January 14, 2020 • 26 mins
In this special episode, Oz and Karah examine our evolving relationship with the technology we create. Karah meets Jason Cohen, CEO of Analytical Flavor Systems, to see if his team can hack her taste preferences, and use AI to create a new flavor of beverage that she will love. Oz and Karah also look ahead to Season 2, previewing stories they are excited to report, including algorithms that promise to optimize end-of-life conversations. And they share highlights from conversations with guests from Season 1 who shaped their thinking about AI: Yuval Noah Harari, Siddhartha Mukherjee, and Regina Barzilay of MIT.
Learn more about your ad-choices at https://www.iheartpodcastnetwork.com
See omnystudio.com/listener for privacy information.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:04):
Sleepwalkers is a production of our Heart Radio and Unusual Productions. Hi,
I'm Aloan and I'm care Price. Welcome to a special
bonus episode of Sleepwalkers. Well, Cara, it's very good to
(00:36):
be back to the office. That you It is good
to be back, as I just want to apologize to
our listeners. We don't have an algorithm that's going to
create season two, so things have taken a little longer. Yes, unfortunately,
unfortunately a I can't do everything yet. But we are
hard at work on season two of Sleepwalkers, and we're
focusing on stories that really contextualize the implications of AI,
(00:57):
what it's doing, what the future is, and how it's
affecting us. You know, we had such a good time
in season one of Sleepwalkers wrapping our heads around the
meaning of artificial intelligence. It becomes basically a principle of
statistics prediction, how we're using data to inform our decisions,
and how that's becoming ingrained in products and services and
(01:18):
everything we do. Really, it's true the Pandora's box of
AI has been opened, but we still have the black
box problem that is true explainable AI. We can't tell
what neural networks are doing yet, but people are working
on it, and that's a story we're going to be
covering closely in season two. But in this bonus episode
of Sleepwalkers, we're going to take a look back at
some of the most poignant stories and interesting applications of
(01:40):
AI that we talked about in the first season, and
later in this episode, I'm going to give you a
preview of a fascinating new company that's using AI to
predict very specific consumer taste, as in preferences, not like
tasting clothing. Before we get through though us, when you
look back at season one, what stands out to you
the most? One of the things that stands out to
(02:01):
me most is the story you did about liar Bird.
Thank you, but seriously, the way they use an algorithm
to create a deep fake of your voice. But in
that particular piece, the questions it raised about mortality and
would you want to hear your father's voice beyond the
grave has stuck with me and I was one of
the most powerful moments in the whole podcast. Yeah so,
(02:22):
Jose Satello, one of the founders of Liar Bird, had
me sit down on a microphone for an hour and
just speak which was a personal dream of mine, and
then using that out me interrupting. Using that voice data,
liar Bird's algorithms created a version of my voice that
could turn any written text into something that sounded like me.
(02:42):
But Jose actually explained it better. So can you hit
the clip? Not an AI scientist, but we do have
the sophistication to roll tape. I know it might sound
a bit like magic, but in reality, the way that
our algorithm's work is basically they are just a battern
matching algorithms, and so it's trying to figure out how
(03:05):
to identify the patterns in your voice by comparing it
against thousands of other voices a shoually tens of thousands
of other voices, and trying to figure out what is
it that makes your voice unique. Once Jose's algorithms identified
what was unique about my voice, obviously everything they had
the building blocks they needed to make a fake. Then
(03:25):
we sent Jose a set of sentences we wanted robot
care to say, and he used another set of algorithms
to turn the text into what we heard. The way
they do this is they use it's called a generative
adversarial network again, which is a system where one neural
net tries to trick another one a thousand times per second,
So each time the second network to texts of fake,
(03:47):
the first one tries again. It basically learns from its mistakes,
and once it tricks its adversary, it's ready to show
its results. In our case, liar bird pits my fake
voice against my real voice until it sounds like this
sub dog Scara Karen. One of the reasons why this
story has stuck with me is because it feels like
we're just at the beginning of tapping the potential and
(04:10):
the potential for harm of deep fakes. May be remembered
as the year of the first significant deep fake crime.
An employee at a UK based energy firm believed he
was on the phone to his boss and followed instructions
to transfer two hundred thousand pounds to a scammers bank account.
That certainly won't be the last deep fatefore we hear of,
(04:30):
and it raises questions about responsibility and accountability. Who's liable
in a case like this. Facebook has actually gone so
far as to create a deep fake detection challenge to
get the best minds thinking about deep fakes and how
we might solve the problem and offering like a million
dollar prize reach prize, which is like a dollar but
it also shows how important the issue is, especially when
(04:52):
a company like Facebook gets behind it. Um. There's another
side to deep fake technology that actually highlights this dichotomy
and tech anology right now, which is that it can
be used for menacing purposes but also really powerful and
beautiful applications. Jose goes on to talk about how liar
Bird can be used with als patients and give them
the ability to speak when they've lost all ability to speak,
(05:14):
when it could give them the opportunity to speak in
a version of their voice to their children again, which
is quite profound. One area where I think technology is
a powerhouse for change is in medicine. Technologists and doctors
alike are looking at AI to predict, treat, and diagnose,
you know, everything from depression to cancer, and that's a
very wide spectrum. And it reminds me of one of
(05:36):
my favorite interviews that you did, which was your interview
with Saddartha. Muker g just so fascinated by this article
he'd written for The New Yoker called AI versus m
D which laid out all of the kind of benefits
and potential applications of using AI in medicine, including some
of the downside such as the black box problem of
AI that you mentioned, not knowing why an algorithm has
(05:57):
made a recommendation, and also another problem, which is that
if we really too much on technology, it can erode
human skills. There is a fear that AI could move
us into a very black and white way of thinking.
The computer says you have cancer, or the computer says
you to have your liver removed, said Arthur, who is
one of the world's foremost oncologists and a Pullet Surprise
winning author. Provided a different perspective. There is something very
(06:21):
fundamental about the human brain, a scientist's brain, a doctor's brain,
and artist's brain that asks questions in a fundamentally different manner,
the why question. Why did this happen in this person
in this time? Why does the melanoma appear in the
first calase? What is the molecular basis of that appearance.
The most interesting mysteries of medicine remain mysteries that have
(06:41):
to do with the why. Once we give up some
of the diagnostic pattern recognition material to machines, it will
be time to play. It will be the time to
play in the arena of human therapeutics, human biology, the
complexity of the human interaction, the art of medicine. My
hope is that medicine, in being more playful, will become
(07:01):
more compassionate, more able to take into account individuals and
their individual destinies rather than bucketing people in big categories.
It means having more time to spend with humans. You know,
we are so constrained by time that even compassion gets
three minutes, We won't become more robotic, will become less
(07:23):
robotic as the robots enter our own So Dartha's point
is that these tools could make doctors more efficient so
that they can provide better care. It sort of takes
the grunt work out of medicine and puts the patient
care work back in the doctor's hands. This idea that
technology can actually allow us to be more human, make
(07:46):
us more empathetic, is fascinating, and it also raises the
questions about new types of skills that may need to
be developed in an age of AI. Yeah, and Regina
Barslay from m I T spoke a lot about this.
How doctors have to now equip themselves with new ways
of translating data to patients, we still do not communicate
it to the patient because I think now there is
(08:10):
a walk to be done, not on computer science or AIPAD,
but really on the clinical side. What is the best
way to communicate it to the patient and what is
um you know, the past that you're going to give them.
It is not just enough to say you know you
are high risk. You need to propose some suggestion and solution.
(08:31):
So currently the clinical stuff is thinking and looking at
the ways of effective you know, clinical engagement with a patient.
You know, I speaking of data. You all know Harri
was another person who made you and me think about
humans as reducible to data. I think he's mostly known
as a historian and for his book Sapiens, but you
(08:52):
spoke with him about the data we produce as humans
and how that influences our relationship with technology. That's right,
which is the topic of his book Daus. And he
has this phrase data is m to describe how we've
kind of come to worship the data we create and
our own technological creations. So what happens when based on
all of our past behavior, AI starts to know us
(09:15):
better than we know ourselves. Here's a clip from You
are talking about exactly that. When we talk about AI,
we tend to greatly exaggerate the potential abilities, but at
the same time we also tend to exaggerate the abilities
of humans. People say that AI is not going to
take over our lives because it's very imperfect and it
(09:39):
won't be able to know us perfectly. But what people
forget is that humans often have a very poor understanding
of themselves, of the desires, of their emotions, of their
mental states. For AI to take over your life, it
doesn't need to know you perfectly, just need to know
(10:01):
you better than you know yourself. And that's not very
difficult because we often don't know the most important things
about ourselves. So, but let's say you could turn back
the clock to being fifteen, would you have wanted to
live in a world where there was sufficiently good sensors
to monitor your eyes, your eye movement, you're breathing, you know,
(10:21):
while you're going about your daily life, and then to
interpret that and say to you you have all more
likely than not you're gay. That's a very good question
which will become very practical questions in a few years.
And the way that I grew up and developed. It
would have been a very bad idea. I wouldn't like
(10:42):
to receive this kind of insight from form a machine.
I'm not sure how I would have dealt with it
when I was fifteen, you know, in Israel in the
nineteen eighties, and maybe partly it was you know, a
defense mechanism. In the future too. It it depends where
you live. Brunei has instituted the death penalty for gay people,
at least for people engaged in homosexual sex. So if
(11:05):
I'm a teenager in Brunei, I don't want to be
told by the computer that I'm gay, because the computer
will then be able to tell that to the police
and to the authorities as well. So the apps we use,
the product we buy, the number of steps we take,
the delivery I ordered last night, that will becomes data,
and that data can feed into neural networks to create
(11:27):
statistical models of us and what we might do next,
sometimes in order to diagnose a medical condition, and other
times to sell us a product. Here's V again, looking
to the future, say ten twenty years. The danger is
if I still don't know that I'm gay, but the
government and Coca Cola and and Amazon and Google. They
(11:47):
already know it. I'm at a huge disadvantage. So it
could be something as frightening as the secret police coming
and taking me to a concentration camp. But it could
also be something Coca Cola knowing that I'm gay, they
want to sell me a new drink, and they choose
the advertisement with the shirtless guy and not the advertisement
(12:10):
with the girl in the bikini. And next day morning
I go and I buy this soft drink and I
don't even know why, and they have this huge advantage
over me and can manipulate me in all kinds of ways. Well,
as you've all brought up soda, I was not allowed
to drink soda as a child. My parents tricked me
into thinking that Seltzer was soda. I later found out
(12:33):
that soda is soda and Seltzer is water. And somehow
the Seltzer of it all is the perfect segue because
AI is not only being used to sell a product,
it's also being used to create products like Seltzer in
the R and D research and development phase. And for
this bombs episode, Julian remember Juliana, a lovely producer, And
(12:57):
I went to meet the company behind the gastrograph app,
which is using consumer preference data to make predictions about
new flavors. People might like, I think you hate Nettle,
think against Analytical Flavor Systems or a f S is
tucked away down a side street in Chinatown, up in
(13:19):
a third story walk up. It does actually and when
we were there, the office was still waiting for furniture.
You know, it had this We're going to disrupt the
industry vibe. We moved in like a month or two ago.
But it turns out you can't just buy office furniture.
We met the founder of the company, Jason Cohen, and
we believe that in the future, in order to be competitive,
(13:41):
you have to be targeted. That that there won't be
any billion dollar brands right in ten years if you
don't have an m b A. Here's what Jason's talking about.
Think about the coffee you drank this morning, is it
Third Wave? Did you get it from Starbucks or an
indie roaster. Food and beverage companies are moving more and
more towards knee markets. The problem is that they have
(14:02):
very old school ways of developing new products. But a
FS is offering another way to reach those customers, and
that's by allowing companies to formulate more specific products using AI.
Here's Jason again. Usually the way that things are done
today is you get some conceptual brief. You might say,
we want to develop a new fruit flavored beverage for
(14:24):
Japanese millennial women. Right, you would look at what other
fruit flavored beverages are out there. You'd look at your
own product lineup and say, well, we already have a
lemon flavor, and we're going to send out these briefs,
and we're gonna send these out to these flavor companies.
We're gonna see what other fruits we can get, and
you're gonna wind up with very mainstream things. You're gonna
wind up with peach and mango and strawberry and grapefruit, right,
(14:45):
and maybe you'll wind up with something interesting like low
quad or uzo or dragon fruit. Right. And then you're
gonna have your own consumer tasting panel internally hopefully uh
and you're gonna have them taste it, and they're gonna
have to like some of those more than your current
offering or more than a competitors. After you've done all
(15:05):
of this work, which sometimes costs in the tens of
thousands of dollars in order to have the samples developed.
How the samples sent to you recruit the consumers, but
the product in front of the consumers, that data is
only ever usable once. Right. All you get from that
as a binary yes no, sixty percent of the population
liked more than the competitors, and so what we're doing
is entirely different. Jason's team wants to take product development
(15:28):
out of the yes no binary. Instead of just saying
coke or pepsi, they can calculate which parts of each
soft drink people liked and disliked, and then a f
s can make entirely new flavors based on what a
person kind of liked about pepsi and kind of liked
about coke. And finally, they can transfer those preferences to
(15:48):
entirely different demographics. So what might someone in Mexico want
to taste in their cola compared to what a Japanese
millennial woman might want to taste in her cola? This
is what Jason believes is truly disruptive, being able to
say to a brand, if you want to launch in Mexico,
you should tweak your flavors in this way because we're
actually able to collect this data, develop a data set,
(16:10):
and make predictions. So we could take data from say, white,
twenty to three year old college educated males, and use
that to predict what every other population in the United
States is going to perceive in that product. And so
we're taking an industry that has never seen predictions of
any kind before and finally being able to actually predict things,
predict who's gonna like it, how much they're gonna like
it right, and what we can do to optimize it
so that they like it more. We can actually create
(16:31):
products that no one would have ever thought of, and
no one ever would have thought that a segment of
the population would have liked. And this is something that
we now do with the companies that we work with.
We did talk quite a bit about developing a pine
flavored beverage in Japan. When we first showed these results
to a company there, they said, natzu desca, do you
mean pineapple? Because it was just so out of the yeah,
out of the ordinary. The way Jason and his team
(16:54):
are able to get such nuanced data is with an
app they developed called Gastrograph. Gastrograph looks a lot like
the flavor wheels they use in coffee and wine tasting
to help people map out what they taste when they
try a new product. We think of every flavor, romance,
texture as a signature. You can have the five basic
flavors bitter, sweet, salty, sour mommy, and then underneath that
you're gonna have categorical flavors like fruity earth, the herbaceous nuts,
(17:16):
and seedge roasted. I mentioned that he's a professional taster, right,
and then underneath that you can have subcategorical like citrus,
or specific like lemon are very specific like Meyer lemon
are very very specific like Meyer lemons as right, So
all of those signatures exist in some infinite dimensional space
flavor space. So, car, since you're such a Seltzer fiend,
(17:36):
we demo gaster graph and got a feel for the
app by doing a Seltzer tasting. Yeah, we tasted five
Seltzers that are already on the market that you'd buy,
like seven eleven, and we rated different components of them.
So if I tasted fruit, I'd rate that from zero
to five, and I could add adjectives like strawberry or mango.
You know, I'm not into like Seltzer two point oh
(17:56):
with like kumquat flavored sparkling seltzer. You know, I plane vintage,
but a f s is Resident chemist Ryan On agreed
to formulate a seltzer based on just our extremely small
data set. Um, you'd probably be a pretty fast process.
So we have tons of seltzer data. What we would
want to do is um run through a couple of
(18:19):
different flavors to get an idea of the types of
things that you like, build a model specifically around that,
run an optimization, predict a new flavor that you've never
had before, and then have you tried again. So after
we did all of that, we went home. We sat
in our hands, and then we went back to the office,
which actually a little bit more furniture when we got
(18:41):
back there to see if they actually could create a
seltzer that both Julian and I would like. So it
was a blind tasting. Here we go, all right, drink
one pair. I don't know why I think such a
thing against pears. The number two berries, that's delish. Let's
(19:02):
see what we got here. I wouldn't know huckleberry if
it hit me in the face, honestly, but whatever, Number
three honestly, curry. Do you not taste curry like there's spice?
It's so interesting? One? Oh, I love it like hops
(19:24):
and I'm si care just taste like grass? What are you?
We're done? So the first thing you guys should tell
us is which is your favorite? For me? Is number three?
Just because I like the complexity of flavor. Three was
a beverage too. I'm used to and I've probably had before.
(19:46):
Maybe three. I really enjoyed one I did not like.
I don't like those flavors. We have to reveal. We're
almost at the reveal can to clarify. Your job as
a company is to predict future products that people will
enjoy and come back to. Yes, So in that regard,
(20:07):
three was a winner. All right? Oh my god, tes us.
So we had to do one product that was going
to be optimal for both of you, and we got it.
It was number three. They nailed it. You could say
that we got flavor hacked up. Here. This this blue graph,
(20:27):
this is saying that there's a you know, a seventy
percent chance that you would give this a six? Did
I give it a six? I think I did? You? Did?
You both gave it a six? Um? So we were
pretty confident on this, but we didn't have that level
of confidence that we saw this. What I tasted is
not something I ever predicted I would have liked, but
(20:47):
it's absolutely something I will continue to think about. It
was such a unique flavor and it's actually something I
would buy. It's just that I had never tasted something
like it before. So when I first tasted it, I
was like, this is strange. In the case scenario, gastrographs
AI can help companies create foods that satisfying more specific
tastes and even bring people more joy, and that's good
(21:08):
for business. Instead of making huge bets and trying to
market a product to an entire country, a f S
has created a way to make more specialized bets and
help companies tap into those niches. And this isn't about
AI reducing our experiences to data. AI is being used
to change how we experience the world. More sleepwalkers after
(21:29):
the break So, Karen, I would have been quite nervous
to stand in the shoes of Jason and Ryan and
the gastrograph team because I know that you're such a
connoisseur of Seltzer. How do they do? They did really
(21:54):
well actually, and I think it's important to mention that
they weren't trying to create something they thought I would
already like, like I love cranberry, right. They were trying
to create something that I hadn't tasted before and also liked.
So it was really difficult, and I think it also
shows that there's a bit of reversal in the way
(22:14):
that we do things. Companies have always used market research
to predict consumer preference, but it's often based on things
like focus groups or survey research. What we have now
is massive amounts of data being funneled through an algorithm
to deliver the perfect product for a very specific type
of person exactly. That's the or yeah, age, demographic, socio
(22:37):
economic race. They can target it to all these particular categories,
and I think this is cool and also a little unsettling.
I think as humans, we like to be in control,
you know. I like to think that my preferences are
just that, my own preferences, and this sort of up
ends that notion. You know, using pre existing data, I
(22:58):
can kind of be read they're reading me, and that
makes me feel a little less special. I do think
it's cool that companies are trying to service niche markets,
and I think that's a trend I would definitely get
on board with as far as AI being used to
make predictions. And the reason this gastrograph pieces interesting is
because it's a perfect demonstration that AI is not some
(23:20):
thing which is going to happen in the future. It's
here with us today. We can literally taste it already
AIS in our lives. It's interpreting our data, is analyzing
our preferences, is predicting our behaviors. But we're just starting
(23:41):
to respond to what that means culturally. And so there's
a lot of new technologies and new issues that were
very excited to get our hands dirty with on season two. Absolutely,
and one of the important issues we're going to explore
is bias and technology. It's easy to think that algorithms
are neutral, but the reality is that technology is built
by someone, and that person's bias can be built into
(24:02):
a system. This Princeton professor named Ruha Benjamin has introduced
a concept directly related to this, called the new Gym code,
which asks us to consider the inequities encoded in algorithms. Well,
it's the algorithms and also the data they learned from right.
I mean AI harnesses the power of processing huge amounts
(24:22):
of data about things that have happened in the past
in order to predict a future, and so we have
to be very careful about what that data contains, or
we might not like the future. It's bits out. I
think it's a particularly relevant in the air of medicine.
We see huge, huge promise about honesting AI for better
medical outcomes, but we also need to be very careful
about how the data is being used and who has
(24:43):
access to it, and how can you prevent your data
from being used against you. Well, we have to think
about the potential for data to harm but also to
provide comfort and to drive innovation, sometimes extraordinary, very unexpected innovation.
There are two stories I can't wait to dive into
in season two. One is about a doctor using AI
(25:03):
to record and optimize conversational strategies with very sick patients.
What should they say, how and when? Another is about
using natural language processing to enable immersive conversations with holograms
of people from history, everyone from astronauts to Holocaust survivors.
In other words, using technology to bring history into the
(25:25):
present and ensuring we never forget our past it's wild,
so we're obviously looking forward to seeing you in the
next season. We have a lot of amazing stories lined up.
We'd love to hear from you guys about stories that
you want to hear or subjects that you want to
hear about, So tweet us at Sleepwalker's Pod on Twitter
obviously and on Instagram. We are Sleepwalkers Podcast. Yeah, thank
(25:49):
you so much. We love your feedback and we're really
looking forward to seeing you for season two very soon.
Sleepwalkers is a production of I Heart Radio and Unusual Productions.
(26:10):
For the latest AI news, live interviews, and behind the
scenes footage, find us on Instagram, at Sleepwalker's Podcast or
at Sleepwalker's podcast dot com. Sleepwalkers is hosted by me
Ozveloshin and co hosted by me Kara Price. Were produced
by Julian Weller with help from Jacopo Penzo and Taylor Chcogin,
mixing by Tristan McNeil and Julian Weller. Our story editor
(26:32):
is Matthew Riddle. Sleepwalkers is executive produced by me Ozveloshin
and Mangesh hat Together. For more podcasts from my Heart Radio,
visit the I Heart Radio app Apple podcasts, or wherever
you listen to your favorite shows.
Sleepwalkers is a production of our Heart Radio and Unusual Productions. Hi,
I'm Aloan and I'm care Price. Welcome to a special
bonus episode of Sleepwalkers. Well, Cara, it's very good to
(00:36):
be back to the office. That you It is good
to be back, as I just want to apologize to
our listeners. We don't have an algorithm that's going to
create season two, so things have taken a little longer. Yes, unfortunately,
unfortunately a I can't do everything yet. But we are
hard at work on season two of Sleepwalkers, and we're
focusing on stories that really contextualize the implications of AI,
(00:57):
what it's doing, what the future is, and how it's
affecting us. You know, we had such a good time
in season one of Sleepwalkers wrapping our heads around the
meaning of artificial intelligence. It becomes basically a principle of
statistics prediction, how we're using data to inform our decisions,
and how that's becoming ingrained in products and services and
(01:18):
everything we do. Really, it's true the Pandora's box of
AI has been opened, but we still have the black
box problem that is true explainable AI. We can't tell
what neural networks are doing yet, but people are working
on it, and that's a story we're going to be
covering closely in season two. But in this bonus episode
of Sleepwalkers, we're going to take a look back at
some of the most poignant stories and interesting applications of
(01:40):
AI that we talked about in the first season, and
later in this episode, I'm going to give you a
preview of a fascinating new company that's using AI to
predict very specific consumer taste, as in preferences, not like
tasting clothing. Before we get through though us, when you
look back at season one, what stands out to you
the most? One of the things that stands out to
(02:01):
me most is the story you did about liar Bird.
Thank you, but seriously, the way they use an algorithm
to create a deep fake of your voice. But in
that particular piece, the questions it raised about mortality and
would you want to hear your father's voice beyond the
grave has stuck with me and I was one of
the most powerful moments in the whole podcast. Yeah so,
(02:22):
Jose Satello, one of the founders of Liar Bird, had
me sit down on a microphone for an hour and
just speak which was a personal dream of mine, and
then using that out me interrupting. Using that voice data,
liar Bird's algorithms created a version of my voice that
could turn any written text into something that sounded like me.
(02:42):
But Jose actually explained it better. So can you hit
the clip? Not an AI scientist, but we do have
the sophistication to roll tape. I know it might sound
a bit like magic, but in reality, the way that
our algorithm's work is basically they are just a battern
matching algorithms, and so it's trying to figure out how
(03:05):
to identify the patterns in your voice by comparing it
against thousands of other voices a shoually tens of thousands
of other voices, and trying to figure out what is
it that makes your voice unique. Once Jose's algorithms identified
what was unique about my voice, obviously everything they had
the building blocks they needed to make a fake. Then
(03:25):
we sent Jose a set of sentences we wanted robot
care to say, and he used another set of algorithms
to turn the text into what we heard. The way
they do this is they use it's called a generative
adversarial network again, which is a system where one neural
net tries to trick another one a thousand times per second,
So each time the second network to texts of fake,
(03:47):
the first one tries again. It basically learns from its mistakes,
and once it tricks its adversary, it's ready to show
its results. In our case, liar bird pits my fake
voice against my real voice until it sounds like this
sub dog Scara Karen. One of the reasons why this
story has stuck with me is because it feels like
we're just at the beginning of tapping the potential and
(04:10):
the potential for harm of deep fakes. May be remembered
as the year of the first significant deep fake crime.
An employee at a UK based energy firm believed he
was on the phone to his boss and followed instructions
to transfer two hundred thousand pounds to a scammers bank account.
That certainly won't be the last deep fatefore we hear of,
(04:30):
and it raises questions about responsibility and accountability. Who's liable
in a case like this. Facebook has actually gone so
far as to create a deep fake detection challenge to
get the best minds thinking about deep fakes and how
we might solve the problem and offering like a million
dollar prize reach prize, which is like a dollar but
it also shows how important the issue is, especially when
(04:52):
a company like Facebook gets behind it. Um. There's another
side to deep fake technology that actually highlights this dichotomy
and tech anology right now, which is that it can
be used for menacing purposes but also really powerful and
beautiful applications. Jose goes on to talk about how liar
Bird can be used with als patients and give them
the ability to speak when they've lost all ability to speak,
(05:14):
when it could give them the opportunity to speak in
a version of their voice to their children again, which
is quite profound. One area where I think technology is
a powerhouse for change is in medicine. Technologists and doctors
alike are looking at AI to predict, treat, and diagnose,
you know, everything from depression to cancer, and that's a
very wide spectrum. And it reminds me of one of
(05:36):
my favorite interviews that you did, which was your interview
with Saddartha. Muker g just so fascinated by this article
he'd written for The New Yoker called AI versus m
D which laid out all of the kind of benefits
and potential applications of using AI in medicine, including some
of the downside such as the black box problem of
AI that you mentioned, not knowing why an algorithm has
(05:57):
made a recommendation, and also another problem, which is that
if we really too much on technology, it can erode
human skills. There is a fear that AI could move
us into a very black and white way of thinking.
The computer says you have cancer, or the computer says
you to have your liver removed, said Arthur, who is
one of the world's foremost oncologists and a Pullet Surprise
winning author. Provided a different perspective. There is something very
(06:21):
fundamental about the human brain, a scientist's brain, a doctor's brain,
and artist's brain that asks questions in a fundamentally different manner,
the why question. Why did this happen in this person
in this time? Why does the melanoma appear in the
first calase? What is the molecular basis of that appearance.
The most interesting mysteries of medicine remain mysteries that have
(06:41):
to do with the why. Once we give up some
of the diagnostic pattern recognition material to machines, it will
be time to play. It will be the time to
play in the arena of human therapeutics, human biology, the
complexity of the human interaction, the art of medicine. My
hope is that medicine, in being more playful, will become
(07:01):
more compassionate, more able to take into account individuals and
their individual destinies rather than bucketing people in big categories.
It means having more time to spend with humans. You know,
we are so constrained by time that even compassion gets
three minutes, We won't become more robotic, will become less
(07:23):
robotic as the robots enter our own So Dartha's point
is that these tools could make doctors more efficient so
that they can provide better care. It sort of takes
the grunt work out of medicine and puts the patient
care work back in the doctor's hands. This idea that
technology can actually allow us to be more human, make
(07:46):
us more empathetic, is fascinating, and it also raises the
questions about new types of skills that may need to
be developed in an age of AI. Yeah, and Regina
Barslay from m I T spoke a lot about this.
How doctors have to now equip themselves with new ways
of translating data to patients, we still do not communicate
it to the patient because I think now there is
(08:10):
a walk to be done, not on computer science or AIPAD,
but really on the clinical side. What is the best
way to communicate it to the patient and what is
um you know, the past that you're going to give them.
It is not just enough to say you know you
are high risk. You need to propose some suggestion and solution.
(08:31):
So currently the clinical stuff is thinking and looking at
the ways of effective you know, clinical engagement with a patient.
You know, I speaking of data. You all know Harri
was another person who made you and me think about
humans as reducible to data. I think he's mostly known
as a historian and for his book Sapiens, but you
(08:52):
spoke with him about the data we produce as humans
and how that influences our relationship with technology. That's right,
which is the topic of his book Daus. And he
has this phrase data is m to describe how we've
kind of come to worship the data we create and
our own technological creations. So what happens when based on
all of our past behavior, AI starts to know us
(09:15):
better than we know ourselves. Here's a clip from You
are talking about exactly that. When we talk about AI,
we tend to greatly exaggerate the potential abilities, but at
the same time we also tend to exaggerate the abilities
of humans. People say that AI is not going to
take over our lives because it's very imperfect and it
(09:39):
won't be able to know us perfectly. But what people
forget is that humans often have a very poor understanding
of themselves, of the desires, of their emotions, of their
mental states. For AI to take over your life, it
doesn't need to know you perfectly, just need to know
(10:01):
you better than you know yourself. And that's not very
difficult because we often don't know the most important things
about ourselves. So, but let's say you could turn back
the clock to being fifteen, would you have wanted to
live in a world where there was sufficiently good sensors
to monitor your eyes, your eye movement, you're breathing, you know,
(10:21):
while you're going about your daily life, and then to
interpret that and say to you you have all more
likely than not you're gay. That's a very good question
which will become very practical questions in a few years.
And the way that I grew up and developed. It
would have been a very bad idea. I wouldn't like
(10:42):
to receive this kind of insight from form a machine.
I'm not sure how I would have dealt with it
when I was fifteen, you know, in Israel in the
nineteen eighties, and maybe partly it was you know, a
defense mechanism. In the future too. It it depends where
you live. Brunei has instituted the death penalty for gay people,
at least for people engaged in homosexual sex. So if
(11:05):
I'm a teenager in Brunei, I don't want to be
told by the computer that I'm gay, because the computer
will then be able to tell that to the police
and to the authorities as well. So the apps we use,
the product we buy, the number of steps we take,
the delivery I ordered last night, that will becomes data,
and that data can feed into neural networks to create
(11:27):
statistical models of us and what we might do next,
sometimes in order to diagnose a medical condition, and other
times to sell us a product. Here's V again, looking
to the future, say ten twenty years. The danger is
if I still don't know that I'm gay, but the
government and Coca Cola and and Amazon and Google. They
(11:47):
already know it. I'm at a huge disadvantage. So it
could be something as frightening as the secret police coming
and taking me to a concentration camp. But it could
also be something Coca Cola knowing that I'm gay, they
want to sell me a new drink, and they choose
the advertisement with the shirtless guy and not the advertisement
(12:10):
with the girl in the bikini. And next day morning
I go and I buy this soft drink and I
don't even know why, and they have this huge advantage
over me and can manipulate me in all kinds of ways. Well,
as you've all brought up soda, I was not allowed
to drink soda as a child. My parents tricked me
into thinking that Seltzer was soda. I later found out
(12:33):
that soda is soda and Seltzer is water. And somehow
the Seltzer of it all is the perfect segue because
AI is not only being used to sell a product,
it's also being used to create products like Seltzer in
the R and D research and development phase. And for
this bombs episode, Julian remember Juliana, a lovely producer, And
(12:57):
I went to meet the company behind the gastrograph app,
which is using consumer preference data to make predictions about
new flavors. People might like, I think you hate Nettle,
think against Analytical Flavor Systems or a f S is
tucked away down a side street in Chinatown, up in
(13:19):
a third story walk up. It does actually and when
we were there, the office was still waiting for furniture.
You know, it had this We're going to disrupt the
industry vibe. We moved in like a month or two ago.
But it turns out you can't just buy office furniture.
We met the founder of the company, Jason Cohen, and
we believe that in the future, in order to be competitive,
(13:41):
you have to be targeted. That that there won't be
any billion dollar brands right in ten years if you
don't have an m b A. Here's what Jason's talking about.
Think about the coffee you drank this morning, is it
Third Wave? Did you get it from Starbucks or an
indie roaster. Food and beverage companies are moving more and
more towards knee markets. The problem is that they have
(14:02):
very old school ways of developing new products. But a
FS is offering another way to reach those customers, and
that's by allowing companies to formulate more specific products using AI.
Here's Jason again. Usually the way that things are done
today is you get some conceptual brief. You might say,
we want to develop a new fruit flavored beverage for
(14:24):
Japanese millennial women. Right, you would look at what other
fruit flavored beverages are out there. You'd look at your
own product lineup and say, well, we already have a
lemon flavor, and we're going to send out these briefs,
and we're gonna send these out to these flavor companies.
We're gonna see what other fruits we can get, and
you're gonna wind up with very mainstream things. You're gonna
wind up with peach and mango and strawberry and grapefruit, right,
(14:45):
and maybe you'll wind up with something interesting like low
quad or uzo or dragon fruit. Right. And then you're
gonna have your own consumer tasting panel internally hopefully uh
and you're gonna have them taste it, and they're gonna
have to like some of those more than your current
offering or more than a competitors. After you've done all
(15:05):
of this work, which sometimes costs in the tens of
thousands of dollars in order to have the samples developed.
How the samples sent to you recruit the consumers, but
the product in front of the consumers, that data is
only ever usable once. Right. All you get from that
as a binary yes no, sixty percent of the population
liked more than the competitors, and so what we're doing
is entirely different. Jason's team wants to take product development
(15:28):
out of the yes no binary. Instead of just saying
coke or pepsi, they can calculate which parts of each
soft drink people liked and disliked, and then a f
s can make entirely new flavors based on what a
person kind of liked about pepsi and kind of liked
about coke. And finally, they can transfer those preferences to
(15:48):
entirely different demographics. So what might someone in Mexico want
to taste in their cola compared to what a Japanese
millennial woman might want to taste in her cola? This
is what Jason believes is truly disruptive, being able to
say to a brand, if you want to launch in Mexico,
you should tweak your flavors in this way because we're
actually able to collect this data, develop a data set,
(16:10):
and make predictions. So we could take data from say, white,
twenty to three year old college educated males, and use
that to predict what every other population in the United
States is going to perceive in that product. And so
we're taking an industry that has never seen predictions of
any kind before and finally being able to actually predict things,
predict who's gonna like it, how much they're gonna like
it right, and what we can do to optimize it
so that they like it more. We can actually create
(16:31):
products that no one would have ever thought of, and
no one ever would have thought that a segment of
the population would have liked. And this is something that
we now do with the companies that we work with.
We did talk quite a bit about developing a pine
flavored beverage in Japan. When we first showed these results
to a company there, they said, natzu desca, do you
mean pineapple? Because it was just so out of the yeah,
out of the ordinary. The way Jason and his team
(16:54):
are able to get such nuanced data is with an
app they developed called Gastrograph. Gastrograph looks a lot like
the flavor wheels they use in coffee and wine tasting
to help people map out what they taste when they
try a new product. We think of every flavor, romance,
texture as a signature. You can have the five basic
flavors bitter, sweet, salty, sour mommy, and then underneath that
you're gonna have categorical flavors like fruity earth, the herbaceous nuts,
(17:16):
and seedge roasted. I mentioned that he's a professional taster, right,
and then underneath that you can have subcategorical like citrus,
or specific like lemon are very specific like Meyer lemon
are very very specific like Meyer lemons as right, So
all of those signatures exist in some infinite dimensional space
flavor space. So, car, since you're such a Seltzer fiend,
(17:36):
we demo gaster graph and got a feel for the
app by doing a Seltzer tasting. Yeah, we tasted five
Seltzers that are already on the market that you'd buy,
like seven eleven, and we rated different components of them.
So if I tasted fruit, I'd rate that from zero
to five, and I could add adjectives like strawberry or mango.
You know, I'm not into like Seltzer two point oh
(17:56):
with like kumquat flavored sparkling seltzer. You know, I plane vintage,
but a f s is Resident chemist Ryan On agreed
to formulate a seltzer based on just our extremely small
data set. Um, you'd probably be a pretty fast process.
So we have tons of seltzer data. What we would
want to do is um run through a couple of
(18:19):
different flavors to get an idea of the types of
things that you like, build a model specifically around that,
run an optimization, predict a new flavor that you've never
had before, and then have you tried again. So after
we did all of that, we went home. We sat
in our hands, and then we went back to the office,
which actually a little bit more furniture when we got
(18:41):
back there to see if they actually could create a
seltzer that both Julian and I would like. So it
was a blind tasting. Here we go, all right, drink
one pair. I don't know why I think such a
thing against pears. The number two berries, that's delish. Let's
(19:02):
see what we got here. I wouldn't know huckleberry if
it hit me in the face, honestly, but whatever, Number
three honestly, curry. Do you not taste curry like there's spice?
It's so interesting? One? Oh, I love it like hops
(19:24):
and I'm si care just taste like grass? What are you?
We're done? So the first thing you guys should tell
us is which is your favorite? For me? Is number three?
Just because I like the complexity of flavor. Three was
a beverage too. I'm used to and I've probably had before.
(19:46):
Maybe three. I really enjoyed one I did not like.
I don't like those flavors. We have to reveal. We're
almost at the reveal can to clarify. Your job as
a company is to predict future products that people will
enjoy and come back to. Yes, So in that regard,
(20:07):
three was a winner. All right? Oh my god, tes us.
So we had to do one product that was going
to be optimal for both of you, and we got it.
It was number three. They nailed it. You could say
that we got flavor hacked up. Here. This this blue graph,
(20:27):
this is saying that there's a you know, a seventy
percent chance that you would give this a six? Did
I give it a six? I think I did? You? Did?
You both gave it a six? Um? So we were
pretty confident on this, but we didn't have that level
of confidence that we saw this. What I tasted is
not something I ever predicted I would have liked, but
(20:47):
it's absolutely something I will continue to think about. It
was such a unique flavor and it's actually something I
would buy. It's just that I had never tasted something
like it before. So when I first tasted it, I
was like, this is strange. In the case scenario, gastrographs
AI can help companies create foods that satisfying more specific
tastes and even bring people more joy, and that's good
(21:08):
for business. Instead of making huge bets and trying to
market a product to an entire country, a f S
has created a way to make more specialized bets and
help companies tap into those niches. And this isn't about
AI reducing our experiences to data. AI is being used
to change how we experience the world. More sleepwalkers after
(21:29):
the break So, Karen, I would have been quite nervous
to stand in the shoes of Jason and Ryan and
the gastrograph team because I know that you're such a
connoisseur of Seltzer. How do they do? They did really
(21:54):
well actually, and I think it's important to mention that
they weren't trying to create something they thought I would
already like, like I love cranberry, right. They were trying
to create something that I hadn't tasted before and also liked.
So it was really difficult, and I think it also
shows that there's a bit of reversal in the way
(22:14):
that we do things. Companies have always used market research
to predict consumer preference, but it's often based on things
like focus groups or survey research. What we have now
is massive amounts of data being funneled through an algorithm
to deliver the perfect product for a very specific type
of person exactly. That's the or yeah, age, demographic, socio
(22:37):
economic race. They can target it to all these particular categories,
and I think this is cool and also a little unsettling.
I think as humans, we like to be in control,
you know. I like to think that my preferences are
just that, my own preferences, and this sort of up
ends that notion. You know, using pre existing data, I
(22:58):
can kind of be read they're reading me, and that
makes me feel a little less special. I do think
it's cool that companies are trying to service niche markets,
and I think that's a trend I would definitely get
on board with as far as AI being used to
make predictions. And the reason this gastrograph pieces interesting is
because it's a perfect demonstration that AI is not some
(23:20):
thing which is going to happen in the future. It's
here with us today. We can literally taste it already
AIS in our lives. It's interpreting our data, is analyzing
our preferences, is predicting our behaviors. But we're just starting
(23:41):
to respond to what that means culturally. And so there's
a lot of new technologies and new issues that were
very excited to get our hands dirty with on season two. Absolutely,
and one of the important issues we're going to explore
is bias and technology. It's easy to think that algorithms
are neutral, but the reality is that technology is built
by someone, and that person's bias can be built into
(24:02):
a system. This Princeton professor named Ruha Benjamin has introduced
a concept directly related to this, called the new Gym code,
which asks us to consider the inequities encoded in algorithms. Well,
it's the algorithms and also the data they learned from right.
I mean AI harnesses the power of processing huge amounts
(24:22):
of data about things that have happened in the past
in order to predict a future, and so we have
to be very careful about what that data contains, or
we might not like the future. It's bits out. I
think it's a particularly relevant in the air of medicine.
We see huge, huge promise about honesting AI for better
medical outcomes, but we also need to be very careful
about how the data is being used and who has
(24:43):
access to it, and how can you prevent your data
from being used against you. Well, we have to think
about the potential for data to harm but also to
provide comfort and to drive innovation, sometimes extraordinary, very unexpected innovation.
There are two stories I can't wait to dive into
in season two. One is about a doctor using AI
(25:03):
to record and optimize conversational strategies with very sick patients.
What should they say, how and when? Another is about
using natural language processing to enable immersive conversations with holograms
of people from history, everyone from astronauts to Holocaust survivors.
In other words, using technology to bring history into the
(25:25):
present and ensuring we never forget our past it's wild,
so we're obviously looking forward to seeing you in the
next season. We have a lot of amazing stories lined up.
We'd love to hear from you guys about stories that
you want to hear or subjects that you want to
hear about, So tweet us at Sleepwalker's Pod on Twitter
obviously and on Instagram. We are Sleepwalkers Podcast. Yeah, thank
(25:49):
you so much. We love your feedback and we're really
looking forward to seeing you for season two very soon.
Sleepwalkers is a production of I Heart Radio and Unusual Productions.
(26:10):
For the latest AI news, live interviews, and behind the
scenes footage, find us on Instagram, at Sleepwalker's Podcast or
at Sleepwalker's podcast dot com. Sleepwalkers is hosted by me
Ozveloshin and co hosted by me Kara Price. Were produced
by Julian Weller with help from Jacopo Penzo and Taylor Chcogin,
mixing by Tristan McNeil and Julian Weller. Our story editor
(26:32):
is Matthew Riddle. Sleepwalkers is executive produced by me Ozveloshin
and Mangesh hat Together. For more podcasts from my Heart Radio,
visit the I Heart Radio app Apple podcasts, or wherever
you listen to your favorite shows.
kill switch News
Hosts And Creators
Oz Woloshyn
Karah Preiss
Show Links
AboutPopular Podcasts
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
24/7 News: The Latest
The latest news in 4 minutes updated every hour, every day.
Therapy Gecko
An unlicensed lizard psychologist travels the universe talking to strangers about absolutely nothing. TO CALL THE GECKO: follow me on https://www.twitch.tv/lyleforever to get a notification for when I am taking calls. I am usually live Mondays, Wednesdays, and Fridays but lately a lot of other times too. I am a gecko.