Daphne Koller: Biomedicine and Machine Learning

00:00:00.000 | The following is a conversation with Daphne Kohler,

00:00:03.320 | a professor of computer science at Stanford University,

00:00:06.280 | a co-founder of Coursera with Andrew Ng,

00:00:09.000 | and founder and CEO of In-Citro,

00:00:11.880 | a company at the intersection

00:00:13.400 | of machine learning and biomedicine.

00:00:15.960 | We're now in the exciting early days

00:00:17.840 | of using the data-driven methods of machine learning

00:00:20.600 | to help discover and develop new drugs

00:00:22.600 | and treatments at scale.

00:00:24.440 | Daphne and In-Citro are leading the way on this

00:00:27.800 | with breakthroughs that may ripple through

00:00:29.960 | all fields of medicine,

00:00:31.600 | including ones most critical for helping

00:00:34.240 | with the current coronavirus pandemic.

00:00:36.360 | This conversation was recorded before the COVID-19 outbreak.

00:00:41.280 | For everyone feeling the medical, psychological,

00:00:43.520 | and financial burden of this crisis,

00:00:45.600 | I'm sending love your way.

00:00:47.680 | Stay strong.

00:00:48.800 | We're in this together.

00:00:50.000 | We'll beat this thing.

00:00:51.760 | This is the Artificial Intelligence Podcast.

00:00:54.240 | If you enjoy it, subscribe on YouTube,

00:00:56.360 | review it with five stars on Apple Podcast,

00:00:58.720 | support it on Patreon,

00:01:00.080 | or simply connect with me on Twitter at Lex Friedman,

00:01:03.440 | spelled F-R-I-D-M-A-N.

00:01:05.920 | As usual, I'll do a few minutes of ads now

00:01:08.080 | and never any ads in the middle

00:01:09.400 | that can break the flow of the conversation.

00:01:11.720 | I hope that works for you

00:01:13.040 | and doesn't hurt the listening experience.

00:01:15.920 | This show is presented by Cash App,

00:01:17.960 | the number one finance app in the App Store.

00:01:20.280 | When you get it, use code LEXPODCAST.

00:01:23.400 | Cash App lets you send money to friends,

00:01:25.600 | buy Bitcoin, and invest in the stock market

00:01:27.880 | with as little as $1.

00:01:30.200 | Since Cash App allows you to send

00:01:31.680 | and receive money digitally, peer-to-peer,

00:01:34.480 | and security in all digital transactions is very important,

00:01:38.080 | let me mention the PCI data security standard

00:01:41.360 | that Cash App is compliant with.

00:01:43.880 | I'm a big fan of standards for safety and security.

00:01:46.720 | PCI DSS is a good example of that,

00:01:49.480 | where a bunch of competitors got together and agreed

00:01:52.080 | that there needs to be a global standard

00:01:53.800 | around the security of transactions.

00:01:56.000 | Now we just need to do the same for autonomous vehicles

00:01:58.400 | and AI systems in general.

00:02:00.560 | So again, if you get Cash App from the App Store or Google Play

00:02:04.240 | and use the code LEXPODCAST, you get $10,

00:02:08.280 | and Cash App will also donate $10 to FIRST,

00:02:11.160 | an organization that is helping to advance robotics

00:02:14.040 | and STEM education for young people around the world.

00:02:17.640 | And now, here's my conversation with Daphne Koller.

00:02:22.360 | So you co-founded Coursera and made a huge impact

00:02:25.000 | in the global education of AI,

00:02:26.640 | and after five years, in August 2016, wrote a blog post

00:02:31.440 | saying that you're stepping away and wrote, quote,

00:02:34.440 | "It is time for me to turn to another critical challenge,

00:02:37.400 | "the development of machine learning

00:02:38.840 | "and its applications to improving human health."

00:02:41.640 | So let me ask two far-out philosophical questions.

00:02:45.080 | One, do you think we'll one day find cures

00:02:47.960 | for all major diseases known today?

00:02:50.720 | And two, do you think we'll one day figure out

00:02:53.520 | a way to extend the human lifespan,

00:02:55.960 | perhaps to the point of immortality?

00:02:57.840 | - So one day is a very long time,

00:03:01.760 | and I don't like to make predictions of the type

00:03:04.920 | we will never be able to do X

00:03:07.280 | because I think that's a, you know, that's smacks of hubris.

00:03:12.280 | Seems that never in the entire eternity of human existence

00:03:17.440 | will we be able to solve a problem.

00:03:19.360 | That being said, curing disease is very hard

00:03:24.200 | because oftentimes by the time you discover the disease,

00:03:28.480 | a lot of damage has already been done,

00:03:30.520 | and so to assume that we would be able to cure disease

00:03:34.960 | at that stage assumes that we would come up with ways

00:03:37.560 | of basically regenerating entire parts of the human body

00:03:41.880 | in the way that actually returns it to its original state,

00:03:45.320 | and that's a very challenging problem.

00:03:47.400 | We have cured very few diseases.

00:03:49.400 | We've been able to provide treatment

00:03:51.460 | for an increasingly large number,

00:03:52.920 | but the number of things that you could actually define

00:03:54.680 | to be cures is actually not that large.

00:03:57.600 | So I think that there's a lot of work

00:04:02.520 | that would need to happen before one could legitimately say

00:04:05.640 | that we have cured even a reasonable number,

00:04:08.800 | far less all diseases.

00:04:10.440 | - On a scale of zero to 100,

00:04:12.760 | where are we in understanding the fundamental mechanisms

00:04:15.560 | of all major diseases?

00:04:18.160 | What's your sense?

00:04:19.260 | So from the computer science perspective

00:04:21.080 | that you've entered the world of health,

00:04:24.160 | how far along are we?

00:04:25.720 | - I think it depends on which disease.

00:04:29.520 | I mean, there are ones where I would say

00:04:31.800 | we're maybe not quite at 100

00:04:33.400 | because biology is really complicated,

00:04:35.560 | and there's always new things that we uncover

00:04:38.960 | that people didn't even realize existed.

00:04:41.200 | But I would say there's diseases

00:04:44.420 | where we might be in the 70s or 80s,

00:04:48.080 | and then there's diseases in which I would say

00:04:52.000 | probably the majority where we're really close to zero.

00:04:55.200 | - Would Alzheimer's and schizophrenia

00:04:57.960 | and type two diabetes fall closer to zero or to the 80?

00:05:02.960 | - I think Alzheimer's is probably closer to zero than to 80.

00:05:09.320 | There are hypotheses,

00:05:12.680 | but I don't think those hypotheses have as of yet

00:05:17.320 | been sufficiently validated that we believe them to be true,

00:05:22.000 | and there's an increasing number of people

00:05:23.800 | who believe that the traditional hypotheses

00:05:25.920 | might not really explain what's going on.

00:05:28.020 | I would also say that Alzheimer's and schizophrenia

00:05:32.400 | and even type two diabetes are not really one disease.

00:05:35.320 | They're almost certainly a heterogeneous collection

00:05:39.400 | of mechanisms that manifest in clinically similar ways.

00:05:43.720 | So in the same way that we now understand

00:05:46.640 | that breast cancer is really not one disease,

00:05:48.920 | it is multitude of cellular mechanisms,

00:05:53.440 | all of which ultimately translate

00:05:55.160 | to uncontrolled proliferation, but it's not one disease,

00:05:59.340 | the same is almost undoubtedly true

00:06:01.160 | for those other diseases as well,

00:06:02.880 | and it's that understanding that needs to precede

00:06:05.780 | any understanding of the specific mechanisms

00:06:08.480 | of any of those other diseases.

00:06:10.120 | Now, in schizophrenia, I would say we're almost certainly

00:06:12.520 | closer to zero than to anything else.

00:06:15.200 | Type two diabetes is a bit of a mix.

00:06:18.240 | There are clear mechanisms that are implicated

00:06:21.400 | that I think have been validated

00:06:22.980 | that have to do with insulin resistance and such,

00:06:25.280 | but there's almost certainly there as well

00:06:28.520 | many mechanisms that we have not yet understood.

00:06:31.280 | - So you've also thought and worked a little bit

00:06:34.400 | on the longevity side.

00:06:35.880 | Do you see the disease and longevity

00:06:38.360 | as overlapping completely, partially,

00:06:42.080 | or not at all as efforts?

00:06:45.280 | - Those mechanisms are certainly overlapping.

00:06:48.620 | There's a well-known phenomenon that says

00:06:51.940 | that for most diseases, other than childhood diseases,

00:06:56.820 | the risk for contracting that disease

00:07:01.280 | increases exponentially year on year every year

00:07:03.840 | from the time you're about 40.

00:07:05.720 | So obviously there is a connection between those two things.

00:07:09.100 | That's not to say that they're identical.

00:07:12.400 | There's clearly aging that happens

00:07:14.960 | that is not really associated with any specific disease.

00:07:18.740 | And there's also diseases and mechanisms of disease

00:07:22.280 | that are not specifically related to aging.

00:07:25.640 | So I think overlap is where we're at.

00:07:29.120 | - Okay.

00:07:30.440 | It is a little unfortunate that we get older,

00:07:32.620 | and it seems that there's some correlation

00:07:34.160 | with the occurrence of diseases

00:07:39.080 | or the fact that we get older.

00:07:40.760 | And both are quite sad.

00:07:43.080 | - I mean, there's processes that happen as cells age

00:07:46.720 | that I think are contributing to disease.

00:07:49.520 | Some of those have to do with DNA damage

00:07:52.800 | that accumulates as cells divide

00:07:54.960 | where the repair mechanisms don't fully correct for those.

00:07:59.640 | There are accumulations of proteins

00:08:03.680 | that are misfolded and potentially aggregate,

00:08:06.360 | and those two contribute to disease

00:08:08.560 | and contribute to inflammation.

00:08:10.560 | There is a multitude of mechanisms that have been uncovered

00:08:14.040 | that are sort of wear and tear at the cellular level

00:08:17.120 | that contribute to disease processes.

00:08:20.060 | And I'm sure there's many that we don't yet understand.

00:08:24.920 | - On a small tangent, perhaps philosophical,

00:08:27.360 | does the fact that things get older

00:08:32.400 | and the fact that things die

00:08:34.760 | is a very powerful feature for the growth of new things.

00:08:38.920 | It's a kind of learning mechanism.

00:08:41.440 | So it's both tragic and beautiful.

00:08:43.740 | So (laughs)

00:08:46.920 | so in trying to fight disease and trying to fight aging,

00:08:53.960 | do you think about sort of the useful fact of our mortality?

00:08:59.920 | Or would you, like if you could be immortal,

00:09:02.680 | would you choose to be immortal?

00:09:04.280 | - Again, I think immortal is a very long time.

00:09:10.680 | (Lex laughs)

00:09:11.520 | And I don't know that that would necessarily be something

00:09:16.040 | that I would want to aspire to,

00:09:17.960 | but I think all of us aspire to an increased health span,

00:09:22.960 | I would say, which is an increased amount of time

00:09:27.680 | where you're healthy and active

00:09:29.920 | and feel as you did when you were 20.

00:09:33.160 | We're nowhere close to that.

00:09:36.800 | People deteriorate physically and mentally over time,

00:09:41.800 | and that is a very sad phenomenon.

00:09:43.720 | So I think a wonderful aspiration would be

00:09:47.360 | if we could all live to the biblical 120,

00:09:51.600 | maybe, in perfect health.

00:09:53.800 | - In high quality of life.

00:09:54.920 | - High quality of life.

00:09:55.920 | I think that would be an amazing goal

00:09:57.840 | for us to achieve as a society.

00:09:59.360 | Now is the right age, 120, or 100, or 150?

00:10:03.720 | I think that's up for debate,

00:10:05.800 | but I think an increased health span

00:10:07.720 | is a really worthy goal.

00:10:09.100 | - And anyway, in the grand time of the age of the universe,

00:10:14.760 | it's all pretty short.

00:10:16.680 | So from the perspective,

00:10:18.520 | you've done, obviously, a lot of incredible work

00:10:21.060 | in machine learning,

00:10:22.120 | so what role do you think data and machine learning

00:10:25.220 | play in this goal of trying to understand diseases

00:10:29.320 | and trying to eradicate diseases?

00:10:31.840 | - Up until now, I don't think it's played

00:10:35.200 | very much of a significant role,

00:10:37.880 | because largely, the data sets that one really needed

00:10:42.440 | to enable a powerful machine learning method,

00:10:47.320 | those data sets haven't really existed.

00:10:49.660 | There's been dribs and drabs

00:10:50.960 | and some interesting machine learning

00:10:53.320 | that has been applied,

00:10:54.680 | I would say machine learning/data science,

00:10:57.640 | but the last few years are starting to change that.

00:11:00.160 | So we now see an increase in some large data sets,

00:11:05.160 | but equally importantly, an increase in technologies

00:11:11.320 | that are able to produce data at scale.

00:11:14.680 | It's not typically the case

00:11:16.360 | that people have deliberately, proactively

00:11:20.160 | used those tools for the purpose

00:11:22.440 | of generating data for machine learning.

00:11:24.200 | They, to the extent that those techniques

00:11:26.560 | have been used for data production,

00:11:28.560 | they've been used for data production

00:11:29.860 | to drive scientific discovery,

00:11:31.320 | and the machine learning came as a sort of byproduct,

00:11:34.440 | second stage of, oh, now we have a data set,

00:11:36.920 | let's do machine learning on that,

00:11:38.260 | rather than a more simplistic data analysis method.

00:11:41.800 | But what we are doing at In-Sitro

00:11:44.440 | is actually flipping that around and saying,

00:11:46.840 | here's this incredible repertoire of methods

00:11:50.280 | that bioengineers, cell biologists have come up with.

00:11:54.560 | Let's see if we can put them together in brand new ways

00:11:57.400 | with the goal of creating data sets

00:12:00.240 | that machine learning can really be applied on productively

00:12:03.360 | to create powerful predictive models

00:12:06.580 | that can help us address fundamental problems

00:12:08.440 | in human health.

00:12:09.440 | - So really focus, to get, make data the primary focus

00:12:14.440 | and the primary goal,

00:12:15.800 | and find, use the mechanisms of biology and chemistry

00:12:18.920 | to create the kinds of data set

00:12:23.360 | that could allow machine learning to benefit the most.

00:12:25.720 | - I wouldn't put it in those terms,

00:12:27.580 | because that says that data is the end goal.

00:12:30.480 | Data is the means.

00:12:32.160 | So for us, the end goal is helping address challenges

00:12:35.760 | in human health, and the method that we've elected

00:12:39.200 | to do that is to apply machine learning

00:12:42.680 | to build predictive models.

00:12:44.120 | And machine learning, in my opinion,

00:12:46.000 | can only be really successfully applied

00:12:48.800 | especially the more powerful models

00:12:50.680 | if you give it data that is of sufficient scale

00:12:53.520 | and sufficient quality.

00:12:54.520 | So how do you create those data sets

00:12:58.560 | so as to drive the ability to generate predictive models

00:13:03.560 | which subsequently help improve human health?

00:13:05.680 | - So before we dive into the details of that,

00:13:08.360 | let me take a step back and ask,

00:13:10.280 | when and where was your interest in human health born?

00:13:16.760 | Are there moments, events, perhaps, if I may ask,

00:13:19.880 | tragedies in your own life that catalyzes passion,

00:13:23.040 | or was it the broader desire to help humankind?

00:13:26.560 | - So I would say it's a bit of both.

00:13:29.160 | So on, I mean, my interest in human health

00:13:32.600 | actually dates back to the early 2000s

00:13:36.520 | when a lot of my peers in machine learning

00:13:42.760 | and I were using data sets

00:13:45.480 | that frankly were not very inspiring.

00:13:47.400 | Some of us old timers still remember

00:13:49.800 | the quote unquote 20 news groups data set

00:13:52.320 | where this was literally a bunch of texts

00:13:55.720 | from 20 news groups,

00:13:57.120 | a concept that doesn't really even exist anymore.

00:13:59.240 | And the question was, can you classify

00:14:01.120 | which news group a particular bag of words came from?

00:14:06.760 | And it wasn't very interesting.

00:14:08.680 | The data sets at the time on the biology side

00:14:12.440 | were much more interesting both from a technical

00:14:14.920 | and also from an aspirational perspective.

00:14:17.520 | They were still pretty small,

00:14:18.800 | but they were better than 20 news groups.

00:14:20.720 | And so I started out, I think,

00:14:23.600 | just by wanting to do something that was more,

00:14:27.280 | I don't know, societally useful and technically interesting.

00:14:30.760 | And then over time became more and more interested

00:14:34.400 | in the biology and the human health aspects for themselves

00:14:39.400 | and began to work even sometimes on papers

00:14:43.400 | that were just in biology

00:14:45.080 | without having a significant machine learning component.

00:14:48.400 | I think my interest in drug discovery

00:14:52.680 | is partly due to an incident I had

00:14:57.320 | when my father sadly passed away about 12 years ago.

00:15:02.520 | He had an autoimmune disease that settled in his lungs

00:15:06.960 | and the doctor's basic said,

00:15:11.360 | well, there's only one thing that we could do,

00:15:13.280 | which is give him prednisone.

00:15:14.960 | At some point, I remember a doctor even came and said,

00:15:17.640 | "Hey, let's do a lung biopsy to figure out

00:15:19.520 | "which autoimmune disease he has."

00:15:20.880 | And I said, "Would that be helpful?

00:15:23.080 | "Would that change treatment?"

00:15:23.920 | He said, "No, there's only prednisone.

00:15:25.360 | "That's the only thing we can give him."

00:15:27.120 | And I had friends who were rheumatologists who said,

00:15:29.880 | "The FDA would never approve prednisone today

00:15:31.920 | "because the ratio of side effects to benefit

00:15:36.920 | "is probably not large enough."

00:15:39.560 | Today, we're in a state where there's probably four or five,

00:15:44.560 | maybe even more, well, it depends for which autoimmune disease

00:15:48.720 | but there are multiple drugs that can help people

00:15:52.880 | with autoimmune disease,

00:15:53.920 | many of which didn't exist 12 years ago.

00:15:56.720 | And I think we're at a golden time in some ways

00:16:00.340 | in drug discovery where there's the ability to create drugs

00:16:06.800 | that are much more safe and much more effective

00:16:10.600 | than we've ever been able to before.

00:16:13.080 | And what's lacking is enough understanding

00:16:16.360 | of biology and mechanism to know where to aim that engine.

00:16:21.360 | And I think that's where machine learning can help.

00:16:25.400 | - So in 2018, you started and now lead a company in Citro,

00:16:29.920 | which is, like you mentioned,

00:16:31.760 | perhaps the focus is drug discovery

00:16:34.760 | and the utilization of machine learning for drug discovery.

00:16:38.140 | So you mentioned that, quote,

00:16:40.540 | "We're really interested in creating what you might call

00:16:43.600 | "a disease in a dish model, disease in a dish models,

00:16:47.320 | "places where diseases are complex,

00:16:49.080 | "where we really haven't had a good model system,

00:16:52.140 | "where typical animal models that have been used for years,

00:16:54.960 | "including testing on mice, just aren't very effective."

00:16:58.860 | So can you try to describe what is an animal model

00:17:02.640 | and what is a disease in a dish model?

00:17:05.360 | - Sure.

00:17:06.280 | So an animal model for disease is where you create

00:17:11.280 | effectively, it's what it sounds like.

00:17:14.880 | It's oftentimes a mouse where we have introduced

00:17:19.280 | some external perturbation that creates the disease,

00:17:22.800 | and then we cure that disease.

00:17:26.280 | And the hope is that by doing that,

00:17:28.720 | we will cure a similar disease in the human.

00:17:31.320 | The problem is that oftentimes the way in which

00:17:35.040 | we generate the disease in the animal

00:17:36.880 | has nothing to do with how that disease

00:17:38.520 | actually comes about in a human.

00:17:40.880 | It's what you might think of as a copy of the phenotype,

00:17:44.360 | a copy of the clinical outcome,

00:17:46.720 | but the mechanisms are quite different.

00:17:48.720 | And so curing the disease in the animal,

00:17:52.080 | which in most cases doesn't happen naturally,

00:17:54.840 | mice don't get Alzheimer's, they don't get diabetes,

00:17:57.160 | they don't get atherosclerosis,

00:17:58.680 | they don't get autism or schizophrenia.

00:18:01.240 | Those cures don't translate over

00:18:05.680 | to what happens in the human.

00:18:08.120 | And that's where most drugs fails,

00:18:10.840 | just because the findings that we had in the mouse

00:18:13.680 | don't translate to a human.

00:18:15.040 | The disease in the dish models is a fairly new approach.

00:18:20.840 | It's been enabled by technologies that have not existed

00:18:25.840 | for more than five to 10 years.

00:18:28.360 | So for instance, the ability for us to take a cell

00:18:32.760 | from any one of us, you or me,

00:18:35.560 | revert that say skin cell to what's called stem cell status,

00:18:39.960 | which is what's called a pluripotent cell

00:18:44.760 | that can then be differentiated

00:18:46.600 | into different types of cells.

00:18:47.840 | So from that pluripotent cell,

00:18:49.800 | one can create a Lex neuron or a Lex cardiomyocyte

00:18:54.280 | or a Lex hepatocyte that has your genetics,

00:18:57.760 | but that right cell type.

00:19:00.320 | And so if there's a genetic burden of disease

00:19:04.800 | that would manifest in that particular cell type,

00:19:07.160 | you might be able to see it by looking at those cells

00:19:10.320 | and saying, oh, that's what potentially sick cells look like

00:19:13.860 | versus healthy cells and understand how,

00:19:18.520 | and then explore what kind of interventions

00:19:20.760 | might revert the unhealthy looking cell to a healthy cell.

00:19:24.860 | Now, of course, curing cells is not the same

00:19:27.720 | as curing people.

00:19:28.860 | And so there's still potentially a translatability gap,

00:19:33.220 | but at least for diseases that are driven, say,

00:19:38.220 | by human genetics and where the human genetics

00:19:41.980 | is what drives the cellular phenotype,

00:19:43.780 | there is some reason to hope that if we revert those cells

00:19:47.960 | in which the disease begins

00:19:49.600 | and where the disease is driven by genetics

00:19:52.220 | and we can revert that cell back to a healthy state,

00:19:55.260 | maybe that will help also revert

00:19:58.140 | the more global clinical phenotypes.

00:20:00.900 | So that's really what we're hoping to do.

00:20:02.760 | - That step, that backward step, I was reading about it,

00:20:06.020 | the Yamanaka factor.

00:20:08.300 | - Yes.

00:20:09.140 | - So it's like that reverse step back to stem cells.

00:20:12.300 | - Yes.

00:20:13.140 | - Seems like magic.

00:20:14.180 | - It is.

00:20:15.020 | Honestly, before that happened,

00:20:17.660 | I think very few people would have predicted

00:20:20.100 | that to be possible.

00:20:21.700 | It's amazing.

00:20:22.540 | - Can you maybe elaborate, is it actually possible?

00:20:25.340 | So this result was maybe, I don't know how many years ago,

00:20:30.580 | maybe 10 years ago was first demonstrated,

00:20:32.700 | something like that.

00:20:34.300 | How hard is this?

00:20:35.780 | How noisy is this backward step?

00:20:37.500 | It seems quite incredible and cool.

00:20:39.460 | - It is incredible and cool.

00:20:42.220 | It was much more, I think, finicky and bespoke

00:20:46.420 | at the early stages when the discovery was first made,

00:20:49.980 | but at this point it's become almost industrialized.

00:20:54.500 | There are what's called contract research organizations,

00:20:59.460 | vendors that will take a sample from a human

00:21:02.300 | and revert it back to stem cell status,

00:21:04.460 | and it works a very good fraction of the time.

00:21:07.100 | Now there are people who will ask, I think, good questions.

00:21:12.020 | Is this really, truly a stem cell or does it remember

00:21:15.340 | certain aspects of changes that were made

00:21:19.420 | in the human beyond the genetics?

00:21:22.500 | - It's passed as a skin cell, yeah.

00:21:24.660 | - It's passed as a skin cell or it's passed

00:21:26.740 | in terms of exposures to different environmental factors

00:21:29.900 | and so on.

00:21:30.940 | So I think the consensus right now

00:21:33.300 | is that these are not always perfect

00:21:36.420 | and there is little bits and pieces of memory sometimes,

00:21:40.020 | but by and large, these are actually pretty good.

00:21:43.580 | - So one of the key things, well, maybe you can correct me,

00:21:48.740 | but one of the useful things for machine learning

00:21:50.900 | is size, scale of data.

00:21:54.180 | How easy it is to do these kinds of reversals to stem cells

00:21:59.100 | and then does using a dish models at scale?

00:22:02.340 | Is that a huge challenge or not?

00:22:05.300 | - So the reversal is not, as of this point,

00:22:11.180 | something that can be done at the scale of tens of thousands

00:22:16.340 | or hundreds of thousands.

00:22:18.500 | I think total number of stem cells or IPS cells

00:22:22.220 | that are what's called induced pluripotent stem cells

00:22:25.220 | in the world, I think is somewhere between five and 10,000

00:22:30.180 | last I looked.

00:22:31.420 | Now, again, that might not count things that exist

00:22:34.420 | in this or that academic center

00:22:36.220 | and they may add up to a bit more,

00:22:37.820 | but that's about the range.

00:22:40.020 | So it's not something that you could this point

00:22:42.140 | generate IPS cells from a million people,

00:22:45.540 | but maybe you don't need to

00:22:47.900 | because maybe that background is enough

00:22:51.820 | because it can also be now perturbed in different ways.

00:22:56.140 | And some people have done really interesting experiments

00:23:00.100 | in, for instance, taking cells from a healthy human

00:23:05.100 | and then introducing a mutation into it

00:23:08.540 | using one of the other miracle technologies

00:23:11.860 | that's emerged in the last decade,

00:23:13.860 | which is CRISPR gene editing

00:23:16.140 | and introduced a mutation that is known to be pathogenic.

00:23:19.660 | And so you can now look at the healthy cells

00:23:22.420 | and unhealthy cells, the one with the mutation

00:23:24.700 | and do a one-on-one comparison

00:23:26.060 | where everything else is held constant.

00:23:28.380 | And so you could really start to understand specifically

00:23:31.820 | what the mutation does at the cellular level.

00:23:34.380 | So the IPS cells are a great starting point

00:23:37.660 | and obviously more diversity is better

00:23:39.780 | 'cause you also wanna capture ethnic background

00:23:42.380 | and how that affects things,

00:23:43.580 | but maybe you don't need one from every single patient

00:23:46.780 | with every single type of disease

00:23:48.140 | because we have other tools at our disposal.

00:23:50.300 | - Well, how much difference is there between people?

00:23:52.580 | I mentioned ethnic background.

00:23:53.860 | In terms of IPS cells,

00:23:54.940 | so we're all, like it seems like these magical cells

00:23:59.380 | that can do anything, create anything

00:24:01.860 | between different populations, different people.

00:24:04.020 | Is there a lot of variability between cell cells?

00:24:07.020 | - Well, first of all, there's the variability

00:24:09.580 | that's driven simply by the fact

00:24:10.980 | that genetically we're different.

00:24:13.420 | So a stem cell that's derived from my genotype

00:24:15.820 | is gonna be different from a stem cell

00:24:18.340 | that's derived from your genotype.

00:24:20.540 | There's also some differences

00:24:21.780 | that have more to do with, for whatever reason,

00:24:25.300 | some people's stem cells differentiate better

00:24:28.540 | than other people's stem cells.

00:24:29.860 | We don't entirely understand why,

00:24:31.540 | so there's certainly some differences there as well.

00:24:34.180 | But the fundamental difference

00:24:35.460 | and the one that we really care about and is a positive

00:24:38.740 | is the fact that the genetics are different

00:24:43.220 | and therefore recapitulate my disease burden

00:24:45.980 | versus your disease burden.

00:24:47.780 | - What's a disease burden?

00:24:49.260 | - Well, a disease burden is just, if you think,

00:24:52.300 | I mean, it's not a well-defined mathematical term,

00:24:55.060 | although there are mathematical formulations of it.

00:24:58.260 | If you think about the fact that some of us

00:25:00.540 | are more likely to get a certain disease than others

00:25:03.460 | because we have more variations in our genome

00:25:07.300 | that are causative of the disease,

00:25:09.500 | maybe fewer that are protective of the disease.

00:25:12.600 | People have quantified that

00:25:14.860 | using what are called polygenic risk scores,

00:25:17.860 | which look at all of the variations

00:25:20.820 | in an individual person's genome

00:25:23.620 | and add them all up in terms of how much risk they confer

00:25:26.620 | for a particular disease,

00:25:27.820 | and then they've put people on a spectrum

00:25:30.540 | of their disease risk.

00:25:32.540 | And for certain diseases

00:25:34.560 | where we've been sufficiently powered

00:25:36.940 | to really understand the connection

00:25:38.740 | between the many, many small variations

00:25:41.580 | that give rise to an increased disease risk,

00:25:44.940 | there is some pretty significant differences

00:25:47.040 | in terms of the risk between the people,

00:25:49.300 | say, at the highest decile of this polygenic risk score

00:25:52.060 | and the people at the lowest decile.

00:25:53.500 | Sometimes those differences are a factor of 10 or 12 higher,

00:25:58.500 | so there's definitely a lot that our genetics

00:26:03.940 | contributes to disease risk,

00:26:06.140 | even if it's not by any stretch the full explanation.

00:26:09.100 | - And from a machine learning perspective,

00:26:10.500 | there's signal there.

00:26:12.020 | - There is definitely signal in the genetics,

00:26:14.780 | and there's even more signal, we believe,

00:26:19.060 | in looking at the cells that are derived

00:26:21.540 | from those different genetics,

00:26:23.460 | because in principle, you could say all the signal

00:26:26.580 | is there at the genetics level,

00:26:28.680 | so we don't need to look at the cells,

00:26:30.180 | but our understanding of the biology

00:26:31.980 | is so limited at this point,

00:26:34.660 | then seeing what actually happens at the cellular level

00:26:37.960 | is a heck of a lot closer to the human clinical outcome

00:26:41.800 | than looking at the genetics directly,

00:26:44.620 | and so we can learn a lot more from it

00:26:47.180 | than we could by looking at genetics alone.

00:26:49.460 | - So just to get a sense, I don't know if it's easy to do,

00:26:51.660 | but what kind of data is useful

00:26:54.260 | in this disease-in-a-dish model?

00:26:56.220 | Like, what's the source of raw data information?

00:26:59.980 | And also, from my outsider's perspective,

00:27:03.880 | biology and cells are squishy things.

00:27:08.880 | - They are.

00:27:10.020 | - How do you connect-- - They're literally

00:27:10.860 | squishy things.

00:27:11.900 | - How do you connect the computer to that?

00:27:14.560 | Which sensory mechanisms, I guess?

00:27:17.780 | - So that's another one of those revolutions

00:27:20.660 | that have happened in the last 10 years,

00:27:22.540 | in that our ability to measure cells very quantitatively

00:27:27.540 | has also dramatically increased.

00:27:30.360 | So back when I started doing biology,

00:27:32.860 | in the late '90s, early 2000s,

00:27:36.940 | that was the initial era where we started to measure biology

00:27:42.820 | in really quantitative ways, using things like microarrays,

00:27:46.740 | where you would measure, in a single experiment,

00:27:50.900 | the activity level, what's called expression level,

00:27:54.180 | of every gene in the genome in that sample.

00:27:57.340 | And that ability is what actually allowed us

00:28:00.660 | to even understand that there are molecular subtypes

00:28:04.500 | of diseases like cancer, where up until that point,

00:28:07.140 | it's like, oh, you have breast cancer.

00:28:09.540 | But then, when we looked at the molecular data,

00:28:13.480 | it was clear that there's different subtypes

00:28:15.240 | of breast cancer that, at the level of gene activity,

00:28:17.780 | look completely different to each other.

00:28:19.780 | So that was the beginning of this process.

00:28:23.380 | Now we have the ability to measure individual cells

00:28:27.460 | in terms of their gene activity,

00:28:28.860 | using what's called single-cell RNA sequencing,

00:28:31.340 | which basically sequences the RNA,

00:28:35.020 | which is that activity level of different genes

00:28:37.980 | for every gene in the genome.

00:28:40.980 | And you could do that at single-cell level.

00:28:42.740 | So that's an incredibly powerful way of measuring cells.

00:28:45.400 | I mean, you literally count the number of transcripts.

00:28:47.860 | So it really turns that squishy thing

00:28:50.060 | into something that's digital.

00:28:51.820 | Another tremendous data source

00:28:54.340 | that's emerged in the last few years is microscopy,

00:28:57.500 | and specifically even super-resolution microscopy,

00:29:00.580 | where you could use digital reconstruction

00:29:03.460 | to look at subcellular structures,

00:29:06.460 | sometimes even things that are below

00:29:08.380 | the diffraction limit of light

00:29:10.540 | by doing a sophisticated reconstruction.

00:29:13.380 | And again, that gives you tremendous amounts of information

00:29:16.500 | at the subcellular level.

00:29:18.420 | There's now more and more ways

00:29:20.660 | that amazing scientists out there are developing

00:29:24.500 | for getting new types of information from even single cells.

00:29:29.500 | And so that is a way of turning those squishy things

00:29:35.500 | into digital data.

00:29:37.300 | - Into beautiful data sets.

00:29:38.660 | But so that data set then with machine learning tools

00:29:42.580 | allows you to maybe understand the developmental,

00:29:45.860 | like the mechanism of a particular disease.

00:29:49.940 | And if it's possible to sort of at a high level,

00:29:53.700 | describe how does that help lead to drug discovery

00:29:58.700 | that can help prevent, reverse that mechanism?

00:30:05.380 | - So I think there's different ways

00:30:06.780 | in which this data could potentially be used.

00:30:10.420 | Some people use it for scientific discovery

00:30:13.860 | and say, "Oh, look, we see this phenotype

00:30:17.060 | "at the cellular level, so let's try

00:30:21.460 | "and work our way backwards

00:30:22.940 | "and think which genes might be involved

00:30:25.180 | "in pathways that give rise to that."

00:30:27.140 | So that's a very sort of analytical method

00:30:32.140 | to sort of work our way backwards

00:30:35.220 | using our understanding of known biology.

00:30:37.700 | Some people use it in a somewhat more sort of forward,

00:30:42.980 | if that was backward, this would be forward,

00:30:47.580 | which is to say, "Okay, if I can perturb this gene,

00:30:51.100 | "does it show a phenotype that is similar

00:30:54.060 | "to what I see in disease patients?"

00:30:56.100 | And so maybe that gene is actually causal of the disease,

00:30:59.060 | so that's a different way.

00:31:00.260 | And then there's what we do,

00:31:01.660 | which is basically to take that very large collection

00:31:06.300 | of data and use machine learning

00:31:08.340 | to uncover the patterns that emerge from it.

00:31:12.420 | So for instance, what are those subtypes

00:31:14.980 | that might be similar at the human clinical outcome,

00:31:18.700 | but quite distinct when you look at the molecular data?

00:31:21.780 | And then if we can identify such a subtype,

00:31:25.180 | are there interventions that if I apply it

00:31:28.020 | to cells that come from this subtype of the disease,

00:31:32.100 | and you apply that intervention, it could be a drug

00:31:34.940 | or it could be a CRISPR gene intervention,

00:31:38.980 | does it revert the disease state

00:31:41.380 | to something that looks more like

00:31:42.700 | normal, happy, healthy cells?

00:31:44.140 | And so hopefully if you see that,

00:31:46.940 | that gives you a certain hope that that intervention

00:31:51.940 | will also have a meaningful clinical benefit to people.

00:31:55.140 | And there's obviously a bunch of things

00:31:56.620 | that you would wanna do after that to validate that,

00:31:58.740 | but it's a very different and much less hypothesis-driven way

00:32:03.740 | of uncovering new potential interventions

00:32:06.140 | and might give rise to things that are not the same things

00:32:10.100 | that everyone else is already looking at.

00:32:12.500 | - That's, I don't know, I'm just like,

00:32:15.940 | to psychoanalyze my own feeling

00:32:17.380 | about our discussion currently,

00:32:18.700 | it's so exciting to talk about sort of a fundamentally,

00:32:22.260 | well, something that's been turned

00:32:23.780 | into a machine learning problem

00:32:25.900 | and that can have so much real-world impact.

00:32:29.140 | - That's how I feel too.

00:32:30.340 | - That's kind of exciting 'cause I'm so,

00:32:32.220 | most of my day is spent with data sets

00:32:35.740 | that I guess closer to the news groups.

00:32:37.900 | So this is a kind of, it just feels good to talk about.

00:32:41.980 | In fact, I almost don't wanna talk to you

00:32:43.500 | about machine learning.

00:32:45.340 | I wanna talk about the fundamentals of the data set,

00:32:47.460 | which is an exciting place to be.

00:32:50.420 | - I agree with you.

00:32:51.740 | It's what gets me up in the morning.

00:32:53.740 | It's also what attracts a lot of the people

00:32:57.140 | who work at In-sitro to In-sitro

00:32:59.140 | because I think all of the,

00:33:01.660 | certainly all of our machine learning people are outstanding

00:33:04.580 | and could go get a job selling ads online

00:33:08.860 | or doing e-commerce or even self-driving cars.

00:33:12.820 | But I think they would want,

00:33:16.660 | they come to us because they want to work on something

00:33:19.980 | that has more of an aspirational nature

00:33:22.340 | and can really benefit humanity.

00:33:24.660 | - What would these approaches,

00:33:27.540 | what do you hope, what kind of diseases can be helped?

00:33:31.100 | We mentioned Alzheimer's, schizophrenia, type 2 diabetes.

00:33:33.900 | Can you just describe the various kinds of diseases

00:33:36.500 | that this approach can help?

00:33:38.540 | - Well, we don't know.

00:33:39.620 | And I try and be very cautious

00:33:42.500 | about making promises about some things.

00:33:44.780 | Like, oh, we will cure X.

00:33:46.620 | People make that promise.

00:33:48.060 | And I think it's, I try to first deliver and then promise

00:33:52.700 | as opposed to the other way around.

00:33:54.460 | There are characteristics of a disease

00:33:57.300 | that make it more likely that this type of approach

00:34:00.580 | can potentially be helpful.

00:34:02.700 | So for instance, diseases that have

00:34:04.580 | a very strong genetic basis

00:34:07.980 | are ones that are more likely to manifest

00:34:10.940 | in a stem cell derived model.

00:34:12.780 | We would want the cellular models

00:34:16.300 | to be relatively reproducible and robust

00:34:19.980 | so that you could actually get enough of those cells

00:34:24.980 | in a way that isn't very highly variable and noisy.

00:34:29.580 | You would want the disease to be relatively contained

00:34:34.140 | in one or a small number of cell types

00:34:36.700 | that you could actually create in vitro in a dish setting.

00:34:40.980 | Whereas if it's something that's really broad and systemic

00:34:43.460 | and involves multiple cells

00:34:45.540 | that are in very distal parts of your body,

00:34:48.460 | putting that all in the dish is really challenging.

00:34:50.980 | So we want to focus on the ones

00:34:53.740 | that are most likely to be successful today

00:34:56.980 | with the hope, I think,

00:34:58.660 | that really smart bioengineers out there

00:35:03.340 | are developing better and better systems all the time

00:35:06.220 | so that diseases that might not be tractable today

00:35:09.380 | might be tractable in three years.

00:35:11.780 | So for instance, five years ago,

00:35:14.900 | these stem cell derived models didn't really exist.

00:35:16.700 | People were doing most of the work in cancer cells,

00:35:19.100 | and cancer cells are very, very poor models

00:35:22.180 | of most human biology because they're,

00:35:24.820 | A, they were cancer to begin with,

00:35:26.340 | and B, as you passage them and they proliferate in a dish,

00:35:30.700 | they become, because of the genomic instability,

00:35:33.140 | even less similar to human biology.

00:35:36.200 | Now we have these stem cell derived models.

00:35:38.540 | We have the capability to reasonably robustly,

00:35:43.120 | not quite at the right scale yet, but close,

00:35:46.320 | to derive what's called organoids,

00:35:48.440 | which are these teeny little sort of multicellular

00:35:53.160 | organ sort of models of an organ system.

00:35:57.160 | So there's cerebral organoids and liver organoids

00:35:59.760 | and kidney organoids and gut organoids.

00:36:01.720 | - Yeah, brain organoids is possibly

00:36:04.040 | the coolest thing I've ever seen.

00:36:05.480 | - Is that not like the coolest thing?

00:36:07.520 | - Yeah.

00:36:08.360 | - And then I think on the horizon,

00:36:09.960 | we're starting to see things like connecting

00:36:11.800 | these organoids to each other so that you could actually

00:36:14.880 | start, and there's some really cool papers

00:36:16.600 | that start to do that, where you can actually start

00:36:18.840 | to say, okay, can we do multi-organ system stuff?

00:36:22.240 | There's many challenges to that.

00:36:23.560 | It's not easy by any stretch, but it might,

00:36:27.840 | I'm sure people will figure it out,

00:36:29.520 | and in three years or five years,

00:36:31.640 | there will be disease models that we could make

00:36:34.040 | for things that we can't make today.

00:36:35.440 | - Yeah, and this conversation would seem almost outdated

00:36:38.760 | with the kind of scale that could be achieved

00:36:40.520 | in like three years.

00:36:41.360 | - I hope so. - That's the hope.

00:36:42.280 | - That would be so cool.

00:36:43.880 | - So you've co-founded Coursera with Andrew Ng,

00:36:48.120 | and were part of the whole MOOC revolution.

00:36:50.460 | So to jump topics a little bit,

00:36:53.960 | can you maybe tell the origin story of the history,

00:36:57.960 | the origin story of MOOCs, of Coursera,

00:37:00.960 | and in general, your teaching to huge audiences

00:37:05.960 | on a very sort of impactful topic of AI in general?

00:37:12.200 | - So I think the origin story of MOOCs emanates

00:37:16.480 | from a number of efforts that occurred

00:37:19.000 | at Stanford University around the late 2000s,

00:37:24.000 | where different individuals within Stanford,

00:37:28.640 | myself included, were getting really excited

00:37:31.560 | about the opportunities of using online technologies

00:37:35.300 | as a way of achieving both improved quality of teaching

00:37:39.040 | and also improved scale.

00:37:41.040 | And so Andrew, for instance,

00:37:44.520 | led the Stanford Engineering Everywhere,

00:37:48.900 | which was sort of an attempt to take 10 Stanford courses

00:37:51.740 | and put them online, just as video lectures.

00:37:56.080 | I led an effort within Stanford to take some of the courses

00:38:00.640 | and really create a very different teaching model

00:38:04.440 | that broke those up into smaller units

00:38:07.400 | and had some of those embedded interactions and so on,

00:38:11.100 | which got a lot of support from university leaders

00:38:14.640 | because they felt like it was potentially a way

00:38:17.440 | of improving the quality of instruction at Stanford

00:38:19.640 | by moving to what's now called the flipped classroom model.

00:38:23.520 | And so those efforts eventually sort of started

00:38:26.680 | to interplay with each other

00:38:28.080 | and created a tremendous sense of excitement and energy

00:38:31.000 | within the Stanford community

00:38:32.840 | about the potential of online teaching

00:38:36.440 | and led in the fall of 2011

00:38:39.280 | to the launch of the first Stanford MOOCs.

00:38:42.540 | - By the way, MOOCs, it's probably impossible

00:38:46.480 | that people don't know, but I guess massive-

00:38:49.080 | - Open online courses.

00:38:50.280 | - Open online courses.

00:38:51.920 | So the- - We did not come up

00:38:53.080 | with the acronym.

00:38:54.360 | I'm not particularly fond of the acronym,

00:38:57.060 | but it is what it is.

00:38:57.900 | - It is what it is.

00:38:58.720 | Big bang is not a great term for the start of the universe,

00:39:01.400 | but it is what it is.

00:39:02.360 | - Probably so.

00:39:03.580 | (Lex laughing)

00:39:05.280 | So anyway, so those courses launched in the fall of 2011

00:39:10.280 | and there were, within a matter of weeks,

00:39:13.800 | with no real publicity campaign,

00:39:15.920 | just a New York Times article that went viral,

00:39:20.360 | about 100,000 students or more in each of those courses.

00:39:24.600 | And I remember this conversation that Andrew and I had,

00:39:29.240 | which is like, wow, this is just,

00:39:31.680 | there's this real need here.

00:39:33.440 | And I think we both felt like,

00:39:35.880 | sure, we were accomplished academics

00:39:39.000 | and we could go back and go back to our labs,

00:39:41.600 | write more papers, but if we did that,

00:39:43.560 | then this wouldn't happen

00:39:45.840 | and it seemed too important not to happen.

00:39:48.720 | And so we spent a fair bit of time debating,

00:39:51.640 | do we wanna do this as a Stanford effort,

00:39:55.320 | kind of building on what we'd started?

00:39:56.840 | Do we wanna do this as a for-profit company?

00:39:59.360 | Do we wanna do this as a non-profit?

00:40:00.800 | And we decided ultimately to do it as we did with Coursera.

00:40:03.940 | And so we started really operating as a company

00:40:09.920 | at the beginning of 2012.

00:40:12.200 | - And the rest is history.

00:40:14.600 | - And the rest is history.

00:40:15.440 | But how did you, was that really surprising to you?

00:40:18.540 | How did you at that time and at this time

00:40:23.440 | make sense of this need

00:40:26.360 | for sort of global education you mentioned,

00:40:28.120 | that you felt that, wow, the popularity indicates

00:40:31.120 | that there's a hunger for sort of globalization of learning?

00:40:36.120 | - I think there is a hunger for learning that,

00:40:45.000 | globalization is part of it,

00:40:46.280 | but I think it's just a hunger for learning.

00:40:48.400 | The world has changed in the last 50 years.

00:40:51.680 | It used to be that you finished college, you got a job,

00:40:56.120 | by and large, the skills that you learned in college

00:40:58.280 | were pretty much what got you

00:41:00.520 | through the rest of your job history.

00:41:02.520 | And yeah, you learned some stuff,

00:41:04.200 | but it wasn't a dramatic change.

00:41:06.760 | Today, we're in a world where the skills

00:41:09.560 | that you need for a lot of jobs,

00:41:11.480 | they didn't even exist when you went to college

00:41:13.680 | and the jobs and many of the jobs that exist

00:41:15.840 | when you went to college don't even exist today or are dying.

00:41:19.920 | So part of that is due to AI, but not only.

00:41:23.920 | And we need to find a way of keeping people,

00:41:28.600 | giving people access to the skills that they need today.

00:41:31.200 | And I think that's really what's driving

00:41:33.280 | a lot of this hunger.

00:41:35.120 | - So I think if we even take a step back,

00:41:37.980 | for you, all of this started in trying to think of new ways

00:41:42.500 | to teach or new ways to sort of organize the material

00:41:47.500 | and present the material in a way

00:41:49.600 | that would help the education process, the pedagogy.

00:41:52.560 | So what have you learned about effective education

00:41:57.560 | from this process of playing,

00:41:58.720 | of experimenting with different ideas?

00:42:01.720 | - So we learned a number of things,

00:42:04.200 | some of which I think could translate back

00:42:06.840 | and have translated back effectively

00:42:08.600 | to how people teach on campus.

00:42:10.120 | And some of which I think are more specific

00:42:11.920 | to people who learn online,

00:42:14.040 | and more sort of people who learn

00:42:17.120 | as part of their daily life.

00:42:19.120 | So we learned, for instance, very quickly

00:42:21.200 | that short is better.

00:42:23.400 | So people who are especially in the workforce

00:42:27.040 | can't do a 15 week semester long course.

00:42:30.240 | They just can't fit that into their lives.

00:42:32.720 | - Sure, can you describe the shortness of what?

00:42:35.760 | The entirety? - Both.

00:42:38.520 | - Every aspect, so the little lecture,

00:42:40.680 | the lecture's short, the course is short.

00:42:43.240 | - Both.

00:42:44.080 | We started out, the first online education efforts

00:42:48.060 | were actually MIT's OpenCourseWare initiatives,

00:42:50.840 | and that was recording of classroom lectures.

00:42:55.200 | - An hour and a half or something like that, yeah.

00:42:57.880 | - And that didn't really work very well.

00:43:00.400 | I mean, some people benefit, I mean, of course they did,

00:43:03.160 | but it's not really a very palatable experience

00:43:06.760 | for someone who has a job and three kids

00:43:11.280 | and they need to run errands and such.

00:43:14.040 | They can't fit 15 weeks into their life,

00:43:17.920 | and the hour and a half is really hard.

00:43:20.760 | So we learned very quickly,

00:43:23.000 | I mean, we started out with short video modules,

00:43:26.600 | and over time we made them shorter

00:43:28.240 | because we realized that 15 minutes was still too long

00:43:31.720 | if you wanna fit in when you're waiting in line

00:43:33.920 | for your kid's doctor's appointment.

00:43:35.560 | It's better if it's five to seven.

00:43:37.260 | We learned that 15 week courses don't work,

00:43:42.580 | and you really wanna break this up into shorter units

00:43:44.880 | so that there is a natural completion point,

00:43:46.840 | gives people a sense of they're really close

00:43:48.720 | to finishing something meaningful.

00:43:50.480 | They can always come back and take part two and part three.

00:43:53.620 | We also learned that compressing the content

00:43:56.160 | works really well, because if some people,

00:43:58.960 | that pace works well, and for others,

00:44:01.080 | they can always rewind and watch again.

00:44:03.280 | And so people have the ability

00:44:05.360 | to then learn at their own pace.

00:44:07.000 | And so that flexibility,

00:44:10.120 | the brevity and the flexibility are both things

00:44:12.620 | that we found to be very important.

00:44:15.420 | We learned that engagement during the content is important,

00:44:18.800 | and the quicker you give people feedback,

00:44:20.640 | the more likely they are to be engaged.

00:44:22.560 | Hence the introduction of these,

00:44:24.560 | which we actually was an intuition that I had going in

00:44:27.480 | and was then validated using data,

00:44:30.880 | that introducing some of these sort of little micro quizzes

00:44:34.320 | into the lectures really helps.

00:44:36.480 | Self-graded, automatically graded assessments

00:44:39.400 | really help too, because it gives people feedback.

00:44:41.880 | See, there you are.

00:44:43.200 | So all of these are valuable.

00:44:45.600 | And then we learned a bunch of other things too.

00:44:47.240 | We did some really interesting experiments,

00:44:48.920 | for instance, on non-gender bias,

00:44:50.600 | and how having a female role model as an instructor

00:44:55.600 | can change the balance of men to women

00:44:59.320 | in terms of, especially in STEM courses.

00:45:02.040 | And you could do that online by doing A/B testing

00:45:04.800 | in ways that would be really difficult to go on campus.

00:45:07.720 | - Oh, that's exciting.

00:45:09.120 | But so the shortness, the compression, I mean,

00:45:12.020 | it has actually, so that probably is true for all,

00:45:16.440 | you know, good editing is always just compressing

00:45:20.360 | the content, making it shorter.

00:45:21.920 | So that puts a lot of burden on the creator of the,

00:45:24.840 | the instructor and the creator of the educational content.

00:45:28.640 | Probably most lectures at MIT or Stanford

00:45:31.240 | could be five times shorter

00:45:34.360 | if the preparation was put enough.

00:45:37.560 | So maybe people might disagree with that,

00:45:41.680 | but like the crispness, the clarity

00:45:44.040 | that a lot of the, like Coursera delivers

00:45:47.280 | is how much effort does that take?

00:45:50.120 | - So first of all, let me say that it's not clear

00:45:54.120 | that that crispness would work as effectively

00:45:57.360 | in a face-to-face setting,

00:45:58.880 | because people need time to absorb the material.

00:46:02.440 | And so you need to at least pause

00:46:04.760 | and give people a chance to reflect and maybe practice.

00:46:07.300 | And that's what MOOCs do,

00:46:08.400 | is that they give you these chunks of content

00:46:10.800 | and then ask you to practice with it.

00:46:13.440 | And that's where I think some of the newer pedagogy

00:46:16.360 | that people are adopting in face-to-face teaching

00:46:19.200 | that have to do with interactive learning and such

00:46:21.600 | can be really helpful.

00:46:23.520 | But both those approaches,

00:46:26.640 | whether you're doing that type of methodology

00:46:29.440 | in online teaching or in that flipped classroom,

00:46:32.880 | interactive teaching.

00:46:34.520 | - What's, sorry to pause, what's flipped classroom?

00:46:37.200 | - Flipped classroom is a way in which online content

00:46:41.560 | is used to supplement face-to-face teaching,

00:46:45.080 | where people watch the videos perhaps

00:46:47.240 | and do some of the exercises before coming to class.

00:46:49.840 | And then when they come to class,

00:46:51.200 | it's actually to do much deeper problem solving,

00:46:53.580 | oftentimes in a group.

00:46:56.120 | But any one of those different pedagogies

00:47:00.440 | that are beyond just standing there and droning on

00:47:03.480 | in front of the classroom for an hour and 15 minutes

00:47:06.280 | require a heck of a lot more preparation.

00:47:09.220 | And so it's one of the challenges I think that people have,

00:47:13.640 | that we had when trying to convince instructors

00:47:15.720 | to teach on Coursera.

00:47:16.680 | And it's part of the challenges

00:47:18.080 | that pedagogy experts on campus have

00:47:21.080 | in trying to get faculty to teach differently,

00:47:22.800 | is that it's actually harder to teach that way

00:47:24.400 | than it is to stand there and drone.

00:47:26.320 | - Do you think MOOCs will replace in-person education

00:47:32.400 | or become the majority of in-person,

00:47:36.920 | of education of the way people learn in the future?

00:47:41.400 | Again, the future could be very far away,

00:47:43.280 | but where's the trend going, do you think?

00:47:46.040 | - So I think it's a nuanced and complicated answer.

00:47:50.240 | I don't think MOOCs will replace face-to-face teaching.

00:47:55.240 | I think learning is in many cases a social experience.

00:48:00.400 | And even at Coursera,

00:48:02.560 | we had people who naturally formed study groups,

00:48:06.680 | even when they didn't have to,

00:48:07.760 | to just come and talk to each other.

00:48:10.280 | And we found that that actually benefited their learning

00:48:14.440 | in very important ways.

00:48:15.760 | So there was more success among learners

00:48:19.640 | who had those study groups than among ones who didn't.

00:48:22.600 | So I don't think it's just gonna,

00:48:23.840 | oh, we're all gonna just suddenly learn online

00:48:26.040 | with a computer and no one else,

00:48:28.360 | in the same way that recorded music

00:48:30.760 | has not replaced live concerts.

00:48:33.160 | But I do think that especially when you are thinking

00:48:39.000 | about continuing education,

00:48:41.680 | the stuff that people get when their traditional,

00:48:44.760 | whatever high school, college education is done,

00:48:47.880 | and they yet have to maintain their level of expertise

00:48:52.560 | and skills in a rapidly changing world,

00:48:54.680 | I think people will consume more and more educational

00:48:57.160 | content in this online format,

00:48:59.720 | because going back to school for formal education

00:49:02.680 | is not an option for most people.

00:49:04.760 | - Briefly, it might be a difficult question to ask,

00:49:07.440 | but there's a lot of people fascinated

00:49:10.000 | by artificial intelligence, by machine learning,

00:49:12.920 | by deep learning.

00:49:14.040 | Is there a recommendation for the next year

00:49:18.160 | or for a lifelong journey of somebody interested in this?

00:49:21.400 | How do they begin?

00:49:23.760 | How do they enter that learning journey?

00:49:27.280 | - I think the important thing is first to just get started.

00:49:31.640 | And there's plenty of online content that one can get

00:49:36.640 | for both the core foundations of mathematics

00:49:41.280 | and statistics and programming.

00:49:43.040 | And then from there to machine learning.

00:49:45.360 | I would encourage people not to skip too quickly

00:49:48.400 | past the foundations, because I find that there's a lot

00:49:51.360 | of people who learn machine learning,

00:49:53.760 | whether it's online or on campus,

00:49:55.240 | without getting those foundations.

00:49:56.880 | And they basically just turn the crank on existing models

00:50:00.760 | in ways that A, don't allow for a lot of innovation

00:50:04.280 | and adjustment to the problem at hand,

00:50:08.440 | but also B, are sometimes just wrong

00:50:10.360 | and they don't even realize that their application

00:50:12.640 | is wrong because there's artifacts

00:50:14.680 | that they haven't fully understood.

00:50:16.600 | So I think the foundations,

00:50:18.480 | machine learning is an important step.

00:50:20.560 | And then actually start solving problems.

00:50:25.560 | Try and find someone to solve them with,

00:50:27.640 | because especially at the beginning,

00:50:29.000 | it's useful to have someone to bounce ideas off

00:50:31.560 | and fix mistakes that you make.

00:50:33.240 | And you can fix mistakes that they make,

00:50:35.960 | but then just find practical problems,

00:50:40.520 | whether it's in your workplace

00:50:41.920 | or if you don't have that,

00:50:43.280 | Kaggle competitions or such are a really great place

00:50:46.120 | to find interesting problems and just practice.

00:50:50.840 | - Practice.

00:50:52.360 | Perhaps a bit of a romanticized question,

00:50:54.520 | but what idea in deep learning do you find,

00:50:59.280 | have you found in your journey,

00:51:01.120 | the most beautiful or surprising or interesting?

00:51:04.080 | Perhaps not just deep learning,

00:51:09.400 | but AI in general, statistics.

00:51:12.600 | - I'm gonna answer with two things.

00:51:16.680 | One would be the foundational concept of end-to-end training,

00:51:23.040 | which is that you start from the raw data

00:51:26.920 | and you train something that is not like a single piece,

00:51:31.920 | but rather the,

00:51:35.080 | towards the actual goal that you're looking to-

00:51:38.960 | - From the raw data to the outcome,

00:51:40.840 | like no details in between.

00:51:43.560 | - Well, not no details,

00:51:44.640 | but the fact that you, I mean,

00:51:45.680 | you could certainly introduce building blocks

00:51:47.520 | that were trained towards other tasks.

00:51:50.240 | I'm actually coming to that in my second half of the answer,

00:51:53.080 | but it doesn't have to be like a single monolithic blob

00:51:57.760 | in the middle, actually, I think that's not ideal,

00:52:00.240 | but rather the fact that at the end of the day,

00:52:02.640 | you can actually train something

00:52:04.160 | that goes all the way from the beginning to the end.

00:52:06.920 | And the other one that I find really compelling

00:52:09.160 | is the notion of learning a representation

00:52:13.200 | that in its turn, even if it was trained to another task,

00:52:18.200 | can potentially be used as a much more rapid starting point

00:52:23.400 | to solving a different task.

00:52:26.960 | And that's, I think,

00:52:29.040 | reminiscent of what makes people successful learners.

00:52:32.560 | It's something that is relatively new

00:52:35.720 | in the machine learning space.

00:52:36.760 | I think it's underutilized

00:52:38.120 | even relative to today's capabilities,

00:52:40.320 | but more and more of how do we learn

00:52:42.800 | sort of reusable representation.

00:52:45.480 | - So end-to-end and transfer learning.

00:52:49.720 | - Yeah.

00:52:51.120 | - Is it surprising to you that neural networks

00:52:53.640 | are able to, in many cases, do these things?

00:52:57.000 | Is it maybe taken back to when you first would dive deep

00:53:02.000 | into neural networks or in general, even today,

00:53:05.440 | is it surprising that neural networks work at all

00:53:07.840 | and work wonderfully to do this kind of raw end-to-end

00:53:12.840 | learning and even transfer learning?

00:53:16.360 | - I think I was surprised by how well

00:53:21.360 | when you have large enough amounts of data,

00:53:25.800 | it's possible to find a meaningful representation

00:53:32.920 | in what is an exceedingly high dimensional space.

00:53:36.080 | And so I find that to be really exciting

00:53:39.280 | and people are still working on the math for that.

00:53:41.600 | There's more papers on that every year.

00:53:43.560 | And I think it would be really cool if we figured that out.

00:53:48.000 | But that to me was a surprise because in the early days

00:53:53.000 | when I was starting my way in machine learning

00:53:56.200 | and the data sets were rather small,

00:53:58.680 | I think we believed,

00:54:01.160 | I believe that you needed to have a much more constrained

00:54:05.480 | and knowledge rich search space

00:54:08.640 | to really get to a meaningful answer.

00:54:11.840 | And I think it was true at the time.

00:54:13.840 | What I think is still a question is,

00:54:18.840 | will a completely knowledge-free approach

00:54:23.180 | where there's no prior knowledge going

00:54:26.000 | into the construction of the model,

00:54:28.980 | is that gonna be the solution or not?

00:54:31.600 | It's not actually the solution today

00:54:34.160 | in the sense that the architecture

00:54:36.520 | of a convolutional neural network that's used for images

00:54:41.480 | is actually quite different to the type of network

00:54:44.720 | that's used for language and yet different

00:54:47.720 | from the one that's used for speech or biology

00:54:51.160 | or any other application.

00:54:52.480 | There's still some insight that goes into the structure

00:54:57.040 | of the network to get to the right performance.

00:55:00.820 | Will you be able to come up

00:55:01.660 | with a universal learning machine?

00:55:03.260 | I don't know.

00:55:04.100 | - I wonder if there's always has to be

00:55:06.580 | some insight injected somewhere or whether it can converge.

00:55:10.300 | So you've done a lot of interesting work

00:55:13.580 | with probabilistic graphical models

00:55:15.540 | and in general Bayesian deep learning and so on.

00:55:18.440 | Can you maybe speak high level,

00:55:21.060 | how can learning systems deal with uncertainty?

00:55:25.500 | - One of the limitations I think

00:55:28.100 | of a lot of machine learning models

00:55:32.360 | is that they come up with an answer

00:55:35.720 | and you don't know how much you can believe that answer.

00:55:40.720 | And oftentimes the answer is actually

00:55:45.820 | quite poorly calibrated relative to its uncertainties.

00:55:50.540 | Even if you look at where the confidence

00:55:55.460 | that comes out of the, say the neural network at the end

00:55:58.660 | and you ask how much more likely is an answer

00:56:02.760 | of 0.8 versus 0.9, it's not really in any way calibrated

00:56:07.640 | to the actual reliability of that network and how true it is

00:56:12.640 | and the further away you move from the training data,

00:56:16.740 | the more, not only the more wrong the network is,

00:56:20.640 | often it's more wrong and more confident

00:56:22.540 | in its wrong answer.

00:56:24.320 | And that is a serious issue in a lot of application areas.

00:56:29.320 | So when you think for instance about medical diagnosis

00:56:31.640 | as being maybe an epitome of how problematic this can be,

00:56:35.680 | if you were training your network on a certain set

00:56:38.920 | of patients, on a certain patient population,

00:56:41.500 | and I have a patient that is an outlier

00:56:44.600 | and there's no human that looks at this

00:56:46.720 | and that patient is put into a neural network

00:56:49.040 | and your network not only gives

00:56:50.320 | a completely incorrect diagnosis,

00:56:51.920 | but is supremely confident in its wrong answer,

00:56:54.960 | you could kill people.

00:56:56.300 | So I think creating more of an understanding

00:57:01.300 | of how do you produce networks that are calibrated

00:57:06.880 | in their uncertainty and can also say, you know what,

00:57:10.280 | I give up, I don't know what to say

00:57:12.680 | about this particular data instance

00:57:14.560 | because I've never seen something

00:57:16.300 | that's sufficiently like it before.

00:57:18.120 | I think it's going to be really important

00:57:20.480 | in mission critical applications,

00:57:23.000 | especially ones where human life is at stake

00:57:25.360 | and that includes medical applications,

00:57:28.280 | but it also includes automated driving

00:57:31.160 | because you'd want the network to be able to say,

00:57:33.280 | you know what, I have no idea what this blob is

00:57:36.000 | that I'm seeing in the middle of the road,

00:57:37.120 | so I'm just gonna stop because I don't wanna

00:57:39.280 | potentially run over a pedestrian that I don't recognize.

00:57:42.800 | - Is there good mechanisms, ideas of how to allow

00:57:47.520 | learning systems to provide that uncertainty

00:57:52.240 | along with their predictions?

00:57:54.040 | - Certainly people have come up with mechanisms

00:57:57.160 | that involve Bayesian deep learning,

00:58:00.680 | deep learning that involves Gaussian processes.

00:58:04.480 | I mean, there's a slew of different approaches

00:58:07.640 | that people have come up with.

00:58:09.160 | There's methods that use ensembles of networks

00:58:12.920 | with trained with different subsets of data

00:58:15.240 | or different random starting points.

00:58:17.640 | Those are actually sometimes surprisingly good

00:58:20.240 | at creating a sort of set of how confident

00:58:24.040 | or not you are in your answer.

00:58:26.600 | It's very much an area of open research.

00:58:29.000 | - Let's cautiously venture back into the land of philosophy

00:58:33.640 | and speaking of AI systems providing uncertainty,

00:58:37.680 | somebody like Stuart Russell believes that

00:58:41.540 | as we create more and more intelligent systems,

00:58:43.440 | it's really important for them to be full of self-doubt

00:58:46.800 | because if they're given more and more power,

00:58:51.960 | we want the way to maintain human control

00:58:54.840 | over AI systems or human supervision,

00:58:57.120 | which is true, like you just mentioned

00:58:58.600 | with autonomous vehicles, it's really important

00:59:00.400 | to get human supervision when the car is not sure

00:59:04.160 | because if it's really confident,

00:59:05.960 | in cases when it can get in trouble,

00:59:07.840 | it's gonna be really problematic.

00:59:09.360 | So let me ask about sort of the questions of AGI

00:59:13.000 | and human level intelligence.

00:59:14.840 | I mean, we've talked about curing diseases,

00:59:17.180 | which is sort of a fundamental thing

00:59:20.160 | we can have an impact today,

00:59:21.800 | but AI people also dream of both understanding

00:59:26.160 | and creating intelligence.

00:59:29.220 | Is that something you think about?

00:59:30.440 | Is that something you dream about?

00:59:32.800 | Is that something you think is within our reach

00:59:37.000 | to be thinking about as computer scientists?

00:59:41.140 | - Boy, let me tease apart different parts of that question.

00:59:45.180 | - The worst question.

00:59:46.320 | (both laughing)

00:59:48.020 | - Yeah, it's a multi-part question.

00:59:50.940 | So let me start with the feasibility of AGI,

00:59:55.940 | then I'll talk about the timelines a little bit

01:00:01.500 | and then talk about, well, what controls does one need

01:00:05.980 | when protecting, when thinking about protections

01:00:09.460 | in the AI space?

01:00:10.540 | So, I think AGI obviously is a longstanding dream

01:00:15.540 | that even our early pioneers in the space had,

01:00:21.340 | the Turing test and so on are the earliest discussions

01:00:26.260 | of that.

01:00:27.600 | We're obviously closer than we were 70 or so years ago,

01:00:32.600 | but I think it's still very far away.

01:00:37.680 | I think machine learning algorithms today

01:00:41.140 | are really exquisitely good pattern recognizers

01:00:46.140 | in very specific problem domains

01:00:49.700 | where they have seen enough training data

01:00:51.800 | to make good predictions.

01:00:54.000 | You take a machine learning algorithm

01:00:58.120 | and you move it to a slightly different version

01:01:00.900 | of even that same problem, far less one that's different,

01:01:04.000 | and it will just completely choke.

01:01:07.200 | So I think we're nowhere close to the versatility

01:01:11.800 | and flexibility of even a human toddler

01:01:15.800 | in terms of their ability to context switch

01:01:19.920 | and solve different problems

01:01:20.940 | using a single knowledge-based single brain.

01:01:24.520 | So am I desperately worried about the machines

01:01:29.520 | taking over the universe and starting to kill people

01:01:35.480 | because they want to have more power?

01:01:37.360 | I don't think so.

01:01:38.440 | - Well, so to pause on that,

01:01:40.480 | so you've kind of intuited that superintelligence

01:01:43.640 | is a very difficult thing to achieve.

01:01:46.320 | - Even intelligence.

01:01:47.160 | - Intelligence.

01:01:48.160 | - Superintelligence, we're not even close to intelligence.

01:01:50.480 | - Even just the greater abilities of generalization

01:01:53.380 | of our current systems.

01:01:55.180 | But we haven't answered all the parts.

01:01:59.200 | - I'm getting to the second part.

01:02:00.800 | - Okay, we'll take it, but maybe another tangent

01:02:03.360 | you can also pick up is can we get in trouble

01:02:06.780 | with much dumber systems?

01:02:08.160 | - Yes, and that is exactly where I was going.

01:02:10.760 | - Okay.

01:02:11.600 | - So just to wrap up on the threats of AGI,

01:02:16.160 | I think that it seems to me a little early today

01:02:21.160 | to figure out protections against a human level

01:02:26.960 | or superhuman level intelligence

01:02:29.400 | where we don't even see the skeleton

01:02:32.360 | of what that would look like.

01:02:33.920 | So it seems that it's very speculative

01:02:36.440 | on how to protect against that.

01:02:40.520 | But we can definitely and have gotten into trouble

01:02:44.640 | on much dumber systems.

01:02:46.680 | And a lot of that has to do with the fact

01:02:49.040 | that the systems that we're building

01:02:51.360 | are increasingly complex,

01:02:54.680 | increasingly poorly understood,

01:02:57.880 | and there's ripple effects that are unpredictable

01:03:02.000 | in changing little things

01:03:04.200 | that can have dramatic consequences on the outcome.

01:03:08.160 | And by the way, that's not unique

01:03:11.080 | to artificial intelligence.

01:03:11.920 | I think artificial intelligence exacerbates that,

01:03:14.400 | brings it to a new level.

01:03:15.660 | But heck, our electric grid is really complicated.

01:03:18.960 | The software that runs our financial markets

01:03:21.360 | is really complicated.

01:03:23.040 | And we've seen those ripple effects translate

01:03:26.360 | to dramatic negative consequences,

01:03:29.060 | like for instance, financial crashes

01:03:32.360 | that have to do with feedback loops

01:03:34.200 | that we didn't anticipate.

01:03:35.520 | So I think that's an issue that we need

01:03:38.080 | to be thoughtful about in many places,

01:03:41.100 | artificial intelligence being one of them.

01:03:44.800 | And we should, and I think it's really important

01:03:48.480 | that people are thinking about ways

01:03:50.680 | in which we can have better interpretability of systems,

01:03:55.640 | better tests for, for instance,

01:03:59.120 | measuring the extent to which a machine learning system

01:04:01.900 | that was trained in one set of circumstances,

01:04:04.880 | how well does it actually work

01:04:07.360 | in a very different set of circumstances

01:04:09.520 | where you might say, for instance,

01:04:12.320 | well, I'm not gonna be able to test my automated vehicle

01:04:14.760 | in every possible city, village,

01:04:17.840 | weather condition and so on.

01:04:20.760 | But if you trained it on this set of conditions

01:04:23.740 | and then tested it on 50 or 100 others

01:04:27.320 | that were quite different from the ones

01:04:29.160 | that you trained it on, and it worked,

01:04:31.960 | then that gives you confidence that the next 50

01:04:34.100 | that you didn't test it on might also work.

01:04:36.080 | So effectively it's testing for generalizability.

01:04:39.020 | So I think there's ways that we should be

01:04:41.300 | constantly thinking about to validate

01:04:44.500 | the robustness of our systems.

01:04:47.500 | I think it's very different from the,

01:04:50.180 | let's make sure robots don't take over the world.

01:04:53.260 | And then the other place where I think we have a threat,

01:04:57.020 | which is also important for us to think about

01:04:59.400 | is the extent to which technology can be abused.

01:05:03.180 | So like any really powerful technology,

01:05:06.540 | machine learning can be very much used badly

01:05:11.900 | as well as to good.

01:05:13.580 | And that goes back to many other technologies

01:05:16.580 | that have come up with when people invented

01:05:20.220 | projectile missiles and it turned into guns.

01:05:23.080 | And people invented nuclear power

01:05:25.580 | and it turned into nuclear bombs.

01:05:27.380 | And I think, honestly, I would say that to me,

01:05:31.280 | gene editing and CRISPR is at least as dangerous

01:05:34.420 | a technology if used badly as machine learning.

01:05:39.420 | You could create really nasty viruses and such

01:05:43.860 | using gene editing that are,

01:05:48.060 | you know, you would be really careful about.

01:05:51.900 | So anyway, that's something that we need

01:05:56.700 | to be really thoughtful about

01:05:58.640 | whenever we have any really powerful new technology.

01:06:02.500 | - Yeah, and in the case of machine learning

01:06:04.140 | is adversarial machine learning,

01:06:06.820 | so all the kinds of attacks like security almost threats.

01:06:09.140 | And there's a social engineering

01:06:10.540 | with machine learning algorithms.

01:06:12.120 | - And there's face recognition

01:06:13.700 | and big brothers watching you.

01:06:15.900 | And there's the killer drones

01:06:19.660 | that can potentially go and targeted execution

01:06:23.400 | of people in a different country.

01:06:25.240 | I don't, you know, one can argue that bombs

01:06:28.600 | are not necessarily that much better,

01:06:30.420 | but, you know, if people wanna kill someone,

01:06:34.020 | they'll find a way to do it.

01:06:35.740 | - So if you, in general, if you look at trends in the data,

01:06:39.060 | there's less wars, there's less violence,

01:06:41.100 | there's more human rights.

01:06:42.940 | So we've been doing overall quite good as a human species.

01:06:48.340 | - Are you optimistic? - Surprisingly sometimes.

01:06:50.660 | - Are you optimistic?

01:06:52.780 | Maybe another way to ask is,

01:06:54.220 | do you think most people are good

01:06:58.020 | and fundamentally we tend towards a better world,

01:07:03.020 | which is underlying the question,

01:07:05.460 | will machine learning with gene editing

01:07:09.180 | ultimately lend us somewhere good?

01:07:12.140 | Are you optimistic?

01:07:15.860 | - I think by and large, I'm optimistic.

01:07:19.140 | I think that most people mean well.

01:07:24.140 | That doesn't mean that most people are, you know,

01:07:28.020 | altruistic do-gooders, but I think most people mean well.

01:07:32.340 | But I think it's also really important for us as a society

01:07:36.300 | to create social norms where doing good

01:07:42.140 | and being perceived well by our peers

01:07:47.140 | are positively correlated.

01:07:51.060 | I mean, it's very easy to create dysfunctional societies.

01:07:55.140 | There are certainly multiple psychological experiments

01:07:58.100 | as well as sadly real world events

01:08:01.940 | where people have devolved to a world

01:08:04.860 | where being perceived well by your peers

01:08:08.900 | is correlated with really atrocious,

01:08:12.260 | often genocidal behaviors.

01:08:16.400 | So we really want to make sure

01:08:19.060 | that we maintain a set of social norms

01:08:21.300 | where people know that to be a successful member of society,

01:08:25.300 | you want to be doing good.

01:08:27.180 | And one of the things that I sometimes worry about

01:08:31.100 | is that some societies don't seem to necessarily

01:08:35.140 | be moving in the forward direction in that regard

01:08:38.020 | where it's not necessarily the case

01:08:40.100 | that being a good person

01:08:44.780 | is what makes you be perceived well by your peers.

01:08:47.700 | And I think that's a really important thing

01:08:49.420 | for us as a society to remember.

01:08:50.980 | It's very easy to degenerate back into a universe

01:08:55.620 | where it's okay to do really bad stuff

01:09:00.220 | and still have your peers think you're amazing.

01:09:03.040 | - It's fun to ask a world-class computer scientist

01:09:07.820 | and engineer a ridiculously philosophical question

01:09:11.020 | like what is the meaning of life?

01:09:13.140 | Let me ask, what gives your life meaning?

01:09:17.180 | What is the source of fulfillment, happiness, joy, purpose?

01:09:22.180 | - When we were starting Coursera in the fall of 2011,

01:09:30.960 | that was right around the time that Steve Jobs passed away.

01:09:37.740 | And so the media was full of various famous quotes

01:09:41.020 | that he uttered.

01:09:42.540 | And one of them that really stuck with me

01:09:45.500 | because it resonated with stuff that I'd been feeling

01:09:48.780 | for even years before that is that our goal in life

01:09:52.380 | should be to make a dent in the universe.

01:09:55.100 | So I think that to me, what gives my life meaning

01:10:00.100 | is that I would hope that when I am lying there

01:10:05.900 | on my deathbed and looking at what I'd done in my life,

01:10:09.660 | that I can point to ways in which I have left the world

01:10:14.660 | a better place than it was when I entered it.

01:10:20.460 | This is something I tell my kids all the time

01:10:23.600 | because I also think that the burden of that

01:10:27.260 | is much greater for those of us

01:10:29.340 | who were born to privilege.

01:10:31.420 | And in some ways I was, I mean, I wasn't born super wealthy

01:10:34.380 | or anything like that, but I grew up in an educated family

01:10:37.900 | with parents who loved me and took care of me

01:10:40.860 | and I had a chance at a great education

01:10:43.100 | and I always had enough to eat.

01:10:46.660 | So I was in many ways born to privilege

01:10:48.940 | more than the vast majority of humanity.

01:10:51.960 | And my kids, I think, are even more so born to privilege

01:10:55.980 | than I was fortunate enough to be.

01:10:57.960 | And I think it's really important that,

01:11:00.500 | especially for those of us who have that opportunity,

01:11:03.920 | that we use our lives to make the world a better place.

01:11:07.420 | - I don't think there's a better way to end it.

01:11:09.620 | Daphne, it was an honor to talk to you.

01:11:11.620 | Thank you so much for talking to me.

01:11:12.460 | - Thank you.

01:11:13.280 | - Thanks for listening to this conversation

01:11:15.900 | with Daphne Koller.

01:11:17.020 | And thank you to our presenting sponsor, Cash App.

01:11:19.900 | Please consider supporting the podcast

01:11:21.660 | by downloading Cash App and using code LEXPODCAST.

01:11:26.180 | Enjoy this podcast, subscribe on YouTube,

01:11:28.620 | review it with five stars on Apple Podcasts,

01:11:31.060 | support it on Patreon,

01:11:32.400 | simply connect with me on Twitter @LexFriedman.

01:11:36.280 | And now let me leave you with some words from Hippocrates,

01:11:39.800 | a physician from ancient Greece

01:11:41.880 | who's considered to be the father of medicine.

01:11:44.360 | Wherever the art of medicine is loved,

01:11:48.320 | there's also a love of humanity.

01:11:50.760 | Thank you for listening and hope to see you next time.

01:11:54.600 | (upbeat music)

01:11:57.180 | (upbeat music)

01:11:59.760 | [BLANK_AUDIO]

Daphne Koller: Biomedicine and Machine Learning | Lex Fridman Podcast #93

Chapters