Daphne Koller: Biomedicine and Machine Learning
生物与进化AI 与机器学习心理与人性技术与编程音乐与艺术
📋 章节目录
暂无章节信息
🔑 关键词
learningdiseasedatamachinedoncellshumancelldiseaseswaysmodelsstemintelligenceimportantmechanismsonlinebiologybettersystemsused
💬 精彩语录
暂无语录
🎙️ 完整对话（1444 条）
Lex Fridman (00:00.000)
The following is a conversation with Daphne Koller,
以下是与达芙妮·科勒的对话，
Lex Fridman (00:03.300)
a professor of computer science at Stanford University,
斯坦福大学计算机科学教授，
Lex Fridman (00:06.260)
a cofounder of Coursera with Andrew Ng,
与 Andrew Ng 共同创办了 Coursera，
Lex Fridman (00:08.980)
and founder and CEO of Incitro,
以及 Incitro 的创始人兼首席执行官，
Lex Fridman (00:11.880)
a company at the intersection
十字路口的一家公司
Daphne Koller (00:13.380)
of machine learning and biomedicine.
机器学习和生物医学。
Lex Fridman (00:15.940)
We're now in the exciting early days
我们现在正处于激动人心的早期阶段
Daphne Koller (00:17.820)
of using the data driven methods of machine learning
使用数据驱动的机器学习方法
Lex Fridman (00:20.580)
to help discover and develop new drugs
帮助发现和开发新药
Lex Fridman (00:22.580)
and treatments at scale.
和大规模的治疗。
Lex Fridman (00:24.420)
Daphne and Incitro are leading the way on this
Daphne 和 Incitro 在这方面处于领先地位
Daphne Koller (00:27.780)
with breakthroughs that may ripple
突破可能会引发连锁反应
Lex Fridman (00:29.660)
through all fields of medicine,
贯穿医学的各个领域，
Daphne Koller (00:31.620)
including ones most critical for helping
包括对帮助最关键的人
Lex Fridman (00:34.260)
with the current coronavirus pandemic.
随着当前的冠状病毒大流行。
Daphne Koller (00:37.220)
This conversation was recorded
这段对话被录音
Lex Fridman (00:38.660)
before the COVID 19 outbreak.
在 COVID 19 爆发之前。
Daphne Koller (00:41.300)
For everyone feeling the medical, psychological,
对于每个感受到医学、心理、
Lex Fridman (00:43.540)
and financial burden of this crisis,
以及这场危机的财务负担，
Daphne Koller (00:45.620)
I'm sending love your way.
我正在用你的方式传递爱。
Lex Fridman (00:47.700)
Stay strong, we're in this together, we'll beat this thing.
Daphne Koller (00:51.740)
This is the Artificial Intelligence Podcast.
Lex Fridman (00:54.260)
If you enjoy it, subscribe on YouTube,
Daphne Koller (00:56.380)
review it with five stars on Apple Podcast,
Lex Fridman (00:58.720)
support it on Patreon,
Daphne Koller (01:00.100)
or simply connect with me on Twitter
Lex Fridman (01:02.060)
at Lex Friedman, spelled F R I D M A N.
Daphne Koller (01:05.940)
As usual, I'll do a few minutes of ads now
Lex Fridman (01:08.100)
and never any ads in the middle
Daphne Koller (01:09.420)
that can break the flow of this conversation.
Lex Fridman (01:11.740)
I hope that works for you
Lex Fridman (01:13.060)
and doesn't hurt the listening experience.
Lex Fridman (01:15.940)
This show is presented by Cash App,
Daphne Koller (01:17.940)
the number one finance app in the app store.
Lex Fridman (01:20.280)
When you get it, use code LEXPODCAST.
Daphne Koller (01:23.420)
Cash App lets you send money to friends,
Lex Fridman (01:25.620)
buy Bitcoin, and invest in the stock market
Daphne Koller (01:27.900)
with as little as one dollar.
Lex Fridman (01:30.220)
Since Cash App allows you to send
Lex Fridman (01:31.700)
and receive money digitally,
Lex Fridman (01:33.420)
peer to peer, and security in all digital transactions
Daphne Koller (01:36.900)
is very important,
Lex Fridman (01:38.120)
let me mention the PCI data security standard
Daphne Koller (01:41.380)
that Cash App is compliant with.
Lex Fridman (01:43.900)
I'm a big fan of standards for safety and security.
Daphne Koller (01:46.780)
PCI DSS is a good example of that,
Lex Fridman (01:49.520)
where a bunch of competitors got together
Lex Fridman (01:51.140)
and agreed that there needs to be a global standard
Lex Fridman (01:53.860)
around the security of transactions.
Daphne Koller (01:56.020)
Now we just need to do the same for autonomous vehicles
Lex Fridman (01:58.420)
and AI systems in general.
Lex Fridman (02:00.620)
So again, if you get Cash App from the App Store
Lex Fridman (02:03.260)
or Google Play and use the code LEXPODCAST,
Daphne Koller (02:07.060)
you get $10 and Cash App will also donate $10 to FIRST,
Lex Fridman (02:11.220)
an organization that is helping to advance robotics
Lex Fridman (02:14.100)
and STEM education for young people around the world.
Lex Fridman (02:17.700)
And now here's my conversation with Daphne Koller.
Lex Fridman (02:22.420)
So you cofounded Coursera and made a huge impact
Lex Fridman (02:25.040)
in the global education of AI.
Lex Fridman (02:26.660)
And after five years in August, 2016,
Lex Fridman (02:29.700)
wrote a blog post saying that you're stepping away
Lex Fridman (02:33.040)
and wrote, quote,
Lex Fridman (02:34.460)
it is time for me to turn to another critical challenge,
Daphne Koller (02:37.500)
the development of machine learning
Lex Fridman (02:38.940)
and its applications to improving human health.
Lex Fridman (02:41.700)
So let me ask two far out philosophical questions.
Lex Fridman (02:45.140)
One, do you think we'll one day find cures
Lex Fridman (02:48.020)
for all major diseases known today?
Lex Fridman (02:50.760)
And two, do you think we'll one day figure out
Daphne Koller (02:53.560)
a way to extend the human lifespan,
Lex Fridman (02:55.980)
perhaps to the point of immortality?
Lex Fridman (02:59.460)
So one day is a very long time
Lex Fridman (03:01.780)
and I don't like to make predictions
Daphne Koller (03:04.260)
of the type we will never be able to do X
Lex Fridman (03:07.300)
because I think that's a smacks of hubris.
Daphne Koller (03:12.740)
It seems that never in the entire eternity
Lex Fridman (03:16.140)
of human existence will we be able to solve a problem.
Daphne Koller (03:19.380)
That being said, curing disease is very hard
Lex Fridman (03:24.260)
because oftentimes by the time you discover the disease,
Daphne Koller (03:28.540)
a lot of damage has already been done.
Lex Fridman (03:30.560)
And so to assume that we would be able to cure disease
Daphne Koller (03:34.980)
at that stage assumes that we would come up with ways
Lex Fridman (03:37.620)
of basically regenerating entire parts of the human body
Daphne Koller (03:41.940)
in the way that actually returns it to its original state.
Lex Fridman (03:45.340)
And that's a very challenging problem.
Daphne Koller (03:47.420)
We have cured very few diseases.
Lex Fridman (03:49.420)
We've been able to provide treatment
Daphne Koller (03:51.460)
for an increasingly large number,
Lex Fridman (03:52.940)
but the number of things that you could actually define
Daphne Koller (03:54.700)
to be cures is actually not that large.
Lex Fridman (03:59.440)
So I think that there's a lot of work
Daphne Koller (04:02.540)
that would need to happen before one could legitimately say
Lex Fridman (04:05.660)
that we have cured even a reasonable number,
Daphne Koller (04:08.820)
far less all diseases.
Lex Fridman (04:10.460)
On the scale of zero to 100,
Daphne Koller (04:12.780)
where are we in understanding the fundamental mechanisms
Lex Fridman (04:15.580)
of all of major diseases?
Lex Fridman (04:18.140)
What's your sense?
Lex Fridman (04:19.260)
So from the computer science perspective
Daphne Koller (04:21.080)
that you've entered the world of health,
Lex Fridman (04:24.160)
how far along are we?
Daphne Koller (04:26.740)
I think it depends on which disease.
Lex Fridman (04:29.520)
I mean, there are ones where I would say
Daphne Koller (04:31.780)
we're maybe not quite at a hundred
Lex Fridman (04:33.420)
because biology is really complicated
Lex Fridman (04:35.580)
and there's always new things that we uncover
Lex Fridman (04:38.960)
that people didn't even realize existed.
Lex Fridman (04:43.040)
But I would say there's diseases
Lex Fridman (04:44.420)
where we might be in the 70s or 80s,
Lex Fridman (04:48.060)
and then there's diseases in which I would say
Lex Fridman (04:51.340)
with probably the majority where we're really close to zero.
Daphne Koller (04:55.220)
Would Alzheimer's and schizophrenia
Lex Fridman (04:57.980)
and type two diabetes fall closer to zero or to the 80?
Daphne Koller (05:04.340)
I think Alzheimer's is probably closer to zero than to 80.
Lex Fridman (05:11.060)
There are hypotheses,
Lex Fridman (05:12.660)
but I don't think those hypotheses have as of yet
Lex Fridman (05:17.300)
been sufficiently validated that we believe them to be true.
Lex Fridman (05:21.980)
And there is an increasing number of people
Lex Fridman (05:23.780)
who believe that the traditional hypotheses
Daphne Koller (05:25.900)
might not really explain what's going on.
Lex Fridman (05:28.020)
I would also say that Alzheimer's and schizophrenia
Lex Fridman (05:31.700)
and even type two diabetes are not really one disease.
Lex Fridman (05:35.300)
They're almost certainly a heterogeneous collection
Daphne Koller (05:39.380)
of mechanisms that manifest in clinically similar ways.
Lex Fridman (05:43.700)
So in the same way that we now understand
Daphne Koller (05:46.640)
that breast cancer is really not one disease,
Lex Fridman (05:48.900)
it is multitude of cellular mechanisms,
Daphne Koller (05:53.420)
all of which ultimately translate
Lex Fridman (05:55.160)
to uncontrolled proliferation, but it's not one disease.
Daphne Koller (05:59.340)
The same is almost undoubtedly true
Lex Fridman (06:01.140)
for those other diseases as well.
Lex Fridman (06:02.900)
And that understanding that needs to precede
Lex Fridman (06:05.780)
any understanding of the specific mechanisms
Daphne Koller (06:08.460)
of any of those other diseases.
Lex Fridman (06:10.100)
Now, in schizophrenia, I would say
Daphne Koller (06:11.580)
we're almost certainly closer to zero than to anything else.
Lex Fridman (06:15.220)
Type two diabetes is a bit of a mix.
Daphne Koller (06:18.260)
There are clear mechanisms that are implicated
Lex Fridman (06:21.380)
that I think have been validated
Daphne Koller (06:22.980)
that have to do with insulin resistance and such,
Lex Fridman (06:25.260)
but there's almost certainly there as well
Daphne Koller (06:28.500)
many mechanisms that we have not yet understood.
Lex Fridman (06:31.300)
You've also thought and worked a little bit
Daphne Koller (06:34.420)
on the longevity side.
Lex Fridman (06:35.860)
Do you see the disease and longevity as overlapping
Lex Fridman (06:40.260)
completely, partially, or not at all as efforts?
Lex Fridman (06:45.260)
Those mechanisms are certainly overlapping.
Daphne Koller (06:48.620)
There's a well known phenomenon that says
Lex Fridman (06:51.940)
that for most diseases, other than childhood diseases,
Daphne Koller (06:56.820)
the risk for contracting that disease
Lex Fridman (07:01.300)
increases exponentially year on year,
Daphne Koller (07:03.260)
every year from the time you're about 40.
Lex Fridman (07:05.700)
So obviously there's a connection between those two things.
Daphne Koller (07:10.380)
That's not to say that they're identical.
Lex Fridman (07:12.420)
There's clearly aging that happens
Daphne Koller (07:14.980)
that is not really associated with any specific disease.
Lex Fridman (07:18.740)
And there's also diseases and mechanisms of disease
Daphne Koller (07:22.300)
that are not specifically related to aging.
Lex Fridman (07:25.660)
So I think overlap is where we're at.
Daphne Koller (07:29.140)
Okay.
Lex Fridman (07:30.420)
It is a little unfortunate that we get older
Lex Fridman (07:32.620)
and it seems that there's some correlation
Lex Fridman (07:34.180)
with the occurrence of diseases
Daphne Koller (07:39.060)
or the fact that we get older.
Lex Fridman (07:40.780)
And both are quite sad.
Daphne Koller (07:43.100)
I mean, there's processes that happen as cells age
Lex Fridman (07:46.700)
that I think are contributing to disease.
Daphne Koller (07:49.580)
Some of those have to do with DNA damage
Lex Fridman (07:52.780)
that accumulates as cells divide
Daphne Koller (07:54.980)
where the repair mechanisms don't fully correct for those.
Lex Fridman (07:59.620)
There are accumulations of proteins
Daphne Koller (08:03.660)
that are misfolded and potentially aggregate
Lex Fridman (08:06.340)
and those too contribute to disease
Lex Fridman (08:08.540)
and will contribute to inflammation.
Lex Fridman (08:10.540)
There's a multitude of mechanisms that have been uncovered
Daphne Koller (08:14.020)
that are sort of wear and tear at the cellular level
Lex Fridman (08:17.100)
that contribute to disease processes
Lex Fridman (08:21.780)
and I'm sure there's many that we don't yet understand.
Lex Fridman (08:24.860)
On a small tangent and perhaps philosophical,
Daphne Koller (08:30.220)
the fact that things get older
Lex Fridman (08:32.340)
and the fact that things die is a very powerful feature
Daphne Koller (08:36.580)
for the growth of new things.
Lex Fridman (08:38.900)
It's a learning, it's a kind of learning mechanism.
Lex Fridman (08:41.380)
So it's both tragic and beautiful.
Lex Fridman (08:44.660)
So do you, so in trying to fight disease
Lex Fridman (08:52.140)
and trying to fight aging,
Lex Fridman (08:55.260)
do you think about sort of the useful fact of our mortality
Daphne Koller (08:58.940)
or would you, like if you were, could be immortal,
Lex Fridman (09:02.660)
would you choose to be immortal?
Daphne Koller (09:07.140)
Again, I think immortal is a very long time
Lex Fridman (09:10.860)
and I don't know that that would necessarily be something
Daphne Koller (09:16.020)
that I would want to aspire to
Lex Fridman (09:17.900)
but I think all of us aspire to an increased health span,
Daphne Koller (09:24.180)
I would say, which is an increased amount of time
Lex Fridman (09:27.620)
where you're healthy and active
Lex Fridman (09:29.860)
and feel as you did when you were 20
Lex Fridman (09:33.300)
and we're nowhere close to that.
Daphne Koller (09:36.780)
People deteriorate physically and mentally over time
Lex Fridman (09:41.820)
and that is a very sad phenomenon.
Lex Fridman (09:43.660)
So I think a wonderful aspiration would be
Lex Fridman (09:47.300)
if we could all live to the biblical 120 maybe
Daphne Koller (09:52.340)
in perfect health.
Lex Fridman (09:53.740)
In high quality of life.
Daphne Koller (09:54.820)
High quality of life.
Lex Fridman (09:55.860)
I think that would be an amazing goal
Daphne Koller (09:57.780)
for us to achieve as a society
Lex Fridman (09:59.300)
now is the right age 120 or 100 or 150.
Daphne Koller (10:03.660)
I think that's up for debate
Lex Fridman (10:05.740)
but I think an increased health span
Daphne Koller (10:07.660)
is a really worthy goal.
Lex Fridman (10:10.100)
And anyway, in a grand time of the age of the universe,
Daphne Koller (10:14.700)
it's all pretty short.
Lex Fridman (10:16.580)
So from the perspective,
Daphne Koller (10:18.460)
you've done obviously a lot of incredible work
Lex Fridman (10:20.980)
in machine learning.
Lex Fridman (10:22.060)
So what role do you think data and machine learning
Lex Fridman (10:25.140)
play in this goal of trying to understand diseases
Lex Fridman (10:29.300)
and trying to eradicate diseases?
Lex Fridman (10:32.940)
Up until now, I don't think it's played
Daphne Koller (10:35.180)
very much of a significant role
Lex Fridman (10:37.860)
because largely the data sets that one really needed
Daphne Koller (10:42.420)
to enable a powerful machine learning methods,
Lex Fridman (10:47.300)
those data sets haven't really existed.
Daphne Koller (10:49.620)
There's been dribs and drabs
Lex Fridman (10:50.940)
and some interesting machine learning
Daphne Koller (10:53.300)
that has been applied, I would say machine learning
Lex Fridman (10:55.700)
slash data science,
Lex Fridman (10:57.660)
but the last few years are starting to change that.
Lex Fridman (11:00.180)
So we now see an increase in some large data sets
Lex Fridman (11:06.300)
but equally importantly, an increase in technologies
Lex Fridman (11:11.340)
that are able to produce data at scale.
Daphne Koller (11:14.700)
It's not typically the case that people have deliberately
Lex Fridman (11:19.340)
proactively used those tools
Daphne Koller (11:21.420)
for the purpose of generating data for machine learning.
Lex Fridman (11:24.180)
They, to the extent that those techniques
Daphne Koller (11:26.540)
have been used for data production,
Lex Fridman (11:28.540)
they've been used for data production
Daphne Koller (11:29.860)
to drive scientific discovery
Lex Fridman (11:31.300)
and the machine learning came as a sort of byproduct
Daphne Koller (11:34.420)
second stage of, oh, you know, now we have a data set,
Lex Fridman (11:36.900)
let's do machine learning on that
Daphne Koller (11:38.260)
rather than a more simplistic data analysis method.
Lex Fridman (11:41.820)
But what we are doing in Citro
Daphne Koller (11:44.420)
is actually flipping that around and saying,
Lex Fridman (11:46.780)
here's this incredible repertoire of methods
Daphne Koller (11:50.300)
that bioengineers, cell biologists have come up with,
Lex Fridman (11:54.580)
let's see if we can put them together in brand new ways
Daphne Koller (11:57.420)
with the goal of creating data sets
Lex Fridman (12:00.260)
that machine learning can really be applied on productively
Daphne Koller (12:03.380)
to create powerful predictive models
Lex Fridman (12:06.580)
that can help us address fundamental problems
Daphne Koller (12:08.460)
in human health.
Lex Fridman (12:09.420)
So really focus to get, make data the primary focus
Lex Fridman (12:14.500)
and the primary goal and find,
Lex Fridman (12:16.460)
use the mechanisms of biology and chemistry
Daphne Koller (12:18.900)
to create the kinds of data set
Lex Fridman (12:23.340)
that could allow machine learning to benefit the most.
Daphne Koller (12:25.700)
I wouldn't put it in those terms
Lex Fridman (12:27.580)
because that says that data is the end goal.
Daphne Koller (12:30.460)
Data is the means.
Lex Fridman (12:32.140)
So for us, the end goal is helping address challenges
Daphne Koller (12:35.740)
in human health and the method that we've elected to do that
Lex Fridman (12:39.980)
is to apply machine learning to build predictive models
Lex Fridman (12:44.140)
and machine learning, in my opinion,
Lex Fridman (12:45.980)
can only be really successfully applied
Daphne Koller (12:48.820)
especially the more powerful models
Lex Fridman (12:50.700)
if you give it data that is of sufficient scale
Lex Fridman (12:53.540)
and sufficient quality.
Lex Fridman (12:54.540)
So how do you create those data sets
Lex Fridman (12:58.580)
so as to drive the ability to generate predictive models
Lex Fridman (13:03.700)
which subsequently help improve human health?
Lex Fridman (13:05.740)
So before we dive into the details of that,
Lex Fridman (13:08.700)
let me take a step back and ask when and where
Lex Fridman (13:13.820)
was your interest in human health born?
Lex Fridman (13:16.780)
Are there moments, events, perhaps if I may ask,
Daphne Koller (13:19.900)
tragedies in your own life that catalyzes passion
Lex Fridman (13:23.060)
or was it the broader desire to help humankind?
Lex Fridman (13:26.580)
So I would say it's a bit of both.
Lex Fridman (13:29.180)
So on, I mean, my interest in human health
Daphne Koller (13:32.620)
actually dates back to the early 2000s
Lex Fridman (13:37.780)
when a lot of my peers in machine learning
Lex Fridman (13:43.940)
and I were using data sets
Lex Fridman (13:45.500)
that frankly were not very inspiring.
Daphne Koller (13:47.420)
Some of us old timers still remember
Lex Fridman (13:49.820)
the quote unquote 20 news groups data set
Daphne Koller (13:52.300)
where this was literally a bunch of texts
Lex Fridman (13:55.740)
from 20 news groups,
Daphne Koller (13:57.100)
a concept that doesn't really even exist anymore.
Lex Fridman (13:59.260)
And the question was, can you classify
Lex Fridman (14:01.660)
which news group a particular bag of words came from?
Lex Fridman (14:06.780)
And it wasn't very interesting.
Daphne Koller (14:08.700)
The data sets at the time on the biology side
Lex Fridman (14:12.460)
were much more interesting,
Daphne Koller (14:14.020)
both from a technical and also from
Lex Fridman (14:15.540)
an aspirational perspective.
Daphne Koller (14:17.540)
They were still pretty small,
Lex Fridman (14:18.860)
but they were better than 20 news groups.
Lex Fridman (14:20.740)
And so I started out, I think just by wanting
Lex Fridman (14:25.620)
to do something that was more, I don't know,
Daphne Koller (14:27.860)
societally useful and technically interesting.
Lex Fridman (14:30.780)
And then over time became more and more interested
Daphne Koller (14:34.420)
in the biology and the human health aspects for themselves
Lex Fridman (14:40.220)
and began to work even sometimes on papers
Daphne Koller (14:43.460)
that were just in biology
Lex Fridman (14:45.140)
without having a significant machine learning component.
Daphne Koller (14:48.460)
I think my interest in drug discovery
Lex Fridman (14:52.740)
is partly due to an incident I had with
Daphne Koller (14:58.580)
when my father sadly passed away about 12 years ago.
Lex Fridman (15:02.580)
He had an autoimmune disease that settled in his lungs
Lex Fridman (15:08.900)
and the doctors basically said,
Lex Fridman (15:11.460)
well, there's only one thing that we could do,
Daphne Koller (15:13.380)
which is give him prednisone.
Lex Fridman (15:15.020)
At some point, I remember a doctor even came and said,
Daphne Koller (15:17.780)
hey, let's do a lung biopsy to figure out
Lex Fridman (15:19.620)
which autoimmune disease he has.
Lex Fridman (15:20.940)
And I said, would that be helpful?
Lex Fridman (15:23.180)
Would that change treatment?
Daphne Koller (15:24.020)
He said, no, there's only prednisone.
Lex Fridman (15:25.500)
That's the only thing we can give him.
Lex Fridman (15:27.180)
And I had friends who were rheumatologists who said
Lex Fridman (15:29.900)
the FDA would never approve prednisone today
Daphne Koller (15:32.060)
because the ratio of side effects to benefit
Lex Fridman (15:37.260)
is probably not large enough.
Daphne Koller (15:39.580)
Today, we're in a state where there's probably four or five,
Lex Fridman (15:44.860)
maybe even more, well, it depends for which autoimmune disease,
Lex Fridman (15:48.740)
but there are multiple drugs that can help people
Lex Fridman (15:52.940)
with autoimmune disease,
Daphne Koller (15:53.980)
many of which didn't exist 12 years ago.
Lex Fridman (15:56.740)
And I think we're at a golden time in some ways
Daphne Koller (16:00.380)
in drug discovery where there's the ability to create drugs
Lex Fridman (16:05.380)
that are much more safe and much more effective
Daphne Koller (16:10.580)
than we've ever been able to before.
Lex Fridman (16:13.060)
And what's lacking is enough understanding
Daphne Koller (16:16.340)
of biology and mechanism to know where to aim that engine.
Lex Fridman (16:22.300)
And I think that's where machine learning can help.
Lex Fridman (16:25.380)
So in 2018, you started and now lead a company in Citro,
Lex Fridman (16:29.900)
which is, like you mentioned,
Daphne Koller (16:32.580)
perhaps the focus is drug discovery
Lex Fridman (16:34.740)
and the utilization of machine learning for drug discovery.
Lex Fridman (16:38.140)
So you mentioned that, quote,
Lex Fridman (16:40.620)
we're really interested in creating
Lex Fridman (16:42.100)
what you might call a disease in a dish model,
Lex Fridman (16:45.580)
disease in a dish models,
Daphne Koller (16:47.380)
places where diseases are complex,
Lex Fridman (16:49.180)
where we really haven't had a good model system,
Daphne Koller (16:52.220)
where typical animal models that have been used for years,
Lex Fridman (16:55.020)
including testing on mice, just aren't very effective.
Lex Fridman (16:58.860)
So can you try to describe what is an animal model
Lex Fridman (17:02.640)
and what is a disease in a dish model?
Daphne Koller (17:05.340)
Sure.
Lex Fridman (17:06.260)
So an animal model for disease
Daphne Koller (17:09.300)
is where you create effectively,
Lex Fridman (17:13.860)
it's what it sounds like.
Daphne Koller (17:14.900)
It's oftentimes a mouse where we have introduced
Lex Fridman (17:19.300)
some external perturbation that creates the disease
Lex Fridman (17:22.780)
and then we cure that disease.
Lex Fridman (17:26.300)
And the hope is that by doing that,
Daphne Koller (17:28.740)
we will cure a similar disease in the human.
Lex Fridman (17:31.340)
The problem is that oftentimes
Daphne Koller (17:33.500)
the way in which we generate the disease in the animal
Lex Fridman (17:36.900)
has nothing to do with how that disease
Daphne Koller (17:38.560)
actually comes about in a human.
Lex Fridman (17:40.900)
It's what you might think of as a copy of the phenotype,
Daphne Koller (17:44.420)
a copy of the clinical outcome,
Lex Fridman (17:46.740)
but the mechanisms are quite different.
Lex Fridman (17:48.740)
And so curing the disease in the animal,
Lex Fridman (17:52.120)
which in most cases doesn't happen naturally,
Daphne Koller (17:54.880)
mice don't get Alzheimer's, they don't get diabetes,
Lex Fridman (17:57.180)
they don't get atherosclerosis,
Daphne Koller (17:58.700)
they don't get autism or schizophrenia.
Lex Fridman (18:02.580)
Those cures don't translate over
Daphne Koller (18:05.700)
to what happens in the human.
Lex Fridman (18:08.140)
And that's where most drugs fails
Daphne Koller (18:10.860)
just because the findings that we had in the mouse
Lex Fridman (18:13.700)
don't translate to a human.
Daphne Koller (18:16.660)
The disease in the dish models is a fairly new approach.
Lex Fridman (18:20.860)
It's been enabled by technologies
Daphne Koller (18:24.140)
that have not existed for more than five to 10 years.
Lex Fridman (18:28.420)
So for instance, the ability for us to take a cell
Daphne Koller (18:32.780)
from any one of us, you or me,
Lex Fridman (18:35.540)
revert that say skin cell to what's called stem cell status,
Daphne Koller (18:39.960)
which is what's called the pluripotent cell
Lex Fridman (18:44.740)
that can then be differentiated
Daphne Koller (18:46.600)
into different types of cells.
Lex Fridman (18:47.860)
So from that pluripotent cell,
Daphne Koller (18:49.800)
one can create a Lex neuron or a Lex cardiomyocyte
Lex Fridman (18:54.300)
or a Lex hepatocyte that has your genetics,
Lex Fridman (18:57.760)
but that right cell type.
Lex Fridman (19:00.020)
And so if there's a genetic burden of disease
Daphne Koller (19:04.780)
that would manifest in that particular cell type,
Lex Fridman (19:07.180)
you might be able to see it by looking at those cells
Lex Fridman (19:10.300)
and saying, oh, that's what potentially sick cells
Lex Fridman (19:13.380)
look like versus healthy cells
Lex Fridman (19:15.620)
and then explore what kind of interventions
Lex Fridman (19:20.740)
might revert the unhealthy looking cell to a healthy cell.
Daphne Koller (19:24.860)
Now, of course, curing cells is not the same
Lex Fridman (19:27.740)
as curing people.
Lex Fridman (19:29.820)
And so there's still potentially a translatability gap,
Lex Fridman (19:33.220)
but at least for diseases that are driven,
Daphne Koller (19:38.500)
say by human genetics and where the human genetics
Lex Fridman (19:41.980)
is what drives the cellular phenotype,
Daphne Koller (19:43.780)
there is some reason to hope that if we revert those cells
Lex Fridman (19:47.960)
in which the disease begins
Lex Fridman (19:49.600)
and where the disease is driven by genetics
Lex Fridman (19:52.180)
and we can revert that cell back to a healthy state,
Daphne Koller (19:55.260)
maybe that will help also revert
Lex Fridman (19:58.140)
the more global clinical phenotype.
Lex Fridman (1:00:01.500)
and then talk about, well, what controls does one need
Lex Fridman (1:00:05.980)
when thinking about protections in the AI space?
Daphne Koller (1:00:10.540)
So, I think AGI obviously is a longstanding dream
Lex Fridman (1:00:17.180)
that even our early pioneers in the space had,
Daphne Koller (1:00:21.300)
the Turing test and so on
Lex Fridman (1:00:23.460)
are the earliest discussions of that.
Daphne Koller (1:00:27.580)
We're obviously closer than we were 70 or so years ago,
Lex Fridman (1:00:32.580)
but I think it's still very far away.
Daphne Koller (1:00:37.420)
I think machine learning algorithms today
Lex Fridman (1:00:40.900)
are really exquisitely good pattern recognizers
Daphne Koller (1:00:46.180)
in very specific problem domains
Lex Fridman (1:00:49.420)
where they have seen enough training data
Daphne Koller (1:00:51.540)
to make good predictions.
Lex Fridman (1:00:53.740)
You take a machine learning algorithm
Lex Fridman (1:00:57.860)
and you move it to a slightly different version
Lex Fridman (1:01:00.660)
of even that same problem, far less one that's different
Lex Fridman (1:01:03.780)
and it will just completely choke.
Lex Fridman (1:01:06.980)
So I think we're nowhere close to the versatility
Lex Fridman (1:01:11.620)
and flexibility of even a human toddler
Lex Fridman (1:01:15.620)
in terms of their ability to context switch
Lex Fridman (1:01:19.740)
and solve different problems
Lex Fridman (1:01:20.740)
using a single knowledge base, single brain.
Lex Fridman (1:01:24.340)
So am I desperately worried about
Lex Fridman (1:01:28.820)
the machines taking over the universe
Lex Fridman (1:01:33.540)
and starting to kill people
Lex Fridman (1:01:35.500)
because they want to have more power?
Daphne Koller (1:01:37.380)
I don't think so.
Lex Fridman (1:01:38.460)
Well, so to pause on that,
Lex Fridman (1:01:40.460)
so you kind of intuited that super intelligence
Lex Fridman (1:01:43.620)
is a very difficult thing to achieve.
Daphne Koller (1:01:46.300)
Even intelligence.
Lex Fridman (1:01:47.140)
Intelligence, intelligence.
Daphne Koller (1:01:48.180)
Super intelligence, we're not even close to intelligence.
Lex Fridman (1:01:50.500)
Even just the greater abilities of generalization
Daphne Koller (1:01:53.380)
of our current systems.
Lex Fridman (1:01:55.180)
But we haven't answered all the parts
Lex Fridman (1:01:59.180)
and we'll take another.
Lex Fridman (1:02:00.020)
I'm getting to the second part.
Daphne Koller (1:02:00.860)
Okay, but maybe another tangent you can also pick up
Lex Fridman (1:02:04.340)
is can we get in trouble with much dumber systems?
Daphne Koller (1:02:08.140)
Yes, and that is exactly where I was going.
Lex Fridman (1:02:11.300)
So just to wrap up on the threats of AGI,
Daphne Koller (1:02:16.140)
I think that it seems to me a little early today
Lex Fridman (1:02:21.140)
to figure out protections against a human level
Daphne Koller (1:02:26.220)
or superhuman level intelligence
Lex Fridman (1:02:28.620)
where we don't even see the skeleton
Daphne Koller (1:02:31.580)
of what that would look like.
Lex Fridman (1:02:33.140)
So it seems that it's very speculative
Daphne Koller (1:02:35.740)
on how to protect against that.
Lex Fridman (1:02:39.820)
But we can definitely and have gotten into trouble
Daphne Koller (1:02:43.940)
on much dumber systems.
Lex Fridman (1:02:45.980)
And a lot of that has to do with the fact
Daphne Koller (1:02:48.340)
that the systems that we're building are increasingly
Lex Fridman (1:02:52.300)
complex, increasingly poorly understood.
Lex Fridman (1:02:57.380)
And there's ripple effects that are unpredictable
Lex Fridman (1:03:01.420)
in changing little things that can have dramatic consequences
Daphne Koller (1:03:06.420)
on the outcome.
Lex Fridman (1:03:08.460)
And by the way, that's not unique to artificial intelligence.
Daphne Koller (1:03:11.620)
I think artificial intelligence exacerbates that,
Lex Fridman (1:03:13.820)
brings it to a new level.
Lex Fridman (1:03:15.100)
But heck, our electric grid is really complicated.
Lex Fridman (1:03:18.420)
The software that runs our financial markets
Daphne Koller (1:03:20.820)
is really complicated.
Lex Fridman (1:03:22.540)
And we've seen those ripple effects translate
Daphne Koller (1:03:25.820)
to dramatic negative consequences,
Lex Fridman (1:03:28.540)
like for instance, financial crashes that have to do
Daphne Koller (1:03:32.820)
with feedback loops that we didn't anticipate.
Lex Fridman (1:03:35.020)
So I think that's an issue that we need to be thoughtful
Daphne Koller (1:03:38.460)
about in many places,
Lex Fridman (1:03:41.940)
artificial intelligence being one of them.
Lex Fridman (1:03:44.300)
And I think it's really important that people are thinking
Lex Fridman (1:03:49.660)
about ways in which we can have better interpretability
Daphne Koller (1:03:54.380)
of systems, better tests for, for instance,
Lex Fridman (1:03:59.140)
measuring the extent to which a machine learning system
Daphne Koller (1:04:01.900)
that was trained in one set of circumstances,
Lex Fridman (1:04:04.860)
how well does it actually work
Daphne Koller (1:04:07.340)
in a very different set of circumstances
Lex Fridman (1:04:09.540)
where you might say, for instance,
Daphne Koller (1:04:12.340)
well, I'm not gonna be able to test my automated vehicle
Lex Fridman (1:04:14.740)
in every possible city, village,
Daphne Koller (1:04:18.980)
weather condition and so on.
Lex Fridman (1:04:20.780)
But if you trained it on this set of conditions
Lex Fridman (1:04:23.740)
and then tested it on 50 or a hundred others
Lex Fridman (1:04:27.340)
that were quite different from the ones
Daphne Koller (1:04:29.140)
that you trained it on and it worked,
Lex Fridman (1:04:31.980)
then that gives you confidence that the next 50
Daphne Koller (1:04:34.100)
that you didn't test it on might also work.
Lex Fridman (1:04:36.100)
So effectively it's testing for generalizability.
Lex Fridman (1:04:39.020)
So I think there's ways that we should be
Lex Fridman (1:04:41.300)
constantly thinking about to validate the robustness
Daphne Koller (1:04:45.900)
of our systems.
Lex Fridman (1:04:47.500)
I think it's very different from the let's make sure
Daphne Koller (1:04:50.980)
robots don't take over the world.
Lex Fridman (1:04:53.260)
And then the other place where I think we have a threat,
Daphne Koller (1:04:57.020)
which is also important for us to think about
Lex Fridman (1:04:59.420)
is the extent to which technology can be abused.
Lex Fridman (1:05:03.180)
So like any really powerful technology,
Lex Fridman (1:05:06.540)
machine learning can be very much used badly
Daphne Koller (1:05:10.900)
as well as to good.
Lex Fridman (1:05:12.700)
And that goes back to many other technologies
Daphne Koller (1:05:15.580)
that have come up with when people invented
Lex Fridman (1:05:19.140)
projectile missiles and it turned into guns
Lex Fridman (1:05:22.140)
and people invented nuclear power
Lex Fridman (1:05:24.660)
and it turned into nuclear bombs.
Lex Fridman (1:05:26.420)
And I think honestly, I would say that to me,
Lex Fridman (1:05:30.340)
gene editing and CRISPR is at least as dangerous
Daphne Koller (1:05:33.500)
as technology if used badly than as machine learning.
Lex Fridman (1:05:39.780)
You could create really nasty viruses and such
Daphne Koller (1:05:43.860)
using gene editing that you would be really careful about.
Lex Fridman (1:05:51.900)
So anyway, that's something that we need
Daphne Koller (1:05:56.700)
to be really thoughtful about whenever we have
Lex Fridman (1:05:59.620)
any really powerful new technology.
Daphne Koller (1:06:02.500)
Yeah, and in the case of machine learning
Lex Fridman (1:06:04.140)
is adversarial machine learning.
Lex Fridman (1:06:06.820)
So all the kinds of attacks like security almost threats
Lex Fridman (1:06:09.140)
and there's a social engineering
Daphne Koller (1:06:10.540)
with machine learning algorithms.
Lex Fridman (1:06:12.100)
And there's face recognition and big brother is watching you
Lex Fridman (1:06:15.900)
and there's the killer drones that can potentially go
Lex Fridman (1:06:20.980)
and targeted execution of people in a different country.
Daphne Koller (1:06:27.180)
One can argue that bombs are not necessarily
Lex Fridman (1:06:29.620)
that much better, but people wanna kill someone,
Daphne Koller (1:06:34.020)
they'll find a way to do it.
Lex Fridman (1:06:35.740)
So in general, if you look at trends in the data,
Daphne Koller (1:06:39.060)
there's less wars, there's less violence,
Lex Fridman (1:06:41.100)
there's more human rights.
Lex Fridman (1:06:42.940)
So we've been doing overall quite good as a human species.
Lex Fridman (1:06:48.340)
Are you optimistic?
Daphne Koller (1:06:49.180)
Surprisingly sometimes.
Lex Fridman (1:06:50.620)
Are you optimistic?
Daphne Koller (1:06:52.740)
Maybe another way to ask is do you think most people
Lex Fridman (1:06:55.540)
are good and fundamentally we tend towards a better world,
Daphne Koller (1:07:03.140)
which is underlying the question,
Lex Fridman (1:07:05.460)
will machine learning with gene editing
Lex Fridman (1:07:09.180)
ultimately land us somewhere good?
Lex Fridman (1:07:12.140)
Are you optimistic?
Daphne Koller (1:07:15.860)
I think by and large, I'm optimistic.
Lex Fridman (1:07:19.140)
I think that most people mean well,
Daphne Koller (1:07:24.140)
that doesn't mean that most people are altruistic do gooders,
Lex Fridman (1:07:28.140)
but I think most people mean well,
Lex Fridman (1:07:31.020)
but I think it's also really important for us as a society
Lex Fridman (1:07:34.980)
to create social norms where doing good
Lex Fridman (1:07:40.820)
and being perceived well by our peers
Lex Fridman (1:07:47.140)
are positively correlated.
Daphne Koller (1:07:49.780)
I mean, it's very easy to create dysfunctional norms
Lex Fridman (1:07:54.060)
in emotional societies.
Daphne Koller (1:07:55.620)
There's certainly multiple psychological experiments
Lex Fridman (1:07:58.540)
as well as sadly real world events
Daphne Koller (1:08:02.420)
where people have devolved to a world
Lex Fridman (1:08:05.300)
where being perceived well by your peers
Daphne Koller (1:08:09.340)
is correlated with really atrocious,
Lex Fridman (1:08:14.100)
often genocidal behaviors.
Lex Fridman (1:08:17.820)
So we really want to make sure
Lex Fridman (1:08:19.500)
that we maintain a set of social norms
Daphne Koller (1:08:21.740)
where people know that to be a successful member of society,
Lex Fridman (1:08:25.660)
you want to be doing good.
Lex Fridman (1:08:27.500)
And one of the things that I sometimes worry about
Lex Fridman (1:08:31.420)
is that some societies don't seem to necessarily
Daphne Koller (1:08:35.420)
be moving in the forward direction in that regard
Lex Fridman (1:08:38.340)
where it's not necessarily the case
Daphne Koller (1:08:43.620)
that being a good person
Lex Fridman (1:08:45.100)
is what makes you be perceived well by your peers.
Lex Fridman (1:08:47.980)
And I think that's a really important thing
Lex Fridman (1:08:49.700)
for us as a society to remember.
Daphne Koller (1:08:51.300)
It's really easy to degenerate back into a universe
Lex Fridman (1:08:55.940)
where it's okay to do really bad stuff
Lex Fridman (1:09:00.540)
and still have your peers think you're amazing.
Lex Fridman (1:09:04.980)
It's fun to ask a world class computer scientist
Lex Fridman (1:09:08.180)
and engineer a ridiculously philosophical question
Lex Fridman (1:09:11.380)
like what is the meaning of life?
Lex Fridman (1:09:13.460)
Let me ask, what gives your life meaning?
Lex Fridman (1:09:17.500)
Or what is the source of fulfillment, happiness,
Lex Fridman (1:09:22.180)
joy, purpose?
Lex Fridman (1:09:26.540)
When we were starting Coursera in the fall of 2011,
Daphne Koller (1:09:32.980)
that was right around the time that Steve Jobs passed away.
Lex Fridman (1:09:37.740)
And so the media was full of various famous quotes
Daphne Koller (1:09:41.020)
that he uttered and one of them that really stuck with me
Lex Fridman (1:09:45.500)
because it resonated with stuff that I'd been feeling
Daphne Koller (1:09:48.780)
for even years before that is that our goal in life
Lex Fridman (1:09:52.380)
should be to make a dent in the universe.
Lex Fridman (1:09:55.100)
So I think that to me, what gives my life meaning
Lex Fridman (1:10:00.620)
is that I would hope that when I am lying there
Daphne Koller (1:10:05.900)
on my deathbed and looking at what I'd done in my life
Lex Fridman (1:10:09.660)
that I can point to ways in which I have left the world
Daphne Koller (1:10:15.860)
a better place than it was when I entered it.
Lex Fridman (1:10:20.460)
This is something I tell my kids all the time
Daphne Koller (1:10:23.620)
because I also think that the burden of that
Lex Fridman (1:10:27.260)
is much greater for those of us who were born to privilege.
Lex Fridman (1:10:31.420)
And in some ways I was, I mean, I wasn't born super wealthy
Lex Fridman (1:10:34.380)
or anything like that, but I grew up in an educated family
Daphne Koller (1:10:37.900)
with parents who loved me and took care of me
Lex Fridman (1:10:40.860)
and I had a chance at a great education
Lex Fridman (1:10:43.060)
and I always had enough to eat.
Lex Fridman (1:10:46.620)
So I was in many ways born to privilege
Daphne Koller (1:10:48.900)
more than the vast majority of humanity.
Lex Fridman (1:10:51.940)
And my kids I think are even more so born to privilege
Daphne Koller (1:10:55.940)
than I was fortunate enough to be.
Lex Fridman (1:10:57.940)
And I think it's really important that especially
Daphne Koller (1:11:01.020)
for those of us who have that opportunity
Lex Fridman (1:11:03.900)
that we use our lives to make the world a better place.
Daphne Koller (1:11:07.420)
I don't think there's a better way to end it.
Lex Fridman (1:11:09.620)
Daphne, it was an honor to talk to you.
Daphne Koller (1:11:11.620)
Thank you so much for talking today.
Lex Fridman (1:11:12.620)
Thank you.
Daphne Koller (1:11:14.420)
Thanks for listening to this conversation
Lex Fridman (1:11:15.900)
with Daphne Koller and thank you
Daphne Koller (1:11:17.780)
to our presenting sponsor, Cash App.
Lex Fridman (1:11:19.900)
Please consider supporting the podcast
Daphne Koller (1:11:21.660)
by downloading Cash App and using code LEXPodcast.
Lex Fridman (1:11:26.180)
If you enjoy this podcast, subscribe on YouTube,
Daphne Koller (1:11:28.620)
review it with five stars on Apple Podcast,
Lex Fridman (1:11:31.060)
support it on Patreon, or simply connect with me
Daphne Koller (1:11:33.340)
on Twitter at LEXFREEDMAN.
Lex Fridman (1:11:36.260)
And now let me leave you with some words from Hippocrates,
Daphne Koller (1:11:39.820)
a physician from ancient Greece
Lex Fridman (1:11:41.900)
who's considered to be the father of medicine.
Daphne Koller (1:11:45.340)
Wherever the art of medicine is loved,
Lex Fridman (1:11:48.340)
there's also a love of humanity.
Daphne Koller (1:11:50.780)
Thank you for listening and hope to see you next time.
Lex Fridman (20:00.860)
So that's really what we're hoping to do.
Daphne Koller (20:02.740)
That step, that backward step, I was reading about it,
Lex Fridman (20:06.020)
the Yamanaka factor.
Daphne Koller (20:08.300)
Yes.
Lex Fridman (20:09.700)
So it's like that reverse step back to stem cells.
Daphne Koller (20:12.280)
Yes.
Lex Fridman (20:13.120)
Seems like magic.
Daphne Koller (20:14.180)
It is.
Lex Fridman (20:15.740)
Honestly, before that happened,
Daphne Koller (20:17.660)
I think very few people would have predicted
Lex Fridman (20:20.120)
that to be possible.
Daphne Koller (20:21.700)
It's amazing.
Lex Fridman (20:22.540)
Can you maybe elaborate, is it actually possible?
Lex Fridman (20:25.180)
Like where, like how stable?
Lex Fridman (20:27.300)
So this result was maybe like,
Daphne Koller (20:29.380)
I don't know how many years ago,
Lex Fridman (20:30.580)
maybe 10 years ago was first demonstrated,
Daphne Koller (20:32.700)
something like that.
Lex Fridman (20:33.860)
Is this, how hard is this?
Lex Fridman (20:35.520)
Like how noisy is this backward step?
Lex Fridman (20:37.500)
It seems quite incredible and cool.
Daphne Koller (20:39.460)
It is, it is incredible and cool.
Lex Fridman (20:42.220)
It was much more, I think finicky and bespoke
Daphne Koller (20:46.420)
at the early stages when the discovery was first made.
Lex Fridman (20:49.980)
But at this point, it's become almost industrialized.
Daphne Koller (20:54.500)
There are what's called contract research organizations,
Lex Fridman (20:59.440)
vendors that will take a sample from a human
Lex Fridman (21:02.300)
and revert it back to stem cell status.
Lex Fridman (21:04.460)
And it works a very good fraction of the time.
Daphne Koller (21:07.120)
Now there are people who will ask,
Lex Fridman (21:10.360)
I think good questions.
Daphne Koller (21:12.060)
Is this really truly a stem cell or does it remember
Lex Fridman (21:15.340)
certain aspects of what,
Lex Fridman (21:17.860)
of changes that were made in the human beyond the genetics?
Lex Fridman (21:22.500)
It's passed as a skin cell, yeah.
Daphne Koller (21:24.660)
It's passed as a skin cell or it's passed
Lex Fridman (21:26.740)
in terms of exposures to different environmental factors
Lex Fridman (21:29.920)
and so on.
Lex Fridman (21:30.920)
So I think the consensus right now
Daphne Koller (21:33.300)
is that these are not always perfect
Lex Fridman (21:36.420)
and there is little bits and pieces of memory sometimes,
Lex Fridman (21:40.020)
but by and large, these are actually pretty good.
Lex Fridman (21:44.780)
So one of the key things,
Daphne Koller (21:47.260)
well, maybe you can correct me,
Lex Fridman (21:48.740)
but one of the useful things for machine learning
Daphne Koller (21:50.900)
is size, scale of data.
Lex Fridman (21:54.180)
How easy it is to do these kinds of reversals to stem cells
Lex Fridman (21:59.100)
and then disease in a dish models at scale.
Lex Fridman (22:02.360)
Is that a huge challenge or not?
Lex Fridman (22:06.180)
So the reversal is not as of this point
Lex Fridman (22:11.660)
something that can be done at the scale
Daphne Koller (22:14.220)
of tens of thousands or hundreds of thousands.
Lex Fridman (22:18.540)
I think total number of stem cells or IPS cells
Daphne Koller (22:22.260)
that are what's called induced pluripotent stem cells
Lex Fridman (22:25.260)
in the world I think is somewhere between five and 10,000
Daphne Koller (22:30.220)
last I looked.
Lex Fridman (22:31.460)
Now again, that might not count things that exist
Daphne Koller (22:34.460)
in this or that academic center
Lex Fridman (22:36.260)
and they may add up to a bit more,
Lex Fridman (22:37.860)
but that's about the range.
Lex Fridman (22:40.060)
So it's not something that you could at this point
Daphne Koller (22:42.180)
generate IPS cells from a million people,
Lex Fridman (22:45.540)
but maybe you don't need to
Daphne Koller (22:47.900)
because maybe that background is enough
Lex Fridman (22:51.820)
because it can also be now perturbed in different ways.
Lex Fridman (22:56.140)
And some people have done really interesting experiments
Lex Fridman (23:00.100)
in for instance, taking cells from a healthy human
Lex Fridman (23:05.660)
and then introducing a mutation into it
Lex Fridman (23:08.540)
using one of the other miracle technologies
Daphne Koller (23:11.860)
that's emerged in the last decade
Lex Fridman (23:13.820)
which is CRISPR gene editing
Lex Fridman (23:16.140)
and introduced a mutation that is known to be pathogenic.
Lex Fridman (23:19.660)
And so you can now look at the healthy cells
Lex Fridman (23:22.420)
and the unhealthy cells, the one with the mutation
Lex Fridman (23:24.740)
and do a one on one comparison
Daphne Koller (23:26.100)
where everything else is held constant.
Lex Fridman (23:28.380)
And so you could really start to understand specifically
Lex Fridman (23:31.820)
what the mutation does at the cellular level.
Lex Fridman (23:34.380)
So the IPS cells are a great starting point
Lex Fridman (23:37.700)
and obviously more diversity is better
Lex Fridman (23:39.820)
because you also wanna capture ethnic background
Lex Fridman (23:42.380)
and how that affects things,
Lex Fridman (23:43.580)
but maybe you don't need one from every single patient
Daphne Koller (23:46.780)
with every single type of disease
Lex Fridman (23:48.100)
because we have other tools at our disposal.
Daphne Koller (23:50.260)
Well, how much difference is there between people
Lex Fridman (23:52.580)
I mentioned ethnic background in terms of IPS cells?
Lex Fridman (23:54.940)
So we're all like, it seems like these magical cells
Lex Fridman (23:59.380)
that can do to create anything
Daphne Koller (24:01.860)
between different populations, different people.
Lex Fridman (24:04.020)
Is there a lot of variability between cell cells?
Daphne Koller (24:07.020)
Well, first of all, there's the variability,
Lex Fridman (24:09.580)
that's driven simply by the fact
Daphne Koller (24:10.980)
that genetically we're different.
Lex Fridman (24:13.420)
So a stem cell that's derived from my genotype
Daphne Koller (24:15.820)
is gonna be different from a stem cell
Lex Fridman (24:18.340)
that's derived from your genotype.
Daphne Koller (24:20.540)
There's also some differences that have more to do with
Lex Fridman (24:23.700)
for whatever reason, some people's stem cells
Daphne Koller (24:27.260)
differentiate better than other people's stem cells.
Lex Fridman (24:29.860)
We don't entirely understand why.
Lex Fridman (24:31.500)
So there's certainly some differences there as well,
Lex Fridman (24:34.180)
but the fundamental difference
Lex Fridman (24:35.460)
and the one that we really care about and is a positive
Lex Fridman (24:38.740)
is that the fact that the genetics are different
Lex Fridman (24:43.220)
and therefore recapitulate my disease burden
Lex Fridman (24:45.980)
versus your disease burden.
Lex Fridman (24:47.780)
What's a disease burden?
Lex Fridman (24:49.260)
Well, a disease burden is just if you think,
Daphne Koller (24:52.300)
I mean, it's not a well defined mathematical term,
Lex Fridman (24:55.060)
although there are mathematical formulations of it.
Daphne Koller (24:58.260)
If you think about the fact that some of us are more likely
Lex Fridman (25:01.500)
to get a certain disease than others
Daphne Koller (25:03.460)
because we have more variations in our genome
Lex Fridman (25:07.300)
that are causative of the disease,
Daphne Koller (25:09.500)
maybe fewer that are protective of the disease.
Lex Fridman (25:12.620)
People have quantified that
Daphne Koller (25:14.860)
using what are called polygenic risk scores,
Lex Fridman (25:17.860)
which look at all of the variations
Daphne Koller (25:20.820)
in an individual person's genome
Lex Fridman (25:23.620)
and add them all up in terms of how much risk they confer
Daphne Koller (25:26.620)
for a particular disease.
Lex Fridman (25:27.820)
And then they've put people on a spectrum
Daphne Koller (25:30.540)
of their disease risk.
Lex Fridman (25:32.540)
And for certain diseases where we've been sufficiently
Daphne Koller (25:36.500)
powered to really understand the connection
Lex Fridman (25:38.740)
between the many, many small variations
Daphne Koller (25:41.580)
that give rise to an increased disease risk,
Lex Fridman (25:44.940)
there's some pretty significant differences
Daphne Koller (25:47.060)
in terms of the risk between the people,
Lex Fridman (25:49.300)
say at the highest decile of this polygenic risk score
Lex Fridman (25:52.060)
and the people at the lowest decile.
Lex Fridman (25:53.500)
Sometimes those differences are factor of 10 or 12 higher.
Lex Fridman (25:58.940)
So there's definitely a lot that our genetics
Lex Fridman (26:03.940)
contributes to disease risk, even if it's not
Daphne Koller (26:07.100)
by any stretch the full explanation.
Lex Fridman (26:09.100)
And from a machine learning perspective,
Daphne Koller (26:10.500)
there's signal there.
Lex Fridman (26:12.020)
There is definitely signal in the genetics
Lex Fridman (26:14.780)
and there's even more signal, we believe,
Lex Fridman (26:19.100)
in looking at the cells that are derived
Daphne Koller (26:21.540)
from those different genetics because in principle,
Lex Fridman (26:25.540)
you could say all the signal is there at the genetics level.
Lex Fridman (26:28.660)
So we don't need to look at the cells,
Lex Fridman (26:30.180)
but our understanding of the biology is so limited at this
Daphne Koller (26:34.100)
point than seeing what actually happens at the cellular
Lex Fridman (26:37.100)
level is a heck of a lot closer to the human clinical outcome
Daphne Koller (26:41.780)
than looking at the genetics directly.
Lex Fridman (26:44.620)
And so we can learn a lot more from it
Daphne Koller (26:47.180)
than we could by looking at genetics alone.
Lex Fridman (26:49.420)
So just to get a sense, I don't know if it's easy to do,
Lex Fridman (26:51.660)
but what kind of data is useful
Lex Fridman (26:54.220)
in this disease in a dish model?
Lex Fridman (26:56.220)
Like what's the source of raw data information?
Lex Fridman (26:59.940)
And also from my outsider's perspective,
Lex Fridman (27:03.900)
so biology and cells are squishy things.
Lex Fridman (27:08.620)
And then how do you connect the computer to that?
Daphne Koller (27:15.620)
Which sensory mechanisms, I guess.
Lex Fridman (27:17.780)
So that's another one of those revolutions
Daphne Koller (27:20.660)
that have happened in the last 10 years
Lex Fridman (27:22.540)
in that our ability to measure cells very quantitatively
Daphne Koller (27:27.540)
has also dramatically increased.
Lex Fridman (27:30.020)
So back when I started doing biology in the late 90s,
Daphne Koller (27:35.260)
early 2000s, that was the initial era
Lex Fridman (27:40.820)
where we started to measure biology
Daphne Koller (27:42.500)
in really quantitative ways using things like microarrays,
Lex Fridman (27:46.420)
where you would measure in a single experiment
Daphne Koller (27:50.580)
the activity level, what's called expression level
Lex Fridman (27:53.820)
of every gene in the genome in that sample.
Lex Fridman (27:56.980)
And that ability is what actually allowed us
Lex Fridman (28:00.340)
to even understand that there are molecular subtypes
Daphne Koller (28:04.180)
of diseases like cancer, where up until that point,
Lex Fridman (28:06.820)
it's like, oh, you have breast cancer.
Lex Fridman (28:09.220)
But then when we looked at the molecular data,
Lex Fridman (28:13.180)
it was clear that there's different subtypes
Daphne Koller (28:14.940)
of breast cancer that at the level of gene activity
Lex Fridman (28:17.460)
look completely different to each other.
Lex Fridman (28:20.660)
So that was the beginning of this process.
Lex Fridman (28:23.100)
Now we have the ability to measure individual cells
Daphne Koller (28:26.900)
in terms of their gene activity
Lex Fridman (28:28.860)
using what's called single cell RNA sequencing,
Daphne Koller (28:31.340)
which basically sequences the RNA,
Lex Fridman (28:35.020)
which is that activity level of different genes
Daphne Koller (28:39.300)
for every gene in the genome.
Lex Fridman (28:40.980)
And you could do that at single cell level.
Lex Fridman (28:42.700)
So that's an incredibly powerful way of measuring cells.
Lex Fridman (28:45.380)
I mean, you literally count the number of transcripts.
Lex Fridman (28:47.860)
So it really turns that squishy thing
Lex Fridman (28:50.020)
into something that's digital.
Daphne Koller (28:51.820)
Another tremendous data source that's emerged
Lex Fridman (28:55.100)
in the last few years is microscopy
Lex Fridman (28:57.860)
and specifically even super resolution microscopy,
Lex Fridman (29:00.580)
where you could use digital reconstruction
Daphne Koller (29:03.460)
to look at subcellular structures,
Lex Fridman (29:06.460)
sometimes even things that are below
Daphne Koller (29:08.380)
the diffraction limit of light
Lex Fridman (29:10.540)
by doing a sophisticated reconstruction.
Lex Fridman (29:13.340)
And again, that gives you a tremendous amount of information
Lex Fridman (29:16.500)
at the subcellular level.
Daphne Koller (29:18.420)
There's now more and more ways that amazing scientists
Lex Fridman (29:22.860)
out there are developing for getting new types
Daphne Koller (29:27.540)
of information from even single cells.
Lex Fridman (29:30.820)
And so that is a way of turning those squishy things
Daphne Koller (29:35.500)
into digital data.
Lex Fridman (29:37.260)
Into beautiful data sets.
Lex Fridman (29:38.660)
But so that data set then with machine learning tools
Lex Fridman (29:42.540)
allows you to maybe understand the developmental,
Daphne Koller (29:45.820)
like the mechanism of a particular disease.
Lex Fridman (29:49.900)
And if it's possible to sort of at a high level describe,
Lex Fridman (29:54.300)
how does that help lead to a drug discovery
Lex Fridman (30:01.180)
that can help prevent, reverse that mechanism?
Lex Fridman (30:05.380)
So I think there's different ways in which this data
Lex Fridman (30:08.180)
could potentially be used.
Daphne Koller (30:10.420)
Some people use it for scientific discovery
Lex Fridman (30:13.820)
and say, oh, look, we see this phenotype
Daphne Koller (30:17.060)
at the cellular level.
Lex Fridman (30:20.060)
So let's try and work our way backwards
Lex Fridman (30:22.940)
and think which genes might be involved in pathways
Lex Fridman (30:26.100)
that give rise to that.
Lex Fridman (30:27.060)
So that's a very sort of analytical method
Lex Fridman (30:32.380)
to sort of work our way backwards
Daphne Koller (30:35.140)
using our understanding of known biology.
Lex Fridman (30:38.500)
Some people use it in a somewhat more,
Daphne Koller (30:44.100)
sort of forward, if that was a backward,
Lex Fridman (30:46.580)
this would be forward, which is to say,
Daphne Koller (30:48.140)
okay, if I can perturb this gene,
Lex Fridman (30:51.140)
does it show a phenotype that is similar
Lex Fridman (30:54.060)
to what I see in disease patients?
Lex Fridman (30:56.020)
And so maybe that gene is actually causal of the disease.
Lex Fridman (30:58.980)
So that's a different way.
Lex Fridman (31:00.180)
And then there's what we do,
Daphne Koller (31:01.580)
which is basically to take that very large collection
Lex Fridman (31:06.260)
of data and use machine learning to uncover the patterns
Daphne Koller (31:10.660)
that emerge from it.
Lex Fridman (31:12.340)
So for instance, what are those subtypes
Daphne Koller (31:14.900)
that might be similar at the human clinical outcome,
Lex Fridman (31:18.620)
but quite distinct when you look at the molecular data?
Lex Fridman (31:21.740)
And then if we can identify such a subtype,
Lex Fridman (31:25.140)
are there interventions that if I apply it
Daphne Koller (31:27.980)
to cells that come from this subtype of the disease
Lex Fridman (31:32.060)
and you apply that intervention,
Daphne Koller (31:34.140)
it could be a drug or it could be a CRISPR gene intervention,
Lex Fridman (31:38.820)
does it revert the disease state
Daphne Koller (31:41.340)
to something that looks more like normal,
Lex Fridman (31:42.980)
happy, healthy cells?
Lex Fridman (31:44.100)
And so hopefully if you see that,
Lex Fridman (31:46.900)
that gives you a certain hope
Daphne Koller (31:50.380)
that that intervention will also have
Lex Fridman (31:53.100)
a meaningful clinical benefit to people.
Lex Fridman (31:55.100)
And there's obviously a bunch of things
Lex Fridman (31:56.580)
that you would wanna do after that to validate that,
Lex Fridman (31:58.740)
but it's a very different and much less hypothesis driven way
Lex Fridman (32:03.900)
of uncovering new potential interventions
Lex Fridman (32:06.100)
and might give rise to things that are not the same things
Lex Fridman (32:10.100)
that everyone else is already looking at.
Daphne Koller (32:12.460)
That's, I don't know, I'm just like to psychoanalyze
Lex Fridman (32:16.780)
my own feeling about our discussion currently.
Daphne Koller (32:18.700)
It's so exciting to talk about sort of a machine,
Lex Fridman (32:21.500)
fundamentally, well, something that's been turned
Daphne Koller (32:23.780)
into a machine learning problem
Lex Fridman (32:25.900)
and that says can have so much real world impact.
Daphne Koller (32:29.140)
That's how I feel too.
Lex Fridman (32:30.340)
That's kind of exciting because I'm so,
Daphne Koller (32:32.260)
most of my day is spent with data sets
Lex Fridman (32:35.740)
that I guess closer to the news groups.
Lex Fridman (32:39.060)
So this is a kind of, it just feels good to talk about.
Lex Fridman (32:41.980)
In fact, I almost don't wanna talk about machine learning.
Daphne Koller (32:45.340)
I wanna talk about the fundamentals of the data set,
Lex Fridman (32:47.460)
which is an exciting place to be.
Daphne Koller (32:50.420)
I agree with you.
Lex Fridman (32:51.740)
It's what gets me up in the morning.
Daphne Koller (32:53.740)
It's also what attracts a lot of the people
Lex Fridman (32:57.140)
who work at InCetro to InCetro
Daphne Koller (32:59.140)
because I think all of the,
Lex Fridman (33:01.660)
certainly all of our machine learning people
Daphne Koller (33:03.220)
are outstanding and could go get a job selling ads online
Lex Fridman (33:08.220)
or doing eCommerce or even self driving cars.
Lex Fridman (33:12.500)
But I think they would want, they come to us
Lex Fridman (33:17.860)
because they want to work on something
Daphne Koller (33:20.020)
that has more of an aspirational nature
Lex Fridman (33:22.380)
and can really benefit humanity.
Daphne Koller (33:24.740)
What, with these approaches, what do you hope,
Lex Fridman (33:28.300)
what kind of diseases can be helped?
Daphne Koller (33:31.140)
We mentioned Alzheimer's, schizophrenia, type 2 diabetes.
Lex Fridman (33:33.940)
Can you just describe the various kinds of diseases
Lex Fridman (33:36.540)
that this approach can help?
Lex Fridman (33:38.580)
Well, we don't know.
Lex Fridman (33:39.620)
And I try and be very cautious about making promises
Lex Fridman (33:43.900)
about some things that, oh, we will cure X.
Daphne Koller (33:46.620)
People make that promise.
Lex Fridman (33:48.060)
And I think it's, I tried to first deliver and then promise
Daphne Koller (33:52.700)
as opposed to the other way around.
Lex Fridman (33:54.460)
There are characteristics of a disease
Daphne Koller (33:57.340)
that make it more likely that this type of approach
Lex Fridman (34:00.580)
can potentially be helpful.
Lex Fridman (34:02.700)
So for instance, diseases that have
Lex Fridman (34:04.580)
a very strong genetic basis are ones
Daphne Koller (34:08.820)
that are more likely to manifest
Lex Fridman (34:10.940)
in a stem cell derived model.
Daphne Koller (34:13.860)
We would want the cellular models
Lex Fridman (34:16.300)
to be relatively reproducible and robust
Lex Fridman (34:19.940)
so that you could actually get enough of those cells
Lex Fridman (34:25.380)
and in a way that isn't very highly variable and noisy.
Daphne Koller (34:30.740)
You would want the disease to be relatively contained
Lex Fridman (34:34.140)
in one or a small number of cell types
Daphne Koller (34:36.700)
that you could actually create in an in vitro,
Lex Fridman (34:40.020)
in a dish setting.
Daphne Koller (34:40.980)
Whereas if it's something that's really broad and systemic
Lex Fridman (34:43.460)
and involves multiple cells
Daphne Koller (34:45.540)
that are in very distal parts of your body,
Lex Fridman (34:48.460)
putting that all in the dish is really challenging.
Lex Fridman (34:50.980)
So we want to focus on the ones
Lex Fridman (34:53.740)
that are most likely to be successful today
Daphne Koller (34:56.980)
with the hope, I think, that really smart bioengineers
Lex Fridman (35:01.980)
out there are developing better and better systems
Daphne Koller (35:04.900)
all the time so that diseases that might not be tractable
Lex Fridman (35:07.900)
today might be tractable in three years.
Lex Fridman (35:11.220)
So for instance, five years ago,
Lex Fridman (35:14.340)
these stem cell derived models didn't really exist.
Daphne Koller (35:16.140)
People were doing most of the work in cancer cells
Lex Fridman (35:18.540)
and cancer cells are very, very poor models
Daphne Koller (35:21.660)
of most human biology because they're,
Lex Fridman (35:24.300)
A, they were cancer to begin with
Lex Fridman (35:25.820)
and B, as you passage them and they proliferate in a dish,
Lex Fridman (35:30.140)
they become, because of the genomic instability,
Daphne Koller (35:32.660)
even less similar to human biology.
Lex Fridman (35:35.700)
Now we have these stem cell derived models.
Daphne Koller (35:39.340)
We have the capability to reasonably robustly,
Lex Fridman (35:42.620)
not quite at the right scale yet, but close,
Daphne Koller (35:45.820)
to derive what's called organoids,
Lex Fridman (35:47.940)
which are these teeny little sort of multicellular organ,
Daphne Koller (35:54.820)
sort of models of an organ system.
Lex Fridman (35:56.660)
So there's cerebral organoids and liver organoids
Lex Fridman (35:59.300)
and kidney organoids and.
Lex Fridman (36:01.620)
Yeah, brain organoids.
Daphne Koller (36:03.460)
That's organoids.
Lex Fridman (36:04.300)
It's possibly the coolest thing I've ever seen.
Lex Fridman (36:05.500)
Is that not like the coolest thing?
Lex Fridman (36:07.500)
Yeah.
Lex Fridman (36:08.380)
And then I think on the horizon,
Lex Fridman (36:09.940)
we're starting to see things like connecting
Daphne Koller (36:11.780)
these organoids to each other
Lex Fridman (36:13.900)
so that you could actually start,
Lex Fridman (36:15.140)
and there's some really cool papers that start to do that
Lex Fridman (36:17.620)
where you can actually start to say,
Lex Fridman (36:19.020)
okay, can we do multi organ system stuff?
Lex Fridman (36:22.180)
There's many challenges to that.
Daphne Koller (36:23.500)
It's not easy by any stretch, but it might,
Lex Fridman (36:27.780)
I'm sure people will figure it out.
Lex Fridman (36:29.460)
And in three years or five years,
Lex Fridman (36:31.580)
there will be disease models that we could make
Daphne Koller (36:34.020)
for things that we can't make today.
Lex Fridman (36:35.420)
Yeah, and this conversation would seem almost outdated
Daphne Koller (36:38.700)
with the kind of scale that could be achieved
Lex Fridman (36:40.460)
in like three years.
Daphne Koller (36:41.300)
I hope so.
Lex Fridman (36:42.140)
That's the hope.
Daphne Koller (36:42.980)
That would be so cool.
Lex Fridman (36:43.820)
So you've cofounded Coursera with Andrew Ng
Lex Fridman (36:48.060)
and were part of the whole MOOC revolution.
Lex Fridman (36:51.380)
So to jump topics a little bit,
Lex Fridman (36:53.900)
can you maybe tell the origin story of the history,
Lex Fridman (36:57.900)
the origin story of MOOCs, of Coursera,
Lex Fridman (37:00.900)
and in general, your teaching to huge audiences
Lex Fridman (37:07.100)
on a very sort of impactful topic of AI in general?
Lex Fridman (37:12.100)
So I think the origin story of MOOCs
Lex Fridman (37:15.860)
emanates from a number of efforts
Daphne Koller (37:17.940)
that occurred at Stanford University
Lex Fridman (37:20.580)
around the late 2000s
Daphne Koller (37:25.420)
where different individuals within Stanford,
Lex Fridman (37:28.580)
myself included, were getting really excited
Daphne Koller (37:31.500)
about the opportunities of using online technologies
Lex Fridman (37:35.220)
as a way of achieving both improved quality of teaching
Lex Fridman (37:38.980)
and also improved scale.
Lex Fridman (37:40.940)
And so Andrew, for instance,
Daphne Koller (37:44.420)
led the Stanford Engineering Everywhere,
Lex Fridman (37:48.820)
which was sort of an attempt to take 10 Stanford courses
Lex Fridman (37:51.660)
and put them online just as video lectures.
Lex Fridman (37:55.980)
I led an effort within Stanford to take some of the courses
Lex Fridman (38:00.620)
and really create a very different teaching model
Lex Fridman (38:04.380)
that broke those up into smaller units
Lex Fridman (38:07.340)
and had some of those embedded interactions and so on,
Lex Fridman (38:11.060)
which got a lot of support from university leaders
Daphne Koller (38:14.620)
because they felt like it was potentially a way
Lex Fridman (38:17.380)
of improving the quality of instruction at Stanford
Daphne Koller (38:19.580)
by moving to what's now called the flipped classroom model.
Lex Fridman (38:22.980)
And so those efforts eventually sort of started
Daphne Koller (38:26.620)
to interplay with each other
Lex Fridman (38:28.020)
and created a tremendous sense of excitement and energy
Daphne Koller (38:30.940)
within the Stanford community
Lex Fridman (38:32.780)
about the potential of online teaching
Lex Fridman (38:36.380)
and led in the fall of 2011
Lex Fridman (38:39.260)
to the launch of the first Stanford MOOCs.
Daphne Koller (38:43.740)
By the way, MOOCs, it's probably impossible
Lex Fridman (38:46.420)
that people don't know, but it's, I guess, massive.
Daphne Koller (38:49.020)
Open online courses. Open online courses.
Lex Fridman (38:51.900)
We did not come up with the acronym.
Daphne Koller (38:54.300)
I'm not particularly fond of the acronym,
Lex Fridman (38:57.020)
but it is what it is. It is what it is.
Daphne Koller (38:58.460)
Big bang is not a great term for the start of the universe,
Lex Fridman (39:01.380)
but it is what it is. Probably so.
Lex Fridman (39:05.220)
So anyway, so those courses launched in the fall of 2011,
Lex Fridman (39:10.900)
and there were, within a matter of weeks,
Daphne Koller (39:13.780)
with no real publicity campaign, just a New York Times article
Lex Fridman (39:17.940)
that went viral, about 100,000 students or more
Daphne Koller (39:22.660)
in each of those courses.
Lex Fridman (39:24.580)
And I remember this conversation that Andrew and I had.
Daphne Koller (39:29.180)
We were just like, wow, there's this real need here.
Lex Fridman (39:33.420)
And I think we both felt like, sure,
Daphne Koller (39:36.220)
we were accomplished academics and we could go back
Lex Fridman (39:39.820)
and go back to our labs, write more papers.
Lex Fridman (39:42.620)
But if we did that, then this wouldn't happen.
Lex Fridman (39:45.860)
And it seemed too important not to happen.
Lex Fridman (39:48.700)
And so we spent a fair bit of time debating,
Lex Fridman (39:51.620)
do we wanna do this as a Stanford effort,
Lex Fridman (39:55.300)
kind of building on what we'd started?
Lex Fridman (39:56.860)
Do we wanna do this as a for profit company?
Lex Fridman (39:59.340)
Do we wanna do this as a nonprofit?
Lex Fridman (40:00.780)
And we decided ultimately to do it as we did with Coursera.
Lex Fridman (40:04.900)
And so, you know, we started really operating
Lex Fridman (40:09.900)
as a company at the beginning of 2012.
Lex Fridman (40:13.380)
And the rest is history.
Lex Fridman (40:15.340)
But how did you, was that really surprising to you?
Lex Fridman (40:19.580)
How did you at that time and at this time
Lex Fridman (40:23.300)
make sense of this need for sort of global education
Daphne Koller (40:27.580)
you mentioned that you felt that, wow,
Lex Fridman (40:29.380)
the popularity indicates that there's a hunger
Daphne Koller (40:33.260)
for sort of globalization of learning.
Lex Fridman (40:37.620)
I think there is a hunger for learning that,
Daphne Koller (40:43.620)
you know, globalization is part of it,
Lex Fridman (40:45.100)
but I think it's just a hunger for learning.
Daphne Koller (40:47.140)
The world has changed in the last 50 years.
Lex Fridman (40:50.420)
It used to be that you finished college, you got a job,
Daphne Koller (40:54.820)
by and large, the skills that you learned in college
Lex Fridman (40:57.020)
were pretty much what got you through
Daphne Koller (40:59.700)
the rest of your job history.
Lex Fridman (41:01.380)
And yeah, you learn some stuff,
Lex Fridman (41:02.940)
but it wasn't a dramatic change.
Lex Fridman (41:05.500)
Today, we're in a world where the skills that you need
Daphne Koller (41:09.420)
for a lot of jobs, they didn't even exist
Lex Fridman (41:11.260)
when you went to college.
Lex Fridman (41:12.500)
And the jobs, and many of the jobs that existed
Lex Fridman (41:14.540)
when you went to college don't even exist today or are dying.
Lex Fridman (41:18.620)
So part of that is due to AI, but not only.
Lex Fridman (41:22.580)
And we need to find a way of keeping people,
Daphne Koller (41:27.300)
giving people access to the skills that they need today.
Lex Fridman (41:29.900)
And I think that's really what's driving
Daphne Koller (41:32.020)
a lot of this hunger.
Lex Fridman (41:33.900)
So I think if we even take a step back,
Daphne Koller (41:37.020)
for you, all of this started in trying to think
Lex Fridman (41:39.940)
of new ways to teach or to,
Daphne Koller (41:43.140)
new ways to sort of organize the material
Lex Fridman (41:47.100)
and present the material in a way
Daphne Koller (41:48.380)
that would help the education process, the pedagogy, yeah.
Lex Fridman (41:51.380)
So what have you learned about effective education
Daphne Koller (41:56.380)
from this process of playing,
Lex Fridman (41:57.540)
of experimenting with different ideas?
Lex Fridman (42:00.580)
So we learned a number of things.
Lex Fridman (42:03.940)
Some of which I think could translate back
Lex Fridman (42:06.620)
and have translated back effectively
Lex Fridman (42:08.380)
to how people teach on campus.
Lex Fridman (42:09.900)
And some of which I think are more specific
Lex Fridman (42:11.700)
to people who learn online,
Daphne Koller (42:13.820)
more sort of people who learn as part of their daily life.
Lex Fridman (42:18.900)
So we learned, for instance, very quickly
Daphne Koller (42:20.980)
that short is better.
Lex Fridman (42:23.180)
So people who are especially in the workforce
Daphne Koller (42:26.820)
can't do a 15 week semester long course.
Lex Fridman (42:30.020)
They just can't fit that into their lives.
Lex Fridman (42:32.500)
Sure, can you describe the shortness of what?
Lex Fridman (42:35.540)
The entirety, so every aspect,
Lex Fridman (42:39.060)
so the little lecture, the lecture's short,
Lex Fridman (42:41.980)
the course is short.
Daphne Koller (42:43.020)
Both.
Lex Fridman (42:43.860)
We started out, the first online education efforts
Daphne Koller (42:47.820)
were actually MIT's OpenCourseWare initiatives.
Lex Fridman (42:50.620)
And that was recording of classroom lectures and,
Daphne Koller (42:55.860)
Hour and a half or something like that, yeah.
Lex Fridman (42:57.620)
And that didn't really work very well.
Daphne Koller (43:00.380)
I mean, some people benefit.
Lex Fridman (43:01.540)
I mean, of course they did,
Lex Fridman (43:03.140)
but it's not really a very palatable experience
Lex Fridman (43:06.700)
for someone who has a job and three kids
Lex Fridman (43:11.220)
and they need to run errands and such.
Lex Fridman (43:13.980)
They can't fit 15 weeks into their life
Lex Fridman (43:17.900)
and the hour and a half is really hard.
Lex Fridman (43:20.700)
So we learned very quickly.
Daphne Koller (43:22.940)
I mean, we started out with short video modules
Lex Fridman (43:26.540)
and over time we made them shorter
Daphne Koller (43:28.180)
because we realized that 15 minutes was still too long.
Lex Fridman (43:31.660)
If you wanna fit in when you're waiting in line
Daphne Koller (43:33.860)
for your kid's doctor's appointment,
Lex Fridman (43:35.500)
it's better if it's five to seven.
Daphne Koller (43:38.620)
We learned that 15 week courses don't work
Lex Fridman (43:42.540)
and you really wanna break this up into shorter units
Lex Fridman (43:44.820)
so that there is a natural completion point,
Lex Fridman (43:46.820)
gives people a sense of they're really close
Daphne Koller (43:48.660)
to finishing something meaningful.
Lex Fridman (43:50.420)
They can always come back and take part two and part three.
Daphne Koller (43:53.580)
We also learned that compressing the content works
Lex Fridman (43:56.500)
really well because if some people that pace works well
Lex Fridman (44:00.340)
and for others, they can always rewind and watch again.
Lex Fridman (44:03.260)
And so people have the ability
Daphne Koller (44:05.340)
to then learn at their own pace.
Lex Fridman (44:06.980)
And so that flexibility, the brevity and the flexibility
Daphne Koller (44:11.740)
are both things that we found to be very important.
Lex Fridman (44:15.420)
We learned that engagement during the content is important
Lex Fridman (44:18.780)
and the quicker you give people feedback,
Lex Fridman (44:20.620)
the more likely they are to be engaged.
Daphne Koller (44:22.540)
Hence the introduction of these,
Lex Fridman (44:24.540)
which we actually was an intuition that I had going in
Lex Fridman (44:27.740)
and was then validated using data
Lex Fridman (44:30.900)
that introducing some of these sort of little micro quizzes
Daphne Koller (44:34.300)
into the lectures really helps.
Lex Fridman (44:36.500)
Self graded as automatically graded assessments
Daphne Koller (44:39.420)
really helped too because it gives people feedback.
Lex Fridman (44:41.900)
See, there you are.
Lex Fridman (44:43.180)
So all of these are valuable.
Lex Fridman (44:45.620)
And then we learned a bunch of other things too.
Daphne Koller (44:47.260)
We did some really interesting experiments, for instance,
Lex Fridman (44:49.420)
on gender bias and how having a female role model
Daphne Koller (44:54.180)
as an instructor can change the balance of men to women
Lex Fridman (44:59.340)
in terms of, especially in STEM courses.
Lex Fridman (45:02.020)
And you could do that online by doing AB testing
Lex Fridman (45:04.820)
in ways that would be really difficult to go on campus.
Daphne Koller (45:07.740)
Oh, that's exciting.
Lex Fridman (45:09.140)
But so the shortness, the compression,
Daphne Koller (45:11.540)
I mean, that's actually, so that probably is true
Lex Fridman (45:15.700)
for all good editing is always just compressing the content,
Daphne Koller (45:20.980)
making it shorter.
Lex Fridman (45:21.940)
So that puts a lot of burden on the creator of the,
Daphne Koller (45:24.860)
the instructor and the creator of the educational content.
Lex Fridman (45:28.660)
Probably most lectures at MIT or Stanford
Daphne Koller (45:31.260)
could be five times shorter
Lex Fridman (45:34.340)
if the preparation was put enough.
Lex Fridman (45:37.580)
So maybe people might disagree with that,
Lex Fridman (45:41.660)
but like the Christmas, the clarity that a lot of the,
Lex Fridman (45:45.340)
like Coursera delivers is, how much effort does that take?
Lex Fridman (45:50.140)
So first of all, let me say that it's not clear
Daphne Koller (45:54.100)
that that crispness would work as effectively
Lex Fridman (45:57.380)
in a face to face setting
Daphne Koller (45:58.900)
because people need time to absorb the material.
Lex Fridman (46:02.420)
And so you need to at least pause
Lex Fridman (46:04.740)
and give people a chance to reflect and maybe practice.
Lex Fridman (46:07.300)
And that's what MOOCs do is that they give you
Daphne Koller (46:09.500)
these chunks of content and then ask you
Lex Fridman (46:11.780)
to practice with it.
Lex Fridman (46:13.420)
And that's where I think some of the newer pedagogy
Lex Fridman (46:16.300)
that people are adopting in face to face teaching
Daphne Koller (46:19.180)
that have to do with interactive learning and such
Lex Fridman (46:21.580)
can be really helpful.
Lex Fridman (46:23.460)
But both those approaches,
Lex Fridman (46:26.620)
whether you're doing that type of methodology
Daphne Koller (46:29.380)
in online teaching or in that flipped classroom,
Lex Fridman (46:32.820)
interactive teaching.
Lex Fridman (46:34.500)
What's that, sorry to pause, what's flipped classroom?
Lex Fridman (46:37.180)
Flipped classroom is a way in which online content
Daphne Koller (46:41.540)
is used to supplement face to face teaching
Lex Fridman (46:45.060)
where people watch the videos perhaps
Lex Fridman (46:47.220)
and do some of the exercises before coming to class.
Lex Fridman (46:49.860)
And then when they come to class,
Daphne Koller (46:51.180)
it's actually to do much deeper problem solving
Lex Fridman (46:53.580)
oftentimes in a group.
Lex Fridman (46:56.100)
But any one of those different pedagogies
Lex Fridman (47:00.460)
that are beyond just standing there and droning on
Daphne Koller (47:03.500)
in front of the classroom for an hour and 15 minutes
Lex Fridman (47:06.300)
require a heck of a lot more preparation.
Lex Fridman (47:09.260)
And so it's one of the challenges I think that people have
Lex Fridman (47:13.660)
that we had when trying to convince instructors
Daphne Koller (47:15.740)
to teach on Coursera.
Lex Fridman (47:16.700)
And it's part of the challenges that pedagogy experts
Daphne Koller (47:20.380)
on campus have in trying to get faculty
Lex Fridman (47:22.060)
to teach differently is that it's actually harder
Daphne Koller (47:23.740)
to teach that way than it is to stand there and drone.
Lex Fridman (47:27.860)
Do you think MOOCs will replace in person education
Daphne Koller (47:32.420)
or become the majority of in person of education
Lex Fridman (47:37.420)
of the way people learn in the future?
Daphne Koller (47:41.380)
Again, the future could be very far away,
Lex Fridman (47:43.260)
but where's the trend going do you think?
Lex Fridman (47:46.020)
So I think it's a nuanced and complicated answer.
Lex Fridman (47:50.140)
I don't think MOOCs will replace face to face teaching.
Daphne Koller (47:55.780)
I think learning is in many cases a social experience.
Lex Fridman (48:00.300)
And even at Coursera, we had people who naturally formed
Daphne Koller (48:05.300)
study groups, even when they didn't have to,
Lex Fridman (48:07.780)
to just come and talk to each other.
Lex Fridman (48:10.300)
And we found that that actually benefited their learning
Lex Fridman (48:14.420)
in very important ways.
Lex Fridman (48:15.780)
So there was more success among learners
Lex Fridman (48:19.660)
who had those study groups than among ones who didn't.
Lex Fridman (48:22.620)
So I don't think it's just gonna,
Lex Fridman (48:23.860)
oh, we're all gonna just suddenly learn online
Daphne Koller (48:26.060)
with a computer and no one else in the same way
Lex Fridman (48:28.940)
that recorded music has not replaced live concerts.
Lex Fridman (48:33.180)
But I do think that especially when you are thinking
Lex Fridman (48:38.940)
about continuing education, the stuff that people get
Daphne Koller (48:42.740)
when they're traditional,
Lex Fridman (48:44.700)
whatever high school, college education is done,
Lex Fridman (48:47.780)
and they yet have to maintain their level of expertise
Lex Fridman (48:52.500)
and skills in a rapidly changing world,
Daphne Koller (48:54.620)
I think people will consume more and more educational content
Lex Fridman (48:58.180)
in this online format because going back to school
Daphne Koller (49:01.380)
for formal education is not an option for most people.
Lex Fridman (49:04.860)
Briefly, it might be a difficult question to ask,
Lex Fridman (49:07.380)
but there's a lot of people fascinated
Lex Fridman (49:09.940)
by artificial intelligence, by machine learning,
Daphne Koller (49:12.820)
by deep learning.
Lex Fridman (49:13.940)
Is there a recommendation for the next year
Lex Fridman (49:18.140)
or for a lifelong journey of somebody interested in this?
Lex Fridman (49:21.340)
How do they begin?
Lex Fridman (49:23.700)
How do they enter that learning journey?
Lex Fridman (49:27.220)
I think the important thing is first to just get started.
Lex Fridman (49:30.900)
And there's plenty of online content that one can get
Lex Fridman (49:36.580)
for both the core foundations of mathematics
Lex Fridman (49:40.460)
and statistics and programming.
Lex Fridman (49:42.260)
And then from there to machine learning,
Daphne Koller (49:44.580)
I would encourage people not to skip
Lex Fridman (49:47.100)
to quickly pass the foundations
Daphne Koller (49:48.700)
because I find that there's a lot of people
Lex Fridman (49:51.060)
who learn machine learning, whether it's online
Daphne Koller (49:53.740)
or on campus without getting those foundations.
Lex Fridman (49:56.180)
And they basically just turn the crank on existing models
Daphne Koller (50:00.020)
in ways that A, don't allow for a lot of innovation
Lex Fridman (50:03.540)
and an adjustment to the problem at hand,
Lex Fridman (50:07.700)
but also B, are sometimes just wrong
Lex Fridman (50:09.660)
and they don't even realize that their application is wrong
Daphne Koller (50:12.900)
because there's artifacts that they haven't fully understood.
Lex Fridman (50:15.940)
So I think the foundations,
Daphne Koller (50:17.860)
machine learning is an important step.
Lex Fridman (50:19.860)
And then actually start solving problems,
Daphne Koller (50:24.860)
try and find someone to solve them with
Lex Fridman (50:27.620)
because especially at the beginning,
Daphne Koller (50:28.980)
it's useful to have someone to bounce ideas off
Lex Fridman (50:31.580)
and fix mistakes that you make
Lex Fridman (50:33.220)
and you can fix mistakes that they make,
Lex Fridman (50:35.980)
but then just find practical problems,
Daphne Koller (50:40.540)
whether it's in your workplace or if you don't have that,
Lex Fridman (50:43.300)
Kaggle competitions or such are a really great place
Daphne Koller (50:46.100)
to find interesting problems and just practice.
Lex Fridman (50:50.860)
Practice.
Daphne Koller (50:52.340)
Perhaps a bit of a romanticized question,
Lex Fridman (50:54.540)
but what idea in deep learning do you find,
Daphne Koller (50:59.340)
have you found in your journey the most beautiful
Lex Fridman (51:02.220)
or surprising or interesting?
Daphne Koller (51:07.660)
Perhaps not just deep learning,
Lex Fridman (51:09.420)
but AI in general, statistics.
Daphne Koller (51:14.940)
I'm gonna answer with two things.
Lex Fridman (51:19.100)
One would be the foundational concept of end to end training,
Daphne Koller (51:23.100)
which is that you start from the raw data
Lex Fridman (51:26.940)
and you train something that is not like a single piece,
Lex Fridman (51:32.980)
but rather towards the actual goal that you're looking to.
Lex Fridman (51:38.980)
From the raw data to the outcome,
Daphne Koller (51:40.820)
like no details in between.
Lex Fridman (51:43.580)
Well, not no details, but the fact that you,
Daphne Koller (51:45.460)
I mean, you could certainly introduce building blocks
Lex Fridman (51:47.540)
that were trained towards other tasks.
Daphne Koller (51:50.260)
I'm actually coming to that in my second half of the answer,
Lex Fridman (51:53.060)
but it doesn't have to be like a single monolithic blob
Daphne Koller (51:57.740)
in the middle.
Lex Fridman (51:58.580)
Actually, I think that's not ideal,
Lex Fridman (52:00.220)
but rather the fact that at the end of the day,
Lex Fridman (52:02.620)
you can actually train something that goes all the way
Daphne Koller (52:04.780)
from the beginning to the end.
Lex Fridman (52:06.900)
And the other one that I find really compelling
Daphne Koller (52:09.140)
is the notion of learning a representation
Lex Fridman (52:13.180)
that in its turn, even if it was trained to another task,
Daphne Koller (52:18.180)
can potentially be used as a much more rapid starting point
Lex Fridman (52:24.260)
to solving a different task.
Lex Fridman (52:26.700)
And that's, I think, reminiscent
Lex Fridman (52:29.500)
of what makes people successful learners.
Daphne Koller (52:32.300)
It's something that is relatively new
Lex Fridman (52:35.460)
in the machine learning space.
Daphne Koller (52:36.540)
I think it's underutilized even relative
Lex Fridman (52:38.700)
to today's capabilities, but more and more
Lex Fridman (52:41.460)
of how do we learn sort of reusable representation?
Lex Fridman (52:45.220)
And so end to end and transfer learning.
Daphne Koller (52:49.700)
Yeah.
Lex Fridman (52:51.140)
Is it surprising to you that neural networks
Lex Fridman (52:53.660)
are able to, in many cases, do these things?
Lex Fridman (52:56.980)
Is it maybe taken back to when you first would dive deep
Daphne Koller (53:02.260)
into neural networks or in general, even today,
Lex Fridman (53:05.460)
is it surprising that neural networks work at all
Lex Fridman (53:07.860)
and work wonderfully to do this kind of raw end to end
Lex Fridman (53:12.860)
and end to end learning and even transfer learning?
Daphne Koller (53:16.380)
I think I was surprised by how well
Lex Fridman (53:22.540)
when you have large enough amounts of data,
Daphne Koller (53:26.820)
it's possible to find a meaningful representation
Lex Fridman (53:32.940)
in what is an exceedingly high dimensional space.
Lex Fridman (53:36.060)
And so I find that to be really exciting
Lex Fridman (53:39.300)
and people are still working on the math for that.
Daphne Koller (53:41.620)
There's more papers on that every year.
Lex Fridman (53:43.580)
And I think it would be really cool
Daphne Koller (53:46.220)
if we figured that out, but that to me was a surprise
Lex Fridman (53:52.220)
because in the early days when I was starting my way
Daphne Koller (53:55.420)
in machine learning and the data sets were rather small,
Lex Fridman (53:58.700)
I think we believed, I believed that you needed
Daphne Koller (54:02.780)
to have a much more constrained
Lex Fridman (54:05.500)
and knowledge rich search space
Daphne Koller (54:08.620)
to really make, to really get to a meaningful answer.
Lex Fridman (54:11.860)
And I think it was true at the time.
Lex Fridman (54:13.860)
What I think is still a question
Lex Fridman (54:18.220)
is will a completely knowledge free approach
Daphne Koller (54:23.180)
where there's no prior knowledge going
Lex Fridman (54:26.020)
into the construction of the model,
Lex Fridman (54:28.980)
is that gonna be the solution or not?
Lex Fridman (54:31.620)
It's not actually the solution today
Daphne Koller (54:34.180)
in the sense that the architecture of a convolutional
Lex Fridman (54:38.940)
neural network that's used for images
Daphne Koller (54:41.500)
is actually quite different
Lex Fridman (54:43.260)
to the type of network that's used for language
Lex Fridman (54:46.580)
and yet different from the one that's used for speech
Lex Fridman (54:50.220)
or biology or any other application.
Daphne Koller (54:52.500)
There's still some insight that goes
Lex Fridman (54:55.860)
into the structure of the network
Daphne Koller (54:58.180)
to get the right performance.
Lex Fridman (55:00.820)
Will you be able to come up
Lex Fridman (55:01.660)
with a universal learning machine?
Lex Fridman (55:03.220)
I don't know.
Daphne Koller (55:05.100)
I wonder if there's always has to be some insight
Lex Fridman (55:07.300)
injected somewhere or whether it can converge.
Lex Fridman (55:10.300)
So you've done a lot of interesting work
Lex Fridman (55:13.580)
with probabilistic graphical models in general,
Daphne Koller (55:16.340)
Bayesian deep learning and so on.
Lex Fridman (55:18.420)
Can you maybe speak high level,
Lex Fridman (55:21.060)
how can learning systems deal with uncertainty?
Lex Fridman (55:25.500)
One of the limitations I think of a lot
Daphne Koller (55:28.940)
of machine learning models is that
Lex Fridman (55:33.780)
they come up with an answer
Lex Fridman (55:35.780)
and you don't know how much you can believe that answer.
Lex Fridman (55:40.860)
And oftentimes the answer is actually
Daphne Koller (55:47.740)
quite poorly calibrated relative to its uncertainties.
Lex Fridman (55:50.580)
Even if you look at where the confidence
Daphne Koller (55:55.500)
that comes out of say the neural network at the end,
Lex Fridman (55:58.980)
and you ask how much more likely
Daphne Koller (56:01.820)
is an answer of 0.8 versus 0.9,
Lex Fridman (56:04.820)
it's not really in any way calibrated
Daphne Koller (56:07.700)
to the actual reliability of that network
Lex Fridman (56:12.340)
and how true it is.
Lex Fridman (56:13.180)
And the further away you move from the training data,
Lex Fridman (56:16.780)
the more, not only the more wrong the network is,
Daphne Koller (56:20.700)
often it's more wrong and more confident
Lex Fridman (56:22.580)
in its wrong answer.
Lex Fridman (56:24.380)
And that is a serious issue in a lot of application areas.
Lex Fridman (56:29.340)
So when you think for instance,
Daphne Koller (56:30.380)
about medical diagnosis as being maybe an epitome
Lex Fridman (56:33.340)
of how problematic this can be,
Daphne Koller (56:35.700)
if you were training your network
Lex Fridman (56:37.700)
on a certain set of patients
Lex Fridman (56:40.180)
and a certain patient population,
Lex Fridman (56:41.540)
and I have a patient that is an outlier
Lex Fridman (56:44.620)
and there's no human that looks at this,
Lex Fridman (56:46.780)
and that patient is put into a neural network
Lex Fridman (56:49.100)
and your network not only gives
Lex Fridman (56:50.340)
a completely incorrect diagnosis,
Lex Fridman (56:51.940)
but is supremely confident
Lex Fridman (56:53.980)
in its wrong answer, you could kill people.
Lex Fridman (56:56.340)
So I think creating more of an understanding
Lex Fridman (57:01.940)
of how do you produce networks
Daphne Koller (57:05.540)
that are calibrated in their uncertainty
Lex Fridman (57:09.060)
and can also say, you know what, I give up.
Daphne Koller (57:10.940)
I don't know what to say about this particular data instance
Lex Fridman (57:14.580)
because I've never seen something
Daphne Koller (57:16.340)
that's sufficiently like it before.
Lex Fridman (57:18.140)
I think it's going to be really important
Daphne Koller (57:20.540)
in mission critical applications,
Lex Fridman (57:23.060)
especially ones where human life is at stake
Lex Fridman (57:25.380)
and that includes medical applications,
Lex Fridman (57:28.300)
but it also includes automated driving
Daphne Koller (57:31.180)
because you'd want the network to be able to say,
Lex Fridman (57:33.300)
you know what, I have no idea what this blob is
Daphne Koller (57:36.020)
that I'm seeing in the middle of the road.
Lex Fridman (57:37.140)
So I'm just going to stop
Daphne Koller (57:38.380)
because I don't want to potentially run over a pedestrian
Lex Fridman (57:41.540)
that I don't recognize.
Daphne Koller (57:42.820)
Is there good mechanisms, ideas of how to allow
Lex Fridman (57:47.540)
learning systems to provide that uncertainty
Lex Fridman (57:52.260)
along with their predictions?
Lex Fridman (57:54.060)
Certainly people have come up with mechanisms
Daphne Koller (57:57.180)
that involve Bayesian deep learning,
Lex Fridman (58:00.700)
deep learning that involves Gaussian processes.
Daphne Koller (58:04.460)
I mean, there's a slew of different approaches
Lex Fridman (58:07.660)
that people have come up with.
Daphne Koller (58:09.180)
There's methods that use ensembles of networks
Lex Fridman (58:13.660)
trained with different subsets of data
Daphne Koller (58:15.260)
or different random starting points.
Lex Fridman (58:17.620)
Those are actually sometimes surprisingly good
Daphne Koller (58:20.260)
at creating a sort of set of how confident
Lex Fridman (58:24.020)
or not you are in your answer.
Daphne Koller (58:26.580)
It's very much an area of open research.
Lex Fridman (58:30.020)
Let's cautiously venture back into the land of philosophy
Lex Fridman (58:33.660)
and speaking of AI systems providing uncertainty,
Lex Fridman (58:37.660)
somebody like Stuart Russell believes
Daphne Koller (58:41.140)
that as we create more and more intelligence systems,
Lex Fridman (58:43.420)
it's really important for them to be full of self doubt
Daphne Koller (58:46.820)
because if they're given more and more power,
Lex Fridman (58:51.940)
we want the way to maintain human control
Daphne Koller (58:54.820)
over AI systems or human supervision, which is true.
Lex Fridman (58:57.900)
Like you just mentioned with autonomous vehicles,
Daphne Koller (58:59.500)
it's really important to get human supervision
Lex Fridman (59:02.420)
when the car is not sure because if it's really confident
Daphne Koller (59:05.940)
in cases when it can get in trouble,
Lex Fridman (59:07.860)
it's gonna be really problematic.
Lex Fridman (59:09.380)
So let me ask about sort of the questions of AGI
Lex Fridman (59:12.980)
and human level intelligence.
Daphne Koller (59:14.860)
I mean, we've talked about curing diseases,
Lex Fridman (59:18.780)
which is sort of fundamental thing
Daphne Koller (59:20.180)
we can have an impact today,
Lex Fridman (59:21.780)
but AI people also dream of both understanding
Lex Fridman (59:26.180)
and creating intelligence.
Lex Fridman (59:29.220)
Is that something you think about?
Lex Fridman (59:30.420)
Is that something you dream about?
Lex Fridman (59:32.780)
Is that something you think is within our reach
Lex Fridman (59:36.980)
to be thinking about as computer scientists?
Lex Fridman (59:39.660)
Well, boy, let me tease apart different parts
Daphne Koller (59:43.500)
of that question.
Lex Fridman (59:45.180)
The worst question.
Daphne Koller (59:46.420)
Yeah, it's a multi part question.
Lex Fridman (59:50.940)
So let me start with the feasibility of AGI.
Daphne Koller (59:57.500)
Then I'll talk about the timelines a little bit
🔗 相关节目