'm building a lock likelihood discriminator that goes into blah blah blah and does blah blah blah.
39:53
She I don't care about that stuff.
39:53
So if and what are the basic things that people really deeply care about? It's communication.
39:59
it's health, it's education, it's safety.
40:02
Um, these are the things we truly maybe travel experiences we care about.
40:07
Can you relate your innovation to those things?
40:10
And I would say when you do things like transportation or education or most recently I've been working on on shopping and e-commerce.
40:17
These are things people do care about and you can you can explain to somebody who's non-technical.
40:24
>> Tell us about the new startup.
40:24
Yes, I haven't really talked much about it, but we have a a startup that's called Shop on Gold, which has launched just last week.
40:32
Uh, it's a personalization app that goes out on the internet and in the space of fashion finds everyday stuff for you that has been recently discount and you might actually like and use AI to match those millions of things we find every night that are showing up new in the fashion industry and finds those 50 pieces it believes you will love the most.
40:56
And then it also has an agentic AI where once you like something with a single click it does all the shopping for you.
41:03
So you don't have to go and create a password on some obscure website and sign up for some newsletter you didn't want to sign up for and all that stuff.
41:10
So it really changes the interface uh to the worldwide web.
41:15
Instead of having 30, 40, 50, 100 different shops you might actually attend to in your lifetime, there's just one now.
41:21
It does all the shopping for you. >> Fantastic.
41:24
Well, we should be watching the future of that. uh very very closely.
41:27
Maybe if we can just wrap things up, Sebastian, by asking something we ask everyone who comes on Giant Ideas.
41:32
Uh and modesty is not allowed in this answer.
41:34
So what is it about you do you think that has allowed you to have the kind of extraordinary success that we've just talked about for the past half an hour?
41:42
I'm both extremely arrogant and extremely humble.
41:45
Okay, I'm arrogant when I say let's democratize towards education.
41:47
I go to a Stanford deal and say screw this.
41:51
Stanford will admit 10 million people, not just 7,700 people.
41:54
Um, and that's certainly unappreciated by the people who care about the status quo.
41:59
Um, and I do this for everything.
42:02
I do this for government. I do this for health.
42:03
Um, these mountains that have to be climbed, you have to pick them.
42:07
And when you when you proclaim you're going to climb the mountain of making cars drive themselves, you're surrounded by people who basically think you're an idiot.
42:15
Okay, let's be let's be plain in the beginning. That's the way it feels.
42:17
But then you can't just pretend you know everything.
42:21
Um the actual process of climbing this mountain is extremely humbling because you constantly proven wrong and you have to be able to motivate yourself when you constantly do the wrong thing.
42:32
So I give an example for my most recently startup which is now doing extremely well uh shop on gold uh we did five iterations of building something just didn't work.
42:41
We tested it with users and they basically hated it.
42:46
And it's not because I'm not smart enough. Um maybe it is.
42:48
Um but it's very hard to understand what makes people really tick and what makes things work.
42:54
The same is true for the same was true for self-driving cars.
42:56
Did many many many iterations didn't work.
42:58
And to me uh those give me energy.
43:01
I love uncertainty and I love learning.
43:03
I love learning new things.
43:05
I love being proven wrong.
43:07
I have a rule in my life.
43:07
I'm allowed to make every mistake but only once.
43:11
If I make it the second time, I haven't learned anything about it.
43:13
and then I'm actually an idiot word.
43:15
But I I dive enormous energy out of this idea of like you do something, it doesn't work and you make a mistake and you look in this mirror and say I screw this up and it's okay. That's great.
43:27
I allow myself to make mistakes.
43:29
And that's I think where many people fail.
43:31
They they often move in and say, "I know everything."
43:35
And if you take the position to know everything, you got to be extremely lucky to be corrected because you're dealing with something that's never been done before and it's hard.
43:42
It's more likely you believe you know everything but you don't and then you're being stuck with your mistakes and you can't understand that you actually have to learn. Right?
43:52
So I feel like an like a first grader in what I do every single day.
43:57
I feel like a complete novice and I have no clue what I'm doing and I'm learning from the firehouse every single time I build a new company.
44:04
>> One final question, Sebastian.
44:04
What advice would you give to your 10-year-old self? >> Be curious.
44:09
Um I think the world is changing so rapidly.
44:10
Don't get married to anything that exists today.
44:12
And let's say uh develop grit.
44:15
Um grit to me is the one thing that is really driving success.
44:20
Grit means you stick with it, right?
44:23
When you if you're the person who jumps off quickly when the first failures come along, then you're not made for Silicon Valley.
44:29
You're not made for this world of innovation.
44:30
Pick a job like work for the German government or the UK government where things are more predictable.
44:35
But if you want to be in a world where where you invent something, you innovate, you push something forward.
44:41
In fact, pretty much any job.
44:44
If you have grit, you have a much higher chance for success. >> Superb place to end. Couldn't agree more.
44:49
Sebastian, thank you so much for sharing not just one giant idea, but about three or four in in this episode.
44:55
Really really appreciate you taking the time to join us.
44:57
>> Thanks so much for coming, Cameron.
44:57
It was a real pleasure meeting you. >> Likewise. >> Fantastic.
45:04
[Music] That was a fascinating conversation.
45:08
Uh, one one insight I take away from that was just how advanced the technology is today.
45:15
It's essentially pretty close to perfect.
45:17
And uh, by Sebastian's estimation, we have close to a million deaths a year from human drivers.
45:20
Yet, given that the situation, we still as a society, as as as a people aren't quite comfortable with the role of that technology, even though the cost of doing that is is fairly high.
45:31
That's pretty fascinating both philosophically and and uh and socially. >> Yeah, I agree.
45:35
And I think it's so striking when the technology is so far rolled out in San Francisco.
45:39
So we heard there that a quarter of ride shares, right?
45:43
A quarter of the the kind of Uber equivalents are now Whimo, driverless cars. >> It's just working. It's there.
45:48
And you know, go to San Francisco and it's a part of daily life and it's so completely foreign to people in Rome or London where you just don't see that at all. way.
45:57
>> And I think it's unusual because mostly the technology that we get out of Silicon Valley is just widely adopted everywhere all at the same time.
46:01
And it's it's an example of that not being true at all basically because of regulatory differences.
46:06
And I think that will probably increase over time because of this kind of geopolitical differences that are now being accentuated.
46:10
Um >> it's a very tangible example of the future is here but it's not yet equally distributed. >> Yes, totally.
46:16
And I feel like we need in Europe to get on it pretty quickly. Absolutely.
46:19
>> We're going to miss out.
46:19
>> We're going to miss out. I think one thing I would have liked to push Sebastian on more was this question of why exactly AI is going to suddenly deliver free food, free energy, free clothes, which as I said in the
46:30
interview like every single one of these these top AI thinkers are all saying, but no one to to me at least is actually nailing down exactly what is the jump between AI and free energy which obviously would solve climate change and all this and is to me just seems a bit of a fantasy. Do for you do you see
46:44
Do for you do you see that?
46:47
Yeah, I think it's an assumption of this extrapolating from the progress that's happening that it will continue which I think is is an assumption and that there will be this these emergent incredible benefits and we have to use our imagination but I also feel >> yeah it depends on the day that you ask me whether I think that's going to happen or not.
47:02
Um one thing I wish he would have gone a little bit deeper on was the failure modes of building these autonomous systems.
47:07
He's built them from the ground up.
47:08
uh you know it would have been great to go maybe a level deeper on how it could go wrong and how we should design these systems for those edge case scenarios to protect us and protect the planet and and protect uh yeah our lives. >> Agreed.
47:22
But overall I thought it was great and it was really nice.
47:23
You know I saw him 12 years ago in in Silicon Valley in his office and he told me driverless cars are going to be a thing is going to work and at that point no it was a kind of the the pit of despair for for autonomous vehicles and it's now happening and it's kind of down to him.
47:35
Lovely fellow and a clear visionary. [Music]