Latest / What's Your Problem? / Teaching Computers to Smell

Transcript

Teaching Computers to Smell · What's Your Problem?

0:15
Speaker 1: Pushkin, tell me what we don't know about how smell works?
0:23
Speaker 1: Oh jeez, be sure to tell you what we do.
0:27
Speaker 1: This is Alex Wilscoe. He's the co founder and CEO
0:30
Speaker 1: of a company called Osmo, and despite his protests there
0:34
Speaker 1: he did tell me some of the things nobody knows
0:37
Speaker 1: about how smell works.
0:40
Speaker 2: Why do things smell the way that they do. Why
0:43
Speaker 2: can we smell certain things and not other things. What
0:47
Speaker 2: is the logic of how molecules are combined to create
0:51
Speaker 2: beautiful smells? Why do some smells create incredibly powerful emotional
0:56
Speaker 2: associations instantly and others seem neutral? Right? Why do something
1:01
Speaker 2: smell different to people? I think we have a hints
1:04
Speaker 2: in all these directions, but we have nothing like musical scale,
1:09
Speaker 2: where we have nothing like a periodic table. We don't
1:11
Speaker 2: know any structure to why things are the way that
1:14
Speaker 2: they are. It's a ton of mystery, and that's what
1:17
Speaker 2: makes it so exciting to work on this topic, is
1:20
Speaker 2: like there's so much we don't know, and to.
1:22
Speaker 1: Be clear, like with light, we just know whatever. If
1:27
Speaker 1: you tell me the frequency the wavelength, I can know
1:31
Speaker 1: exactly what color you're talking about, or the same thing
1:34
Speaker 1: with a wave form of sound, right, and so but
1:36
Speaker 1: if I give you some random molecule and say what
1:39
Speaker 1: does it smell like?
1:40
Speaker 2: Do you know? So that's what I've spent a lot
1:44
Speaker 2: of my professional life working on. It is exactly that question. Yeah,
1:46
Speaker 2: which is a draw structure of a molecule on a whiteboard,
1:49
Speaker 2: point out it and say, hey, what does this smell
1:51
Speaker 2: like wood or flowers or fruits or whatever? And so
1:56
Speaker 2: there is no way to know that for sure at all.
2:01
Speaker 2: But there's no good way even statistically to predict that
2:05
Speaker 2: without using large data sets, and at least in our hands.
2:09
Speaker 2: You need neural networks, you need deep sorting in order
2:11
Speaker 2: to do that. I'm Jacob Goldstein, and this is what's
2:19
Speaker 2: your problem?
2:20
Speaker 1: The show where I talk to people who are trying
2:22
Speaker 1: to make technological progress.
2:24
Speaker 2: Alex Wilch.
2:25
Speaker 1: Goo's problem is this, Can you use AI to teach
2:29
Speaker 1: computers to smell? And once you figured out how to
2:33
Speaker 1: do that, can you build a profitable business around it?
2:36
Speaker 1: Osmo's fun out of Google in twenty twenty three. The
2:39
Speaker 1: company recently launched a fragrance house to develop new perfumes.
2:43
Speaker 1: They've also done some work using scent to detect counterfeit shoes,
2:47
Speaker 1: and in the long run, they plan to use scent
2:50
Speaker 1: to diagnose disease. Before he started Osmo, Alex worked at
2:55
Speaker 1: Google as an AI researcher. Before that, he got a
2:58
Speaker 1: PhD at Harvard studying how mice respond to scent. But
3:03
Speaker 1: maybe the most important part of his bio came even
3:05
Speaker 1: earlier in his life, specifically when he was twelve years
3:09
Speaker 1: old and went off to summer camp in his home state, Texas.
3:13
Speaker 2: I was from a small town college station, and then
3:16
Speaker 2: most of the kids were from big towns like Houston
3:18
Speaker 2: and Dallas and Austin and San Antonio, and I hadn't
3:23
Speaker 2: really been exposed to, like I don't know fashion trends
3:28
Speaker 2: or you know, what was cool or popular. But everybody's
3:32
Speaker 2: all lumped together in summer camp. And then there is
3:34
Speaker 2: this thing called perfume that some of the richer, frankly
3:38
Speaker 2: richer and more popular kids had, And it was just
3:42
Speaker 2: amazing to me that these boys could spray themselves with
3:48
Speaker 2: this invisible mist, a clear mist, and then for the
3:51
Speaker 2: next four to six hours people around them would treat
3:54
Speaker 2: them differently, and that just blew my mind right, like
3:59
Speaker 2: there's no I can't I can see the clothes. Yeah,
4:01
Speaker 2: I can see how they act and walk and talk
4:03
Speaker 2: and how they you know, posture and all that, but
4:06
Speaker 2: I cannot see the fragrance. Yeah, but yet it is
4:10
Speaker 2: obviously doing something magical. It's like an.
4:13
Speaker 1: Axe body spray ad what what does that cause you
4:17
Speaker 1: to do?
4:18
Speaker 2: When I get home, beg my parents to buy and fail.
4:25
Speaker 2: We shopped at TJ Max and I started to really
4:28
Speaker 2: look out for fragrances there and then it just kind
4:31
Speaker 2: of snowballs from there, which I just realized there was
4:33
Speaker 2: like a whole lot of these things. I guess what,
4:34
Speaker 2: you can just try them, and some of them are
4:38
Speaker 2: actually way better and more opinionated and more beautiful. I
4:41
Speaker 2: don't I didn't have the vocabulary then, but like it
4:44
Speaker 2: just it was clear to me early on that like
4:47
Speaker 2: I never really thought about who made the clothes, but
4:49
Speaker 2: I started to think about who made these perfumes, because
4:53
Speaker 2: it was clear that there were choices that were being made.
4:55
Speaker 2: And like I just remember trying, and this is years later,
4:59
Speaker 2: trying Bulgari black, which really kind of clued me into
5:03
Speaker 2: this world. Bulgary black is not necessarily a great fragrance,
5:07
Speaker 2: but you can experience the top, middle and base notes
5:09
Speaker 2: in like forty minutes. Forty five minutes is pretty short,
5:12
Speaker 2: and so like a bigger fragrance, like a creed, events
5:15
Speaker 2: will last on your skin for a day, and so
5:18
Speaker 2: the whole fragrance unfolds. I mean, top notes will last
5:22
Speaker 2: max Fifteen thirty minutes, but the heart might last for
5:25
Speaker 2: several hours, and the base note might last for ten hours. Right,
5:28
Speaker 2: so it smells different.
5:30
Speaker 1: You can still smell it, but it smells different whatever
5:32
Speaker 1: an hour after.
5:33
Speaker 2: You put it on and four hours after that, because
5:35
Speaker 2: a great fragrance is actually many different fragrances within it. Yeah, Right,
5:39
Speaker 2: There's the first one which peels off or that quickly
5:41
Speaker 2: burns off quickly. There's the second fragrance, which is the
5:43
Speaker 2: heart note, which will last for you know, sometimes hours,
5:46
Speaker 2: but in this case, like another twenty minutes. And then
5:49
Speaker 2: the base note, which is a third fragrance, which is
5:51
Speaker 2: what's left after those two burn off. And it's like
5:55
Speaker 2: three acts of a movie, right, I think it's quite beautiful.
5:58
Speaker 2: So how do we.
5:58
Speaker 1: Get from you being a teenager preoccupied with fragrance to
6:05
Speaker 1: you using AI to predict how malls will smell?
6:10
Speaker 2: Yeah, Like the computer part was always different from the
6:13
Speaker 2: fragrance part I just I love computers. We always had
6:16
Speaker 2: computers at home. I started programming around I don't know,
6:19
Speaker 2: eight years old. It was my life, Like, my entire
6:22
Speaker 2: life was computers for a long time, still lose in
6:24
Speaker 2: a way, and fragrance was not a part of it.
6:28
Speaker 2: I got into you know, statistics, which became machine learning
6:31
Speaker 2: around the same time. Again for totally independent reason. There's
6:33
Speaker 2: this thing called the Netflix Prize. It was like one
6:35
Speaker 2: of the first competitions to build great mL algorithms. I competed.
6:40
Speaker 1: Now, that's basically to tell me what else I'll like
6:42
Speaker 1: on Netflix, right, That's what that contaest was. Like, if
6:44
Speaker 1: I've watched whatever, if I watched Succession and the Sopranos,
6:49
Speaker 1: what should I watch next?
6:51
Speaker 2: Then you're gonna like another kind of dark, but you know, funny,
6:54
Speaker 2: kind of soap opera type of a thing, exactly. And
6:57
Speaker 2: so Netflix did a really bold thing, which is they
6:59
Speaker 2: released a data set and said, here's what good looks like,
7:01
Speaker 2: here's how we measure it. Have at it. And they
7:04
Speaker 2: paid a million dollars to the Winter, which was a
7:06
Speaker 2: combination of a few teams. But what they really did
7:08
Speaker 2: is they brought a particular kind of machine learning into
7:13
Speaker 2: the forefront called collaborative filtering. Really showed that this stuff worked,
7:17
Speaker 2: and by the way, other companies were already racing to
7:20
Speaker 2: use this, So like this recommended systems was a big thing,
7:24
Speaker 2: but Netflix was putting it out into the public and
7:26
Speaker 2: allowed a kid like me I think I was eighteen
7:28
Speaker 2: or nineteen years old to actually compete and pretty well
7:30
Speaker 2: in that. And so I just got exposed to this
7:33
Speaker 2: world through that and it was super fun. I mean
7:36
Speaker 2: they gamified it and had I had a blast. So
7:38
Speaker 2: that was my first exposure to machine learning.
7:40
Speaker 1: Turned out to be a good time to start working
7:42
Speaker 1: on machine learning, Yeah.
7:44
Speaker 2: Totally, because if I had started now, they wouldn't let
7:46
Speaker 2: me in because that's probably ten years ago. Yeah, ten
7:50
Speaker 2: years ago. Yeah. And then, you know, I was doing
7:55
Speaker 2: my undergraduate training in neuroscience, and I was studying more
7:58
Speaker 2: behavior than old action because it actually turned out that
8:02
Speaker 2: olf action was a hyper specialized sub field of neuroscience.
8:05
Speaker 2: I didn't realize how niche it was. I loved smell
8:08
Speaker 2: and I was doing neuroscience, and I knew I wanted
8:10
Speaker 2: to do smell neuroscience the fancy word for nactual factory neuroscience.
8:14
Speaker 2: And so there's really two universities in the world that
8:17
Speaker 2: like have a critical mass of these researchers. It's Columbia
8:20
Speaker 2: and it's Harvard, and I applied to both. I went
8:23
Speaker 2: to Harvard, and I realized nobody cares about this problem.
8:28
Speaker 2: Nobody cares about why molecules smell the way that they do.
8:32
Speaker 2: There's a much longer conversation as so why that's the
8:34
Speaker 2: case and why that's still persistent. Now that's changing. Well,
8:37
Speaker 2: let me ask you this. Let me ask you this.
8:38
Speaker 1: At that time, I mean I get it as a
8:41
Speaker 1: basic research question.
8:42
Speaker 2: I mean, I'll tell you.
8:43
Speaker 1: We was talking with the producer and editor of this show,
8:46
Speaker 1: and we were getting ready for this interview, and we
8:48
Speaker 1: had this interesting conversation talking about scent and what you're
8:50
Speaker 1: working on whatever. And then I went down and I
8:53
Speaker 1: saw my daughter, and so, what are you organized? Said
8:55
Speaker 1: this guy who's trying to figure out scent and teach
8:57
Speaker 1: computers to smell?
8:58
Speaker 2: And she said why, I said, I don't know.
9:01
Speaker 1: I ask, so why was it compelling to I get
9:06
Speaker 1: it as a basic research question, but at that time
9:08
Speaker 1: was like, there were there applications that came to your mind.
9:12
Speaker 3: Look, the the steps of development of this thing that's
9:20
Speaker 3: now OSMO went through those different iterations.
9:23
Speaker 2: You know, I started as an academic scientist and I
9:26
Speaker 2: was trained in that world, and then I left and
9:30
Speaker 2: it did some entrepreneurship, but it ended up in industrial research,
9:33
Speaker 2: and they're like being curious, frankly was enough, and the
9:36
Speaker 2: idea this is this is Google, this is now a
9:38
Speaker 2: Google Brain. Yeah, and there's a few steps in between.
9:41
Speaker 1: But basically as you're an AI researcher at Google at.
9:44
Speaker 2: This moment when you're doing industrial research, right, yeah, exactly,
9:46
Speaker 2: and Google Brain at the time now it's Google Deep
9:49
Speaker 2: Mind very much had like a thousand flowers bloom mentality,
9:52
Speaker 2: and so people were working on crazy stuff, including me
9:54
Speaker 2: working on something Bell Labs.
9:56
Speaker 1: It's basically like Bell Labs of the twenty first century, right, you.
9:59
Speaker 2: Have it, exactly, Bell Labs, Erox Park, that kind of
10:01
Speaker 2: avid and it truly was dreamy, sounds dreamy. It was awesome, right,
10:06
Speaker 2: And it was also a moment in time, and now
10:08
Speaker 2: I think that moment's going for better or for worse.
10:12
Speaker 2: The idea was pretty straightforward for Google, which was, are
10:17
Speaker 2: the products at Google know what the world looks like
10:19
Speaker 2: and know what the world sounds like, and that's useful. Right,
10:21
Speaker 2: that's information that Google's organizing. If we knew what things
10:25
Speaker 2: smelled and tasted like, that would be useful, right. Uh huh.
10:28
Speaker 1: The original mission of Google is organized the world's information, right.
10:32
Speaker 2: Exactly, and make it universally accessible and useful. And there
10:35
Speaker 2: was a whole slice of reality, huh, the chemical slice
10:38
Speaker 2: of reality that was invisible, right, not just to Google
10:41
Speaker 2: but two computers. Yeah. Yeah, and that felt really important,
10:45
Speaker 2: and we had agreement and buying all the way up
10:47
Speaker 2: to the executive level. They're like, yeah, let's go, let's
10:49
Speaker 2: go look at that.
10:50
Speaker 1: So you're doing a like basic AI research at Google
10:53
Speaker 1: and you decide to see if you can basically use
10:57
Speaker 1: AI to figure out scent, to say, here's a molecule,
11:01
Speaker 1: what does it smell like?
11:02
Speaker 2: Right, that's the basic endeavor. How do you do that?
11:05
Speaker 2: What is it that you actually do? Yeah, so first
11:08
Speaker 2: it starts to innovation, which I was like, let's figure
11:10
Speaker 2: out smell. But it actually was a lot more natural
11:14
Speaker 2: than I think it sounds, which is scent is just chemistry.
11:18
Speaker 2: It's molecules, and we got to do AI for molecules. Right.
11:22
Speaker 2: If we're going to do AI for scent, and the
11:25
Speaker 2: thing that had happened in between. You know, there's a
11:28
Speaker 2: five year period between my academic life and my industrial life,
11:34
Speaker 2: and what had happened in those five years is actually
11:35
Speaker 2: some of the people I did my PhD with and
11:37
Speaker 2: then some of the people I ended up working with
11:39
Speaker 2: at Google Brain really cracked machine learning or AI for molecules.
11:45
Speaker 2: But they didn't do it for scent. They did it
11:46
Speaker 2: for a few other things. They did for drug discovery,
11:49
Speaker 2: and they did it for like materials discovery, so like
11:53
Speaker 2: new materials for LEDs.
11:54
Speaker 1: Right, So you happen to be doing essentially basic research
11:59
Speaker 1: at Google at this moment when there is this new
12:03
Speaker 1: way to use AI that is well suited to molecules,
12:06
Speaker 1: and you say, we can do let's do it, Yeah,
12:11
Speaker 1: let's do it.
12:11
Speaker 2: Yeah, we can do it. The other pieces are great.
12:13
Speaker 2: You got the algorithm. Where's the data? Classic? That's the
12:16
Speaker 2: classic AI question, right, like, exactly where's the data? What
12:19
Speaker 2: I did know just from being obsessed and in this
12:23
Speaker 2: world for a long time prior to that, there were
12:25
Speaker 2: these collections of data sets that were honestly really more
12:29
Speaker 2: like magazine catalogs for fragrance ingredients, and so there were
12:33
Speaker 2: these catalogs basically saying this is the ingredient, this is
12:37
Speaker 2: the molecular structure of this ingredient, and here's what it
12:39
Speaker 2: smells like. And by the way, the rating of what
12:42
Speaker 2: it smells like was done by a professional, by a perfumer.
12:45
Speaker 2: And so the special sauce that we added is we
12:48
Speaker 2: we went and we got that data and we fused
12:50
Speaker 2: a few data sets together, and we cleaned it very carefully,
12:54
Speaker 2: and that that hadn't been done.
12:55
Speaker 1: And it's something it's like five thousand dish right, it's
12:58
Speaker 1: five thousand or so different molecules.
13:00
Speaker 2: Okay, yep, exactly. And here is this the one with
13:04
Speaker 2: the list. I love the list. Here I have it.
13:06
Speaker 1: This the one sweet fruity, vanilla, powdery, fluoral, barry, fermented,
13:11
Speaker 1: nutty ozone, buttery musk.
13:12
Speaker 2: It's it's that lispright. Those are they? And there's one
13:15
Speaker 2: hundred and thirty eight of those descriptors I think that
13:18
Speaker 2: we used in that data set. Sometimes we use smaller subsets,
13:21
Speaker 2: but the full set originally is about one hundred and forty.
13:24
Speaker 1: So okay, so you have your whatever, your five thousand
13:28
Speaker 1: molecules labeled with one hundred and forty different sets. You
13:32
Speaker 1: train your AI model on this data set, and then
13:35
Speaker 1: you want to find out does the model work?
13:38
Speaker 2: Does the AI work right?
13:39
Speaker 1: If I give the model some new molecule molecule that
13:42
Speaker 1: wasn't in the training data, will it know what that
13:45
Speaker 1: molecule smells like? And to test that to answer that question,
13:49
Speaker 1: you actually.
13:50
Speaker 2: Do this study.
13:51
Speaker 1: So you get a bunch of people to smell these
13:53
Speaker 1: molecules that are not that your model was not trained on,
13:56
Speaker 1: essentially right, and say what it smells like? It's weird
14:00
Speaker 1: like what? You don't actually care what it fundamentally smells like.
14:03
Speaker 1: You just care what everybody on average thinks it smells.
14:05
Speaker 2: Like, because guess what, that's what it's what smell is? Yeah,
14:08
Speaker 2: that's what do you think of smell? Yeah? Ye?
14:11
Speaker 1: So you asked this panel to what do all these
14:16
Speaker 1: molecule smell like? And then you ask the model what
14:18
Speaker 1: do they smell like? And you compare the results and
14:20
Speaker 1: how does the model do?
14:23
Speaker 2: That was really the threshold of breakthrough in my mind
14:27
Speaker 2: was like are you worse than a person? Yeah? Or
14:29
Speaker 2: are you slightly better than a person? And we got
14:33
Speaker 2: slightly better than a person, which was a breakthrough in
14:36
Speaker 2: my view.
14:36
Speaker 1: Right, and so yes, so that paper you published in
14:39
Speaker 1: Science and you started Ozmo kind of around the same time, Right,
14:43
Speaker 1: you started that study at at Google, is that right?
14:46
Speaker 1: And then by the time it was published you had
14:47
Speaker 1: spun Osmo out of Google, right, that's right. So you
14:51
Speaker 1: have this map, you have this model that can basically
14:54
Speaker 1: given a molecule, predict pretty well what the average person
14:58
Speaker 1: thinks that molecule smell like. But there's still a second problem, right,
15:03
Speaker 1: which is, in the world, in the wild, you don't
15:07
Speaker 1: know what molecules are in the air. You don't you
15:09
Speaker 1: know what molecules somebody's smelling. And so for that second problem,
15:13
Speaker 1: you need to try and build some kind of automated
15:15
Speaker 1: system for figuring out what molecules are in the air
15:18
Speaker 1: at a given that's correct.
15:20
Speaker 2: Getting to one molecule structure is actually not trivial. So
15:26
Speaker 2: to go from a physical thing and know all the
15:28
Speaker 2: molecular structures like not a solveable. So there's a lot
15:31
Speaker 2: of ways to do that. There's a lot of chemical
15:33
Speaker 2: sensors out there, none of them will just tell you
15:37
Speaker 2: the formula, right, So that's hard, really hard.
15:41
Speaker 1: So there's like a chemistry problem of like isolating the
15:44
Speaker 1: molecule basically and deriving the chemical formula.
15:49
Speaker 2: Exactly taking a real smell and it's composed of a
15:51
Speaker 2: bunch of different molecules with different structures, and there's different amounts,
15:55
Speaker 2: there's ratios. You got to get that recipe out of
15:57
Speaker 2: the air. So that's on. That's hard. That was unsolved
16:02
Speaker 2: at the time to do it in an automated way.
16:04
Speaker 2: And by the way, if we're following this story chronologically,
16:08
Speaker 2: we hadn't done this yet, agah, but we knew we
16:10
Speaker 2: had to do that, right, So we knew that, Okay,
16:12
Speaker 2: if we wanted to actually digitize the world of scent
16:15
Speaker 2: and have a record of what the world smelled like
16:18
Speaker 2: and maybe even replay it, we're going to have to
16:20
Speaker 2: do this. We needed to automate that and have it
16:22
Speaker 2: be automatic, and that's what we did.
16:24
Speaker 1: So basically, you can put any smell into the machine
16:28
Speaker 1: and it'll tell you what it's made of. At this point,
16:31
Speaker 1: oh yeah, so you're setting out to start OSMO, Like,
16:33
Speaker 1: what are you thinking of in terms of the set
16:36
Speaker 1: of potential commercial applications.
16:39
Speaker 2: So we really had we had three in mind, and
16:45
Speaker 2: there's still very much present in mind the focuses has
16:49
Speaker 2: become a lot crisper though in terms of what we're
16:52
Speaker 2: concentrating on. We know the fragrance industry is huge and
16:57
Speaker 2: very profitable, and it's also something I personally love. That's
17:02
Speaker 2: a thing we want to automate and understand. And then
17:06
Speaker 2: we know that dogs can detect things, right, and so
17:09
Speaker 2: we know dogs can detect harmful substances like drugs or bombs,
17:13
Speaker 2: or things that just shouldn't be there, like produce where
17:15
Speaker 2: it shouldn't be being shipped. And then we also know that
17:19
Speaker 2: dogs and even in some cases people can detect health
17:22
Speaker 2: or disease states. Right. We know that missus Milner, a
17:25
Speaker 2: nurse in the UK, was able to smell Parkinson's disease
17:31
Speaker 2: and she's since been able to teach that skill to
17:34
Speaker 2: other people, which is really amazing. And then we figured
17:36
Speaker 2: out all the chemistry of what's actually being smelled. We
17:39
Speaker 2: know that there's many many instances where there is a
17:43
Speaker 2: scent signature to a disease or to a wellness or
17:45
Speaker 2: to a health state that hasn't yet been fully figured out, right,
17:50
Speaker 2: but we know that they exist. Those are the three, right,
17:53
Speaker 2: So fragrance industry really security and supply chain and health
17:57
Speaker 2: and wellness, and I view them in that order because
18:03
Speaker 2: that's like the order in which I think we can
18:07
Speaker 2: be useful to the world. Right, Designing fragrances is something
18:11
Speaker 2: that's much more attainable technically, and frankly, it's just a great,
18:14
Speaker 2: much faster sales cycle to be business to be in
18:17
Speaker 2: than ultimately diagnostics, which are so hard. Right, I mean,
18:20
Speaker 2: it is my north star. It's like where I want
18:22
Speaker 2: to take the company. But I also have no illusions
18:25
Speaker 2: about how hard that is, and I just I've seen
18:27
Speaker 2: all the failures of the companies that have attempted it,
18:30
Speaker 2: and I think I've learned from what hasn't worked, and
18:33
Speaker 2: so I'm incorporating those learnings into how I want to
18:35
Speaker 2: build the company, which is, build a great business in fragrance,
18:39
Speaker 2: build beautiful fragrances for the world, and then strike out
18:42
Speaker 2: from that position of strength and to even more ambitious frontiers.
18:49
Speaker 2: We'll be back in just a minute.
19:00
Speaker 1: When I first heard about the work Alex was doing
19:02
Speaker 1: at OSMO, I understood why it would be useful for sensing. Basically,
19:07
Speaker 1: you might be able to build automate sniffing machines that
19:10
Speaker 1: could say detect cancer in a person or detect a
19:14
Speaker 1: bomb in a suitcase. But I couldn't figure out truly
19:18
Speaker 1: what the business case was for perfume. And in fact
19:22
Speaker 1: Osmo has recently launched a perfume business. It's called Generation.
19:27
Speaker 1: So I asked Alex, why is using AI and fancy machines?
19:32
Speaker 1: Why is that better than just designing perfume in the
19:35
Speaker 1: traditional way.
19:36
Speaker 2: We can go from the first kind of client demand. So, hey,
19:41
Speaker 2: I want to create a fragrance, and here's who my
19:42
Speaker 2: brand is, here's what I want to do. So just
19:45
Speaker 2: that description to a starting place of a fragrance in
19:49
Speaker 2: a minute or two.
19:50
Speaker 1: What happens at a traditional perfumer when somebody comes in
19:53
Speaker 1: with that request.
19:54
Speaker 2: Well, so let's say you're an emerging Let's say you're
19:56
Speaker 2: an emerging brand, right, So you're starting out or you
19:59
Speaker 2: have your first product and you want to add a
20:00
Speaker 2: second one. But you're small, right, You're not making a
20:03
Speaker 2: billion dollars in revenue. You're making less than that. So
20:06
Speaker 2: if you want to make a new custom fragrance, good luck. Right,
20:09
Speaker 2: you're not going to be able to get the attention
20:12
Speaker 2: of the big fragrance houses because they want a service
20:15
Speaker 2: business that's like millions and millions of dollars and you're
20:20
Speaker 2: not big enough yet. So if you want a great
20:22
Speaker 2: custom fragrance that your consumers are going to love, and
20:25
Speaker 2: you want to do it quickly so you're responding to trends,
20:29
Speaker 2: you aren't going to be able to get it done.
20:33
Speaker 2: So you have to make compromises, right, So if you
20:35
Speaker 2: want to move fast, you're gonna have to use a
20:37
Speaker 2: regurgitated fragrance. It's also called a library fragrance, which means
20:40
Speaker 2: somebody else in the market has your spell.
20:42
Speaker 1: I'm imagining that people who sell it call it a
20:44
Speaker 1: library fragrance rather than a regurgitated library fragrance.
20:47
Speaker 2: They do, they don't say regurgitated, but that's what it
20:49
Speaker 2: effectively is, right fair.
20:51
Speaker 1: Regurgitated does have a particular old factory connotation, so it's
20:55
Speaker 1: a level.
20:57
Speaker 2: It's vic or all sticks in the mind.
20:59
Speaker 1: What like, And I'm not I just genuinely don't understand, Like,
21:03
Speaker 1: why can't somebody just have a company with a bunch
21:05
Speaker 1: like who knows the molecules? You know, who knows what
21:08
Speaker 1: the five thousand milecules in the book smell like? Because
21:11
Speaker 1: they've got the book and they can just use the
21:13
Speaker 1: book and be like, oh, you want this, let's try that.
21:15
Speaker 2: Do you know what I mean? Like, I'm not trying
21:17
Speaker 2: to be difficult, but I.
21:18
Speaker 1: Genuinely don't understand why you need the technology to do that.
21:24
Speaker 2: Yeah, I genuinely didn't understand this either. And there's there's
21:29
Speaker 2: a class of professional called a perfumer, and their job
21:33
Speaker 2: is to do what you're describing, which is, Hey, I
21:35
Speaker 2: know all these ingredients and I'm going to make them
21:37
Speaker 2: in order to create your fragrance. So they typically there's
21:41
Speaker 2: no perfumer that knows five thousand ingredients, but the best
21:44
Speaker 2: perfumers know a thousand or two thousand ingredients. Most perfumers
21:48
Speaker 2: work with two hundred, one hundred, two hundred ingredients. So
21:52
Speaker 2: already there, like where we're there is very few people
21:55
Speaker 2: in the world that can do what you're saying. Yeah,
21:57
Speaker 2: they can do, and then how what are they going
22:01
Speaker 2: to work on? Right? So it might take them weeks
22:04
Speaker 2: or months to create a fragrance. They're working on a
22:07
Speaker 2: few at a time. Why would they work on an
22:10
Speaker 2: emerging brands fragrance when they can go work on a
22:11
Speaker 2: much larger account. So there's just very very limited number
22:15
Speaker 2: of people who can engage in the fragrance creation process,
22:18
Speaker 2: because it is difficult. It's not so much identifying, Hey,
22:22
Speaker 2: all these molecules smell this particular way, and therefore I
22:25
Speaker 2: should be able to mix them, Like what ratios do
22:27
Speaker 2: you mix them in? Like? What are the rules? Right?
22:28
Speaker 2: And now you're actually getting into designing a system which
22:31
Speaker 2: understands sent well enough to create new fragrance formulas, is
22:34
Speaker 2: starting places, and then of course a perfumer finished system.
22:37
Speaker 2: But you're right, it's like, oh, why shouldn't that exist?
22:40
Speaker 2: And then when you actually start to peel back the
22:42
Speaker 2: layers one by one, you realize, oh, you actually have
22:44
Speaker 2: to build what we built. It's actually in order to
22:46
Speaker 2: answer that question.
22:46
Speaker 1: So presumably now your model cannot only predict what one
22:50
Speaker 1: molecule is going to smell like, but it can predict
22:52
Speaker 1: the combination of molecules. I mean, is it predicting? Does
22:55
Speaker 1: it know concentration? Like does it know oh yeah, yeah,
23:01
Speaker 1: how good is it?
23:02
Speaker 2: I mean you have a perfumer on staff? Why? Well,
23:08
Speaker 2: I think the goal of tools is to have them
23:13
Speaker 2: in the hands of creatives. And there's many steps to perfumery,
23:17
Speaker 2: but I think there's three that are relevant for what
23:19
Speaker 2: we're talking about. The first is a perfumer when they're
23:22
Speaker 2: when they're starting on a project, they have to have
23:24
Speaker 2: a starting place, they have a starting formula, and then
23:26
Speaker 2: they do their creative work step two to evolve that
23:30
Speaker 2: formula to exactly what the customer wants, to a creative
23:34
Speaker 2: expression that the lights the consumer as well. That's the
23:37
Speaker 2: funnest part. Perfumers love that that is actual creation in
23:40
Speaker 2: the creative part. Number three is then it has to
23:43
Speaker 2: be the right price. It has to be compliant with
23:45
Speaker 2: with a regulatory compliance. There cannot be allergens, all that
23:48
Speaker 2: stuff that's more like sound engineering than it is composition
23:51
Speaker 2: or being a rock star. Steps one, the starting place,
23:54
Speaker 2: step three, all the regulatory requirements. That's where we spend
23:58
Speaker 2: the most energy in building these tools. And then a
24:01
Speaker 2: perfumer is the person that is taking the formulas from
24:06
Speaker 2: starting place to creative endpoint and then handing it off
24:09
Speaker 2: for like regulatory finishing. And they're just way more effective
24:12
Speaker 2: with these.
24:13
Speaker 1: Tools, at least for now, right Like, that's the way
24:17
Speaker 1: I feel using an LLM, like I feel like I
24:19
Speaker 1: have a window and me Plus the LLM is better
24:22
Speaker 1: than the LLM alone and we haven't. That window hasn't
24:25
Speaker 1: closed yet, but I'm not optimistic about my long term prospects.
24:29
Speaker 2: We'll see though, I mean, but like listen, I honest
24:33
Speaker 2: belief here, like the tools will get better, but the
24:37
Speaker 2: drive to create will never go away. And I think
24:40
Speaker 2: people will always want to know about the person behind
24:44
Speaker 2: the creation in a way, and it's not uniforms. So
24:47
Speaker 2: I don't think people want to know the perfumer behind
24:49
Speaker 2: the hand soap in the gas station. They just don't,
24:52
Speaker 2: right it. But there I think will always be room
24:58
Speaker 2: for craft and creative use of tools, and the profession
25:01
Speaker 2: that uses those tools might change radically, in the industry
25:05
Speaker 2: in which those tools are used might change radically, but
25:08
Speaker 2: the tools will always be wielded by people, but the
25:11
Speaker 2: work that's being done might be unrecognizable. So you know,
25:15
Speaker 2: we'll see how the world evolves. But like I just
25:17
Speaker 2: like AI is like an engine, it's just a technologist,
25:21
Speaker 2: just a tool.
25:23
Speaker 1: So what's the what's the business model? Just briefly for generation,
25:28
Speaker 1: like how what you know?
25:29
Speaker 2: What's the model? The business model really simply is we
25:34
Speaker 2: all design the fragrance for you and then you'll buy
25:36
Speaker 2: that fragrance to put in your products or will even
25:38
Speaker 2: actually create the full finished product. We'll put in a
25:40
Speaker 2: bottle for you if you if you want. We are
25:43
Speaker 2: behind the scenes. We're an engine supporting brands. We're not
25:46
Speaker 2: a brand ourselves, and we're here to make beautiful fragrance
25:50
Speaker 2: products for for brands. So what's the frontier like you have?
25:56
Speaker 1: You know, on the business side, the generation is kind
25:59
Speaker 1: of the central thing you're working on now, but on
26:02
Speaker 1: the more on the on the research side, like what
26:04
Speaker 1: are you trying to figure out?
26:05
Speaker 2: Now? What are you working on now? So there's there's
26:11
Speaker 2: our starting place, which is why does this molecule smell
26:13
Speaker 2: the way that it does? And we can never stop
26:16
Speaker 2: getting better at that. Then there's the next question of
26:18
Speaker 2: why does this mixture of molecule smell the way that
26:21
Speaker 2: it does? And we can never stop getting better at that.
26:23
Speaker 2: And then there's do you like it? Which is maybe
26:27
Speaker 2: the most important question from a business perspective, or who
26:30
Speaker 2: likes it? And in what context? Yeah, exactly exactly, which
26:34
Speaker 2: is it's not just the formula as the input to
26:37
Speaker 2: this model, but there's also who you are, what are
26:39
Speaker 2: your experiences, where are you from, what are the other
26:41
Speaker 2: things in your life that you've got.
26:42
Speaker 1: That actually goes back to your Netflix collaborative. It gets
26:45
Speaker 1: me like if I watched Succession and the Sopranos and
26:49
Speaker 1: I'm fifty and.
26:51
Speaker 2: Then what's the cologne for me? Yeah? Exactly. And so
26:56
Speaker 2: I was very fortunate to be able to start this
26:59
Speaker 2: company with a guy work with at Twitter name. His
27:02
Speaker 2: name is Rich Witcombe. He's our chief technology officer. His
27:05
Speaker 2: whole professional life has been recommended system. So he was
27:09
Speaker 2: a lead on Spotify's song recommenders system US. So if
27:13
Speaker 2: you like your wrapped playlist or recommended playlist, like, that's
27:16
Speaker 2: his code. And then he also worked on self driving
27:19
Speaker 2: cars at Nvidia. But he's been in this world of like,
27:22
Speaker 2: hey you like these things, what about this thing? Or
27:25
Speaker 2: here's the inputs that the system is getting. What do
27:27
Speaker 2: I do nowt so really really deep into that world.
27:29
Speaker 2: Then we're kind of bringing that spirit, that mindset to
27:34
Speaker 2: sent into fragrance.
27:35
Speaker 1: And then what about beyond you know, for the parts
27:40
Speaker 1: of your work that are the next steps that you
27:43
Speaker 1: alluded to farther in the distance, the essentially sensing right,
27:46
Speaker 1: sensing for security, sensing for health, Like what work are
27:51
Speaker 1: you doing now toward that end?
27:53
Speaker 2: Yeah? So we're incubating this right now. So I'll tell
27:56
Speaker 2: you two things. So one is, we have a partner,
27:59
Speaker 2: We've deployed sensors out in the field. We're detecting inauthentic
28:03
Speaker 2: or counterfeit goods. It's working. What's really the second thing
28:06
Speaker 2: I'll say is it's we've learned something really interesting, which
28:09
Speaker 2: is the molecules that smell really good and fruits and
28:13
Speaker 2: flowers and vegetables that we have to understand to create
28:16
Speaker 2: fragrance are the same molecules in counter for luxury goods
28:20
Speaker 2: and the same molecules in our scent. And by getting
28:24
Speaker 2: really good at understanding and designing fragrance in one domain
28:29
Speaker 2: in the fragrance industry, we're actually strengthening this platform that
28:32
Speaker 2: we're building to get really good at the next frontiers
28:35
Speaker 2: of security detection. And then ultimately, what we care about
28:39
Speaker 2: is healthy. So that's what really surprised us as I
28:43
Speaker 2: thought that by working in fragrance, we're making a trade off,
28:47
Speaker 2: which is we're here to build a great business to
28:50
Speaker 2: make ourselves resilient so that we can work on the
28:52
Speaker 2: much longer haul problems. But in reality, we're making progress
28:56
Speaker 2: on those problems by teaching our platform about what the
28:59
Speaker 2: world smells like. And it's all one it's just scent,
29:02
Speaker 2: it's just molecules in the air. And so the more
29:04
Speaker 2: we learn about really any piece of what the world
29:06
Speaker 2: smells like, the better we get at all of it.
29:09
Speaker 2: I think I'll tell you what I think the big
29:11
Speaker 2: technical frontier is is predicting emotion. Ah, that's interesting. Uh huh.
29:18
Speaker 2: So when you smell something, you obviously perceive something like
29:21
Speaker 2: the first thought or first perception is whatever, fresh cut
29:24
Speaker 2: grass or grapefruit. But then there's another thing that happens
29:28
Speaker 2: almost at the same time, which is I remember or
29:32
Speaker 2: I feel a particular thing. And predicting that is something
29:37
Speaker 2: I don't think anybody's really figured out. But is a
29:41
Speaker 2: beautiful frontier. Well, how do you get the data? You
29:44
Speaker 2: got to ask a lot of people how they feel,
29:46
Speaker 2: what they smell a lot of things, and they have
29:48
Speaker 2: to be able to articulate it. Right.
29:49
Speaker 1: Part of the thing with scent is it's so primordial that, like,
29:54
Speaker 1: you might not even be able to say how you feel,
29:57
Speaker 1: so you need in the computer interface.
30:00
Speaker 2: You might you might, But turns out we have voices
30:02
Speaker 2: and faces that are effectively BCIs there's a lot of
30:06
Speaker 2: information that leaks out of us all the time. And
30:09
Speaker 2: that was what my PhD was in, is how do
30:11
Speaker 2: you interpret body language in a way that makes sense?
30:15
Speaker 2: And by the way, the body language I worked on
30:17
Speaker 2: most closely was body language driven by odors, right, things
30:20
Speaker 2: that make I studied this in animals, but makes animals
30:23
Speaker 2: happy or sad or afraid or calm, And you can
30:27
Speaker 2: read that out. I mean, our behaviors are meant to
30:30
Speaker 2: communicate to other animals. Right, We're very social, we're social species.
30:34
Speaker 2: So I think there's more fundamentals that we have to
30:36
Speaker 2: figure out. But this is I think there's some really
30:40
Speaker 2: fundamental stuff that's still unknown here.
30:43
Speaker 1: I heard you say in another interview that you worry
30:45
Speaker 1: sometimes that you'll hit some barrier in nature to your work,
30:51
Speaker 1: and you said it in passing. But I was very
30:53
Speaker 1: curious about that.
30:54
Speaker 2: What does that mean? I always think about that, which
30:57
Speaker 2: is like, what day will it be when mother nature
31:01
Speaker 2: says you can't figure the next hard thing out? And
31:06
Speaker 2: I just look at this from the history of science.
31:10
Speaker 2: You know, how if somebody cared about how the planets
31:12
Speaker 2: were moving in twelve hundred, well good luck, Like you
31:16
Speaker 2: don't have the right telescopes, you don't have tycho brahe.
31:19
Speaker 2: There's a bunch of stuff you're gonna need, right, And
31:22
Speaker 2: so in a way, it's like mother nature and what
31:24
Speaker 2: our society and species knows conspiring together that basically says
31:28
Speaker 2: progress will have to wait. And so I think about that.
31:31
Speaker 2: I worry about that all the time. And so my
31:34
Speaker 2: mental framework that keeps me super humble is like, I'm
31:38
Speaker 2: just thankful for all the progress we've been able to make.
31:40
Speaker 2: That the tools were around right. So I didn't invent
31:44
Speaker 2: graph neural networks. I didn't even invent the data sets
31:47
Speaker 2: like we are piecing together and curating, cobbling together. All
31:50
Speaker 2: these were standing on the shoulders of so many people
31:53
Speaker 2: and it's just always been the case, and I don't know,
31:57
Speaker 2: it just it makes Maybe this is too philosophical, but
32:03
Speaker 2: for me, when I've been up close and personal with
32:06
Speaker 2: scientific progress, either that I've had a part in or
32:08
Speaker 2: I've observed other people do, it all feels so tenuous.
32:12
Speaker 2: It feels so lucky because once you really dig into
32:14
Speaker 2: the details, you realize, oh my gosh, they had to
32:17
Speaker 2: be right there at that time and have known about
32:19
Speaker 2: that thing.
32:20
Speaker 1: It's amazing that anything happens, whether you think of how
32:23
Speaker 1: confut is everything.
32:24
Speaker 2: The amazing that anything happens, and you know, when you
32:27
Speaker 2: really dig in, you're like, wow, how does anything good
32:30
Speaker 2: happen at all? But nonetheless you persist. And also I
32:34
Speaker 2: think you can create the conditions where it's more likely
32:36
Speaker 2: than not to happen. And so that's what OZMO is,
32:39
Speaker 2: and that's why OSMO Birth Generation is, like, let's create
32:42
Speaker 2: an environment where we're much more likely than not to
32:45
Speaker 2: make both the scientific progress we need to make, but
32:47
Speaker 2: also like really help and change the fragrance industry, which,
32:52
Speaker 2: by the way, will teach us the things we need
32:55
Speaker 2: to know to get to the next thing. So I
32:56
Speaker 2: think there's so much beauty to create in the fragrance
32:59
Speaker 2: industry that I'm going to just enjoy the heck out
33:01
Speaker 2: of it and do it for the rest of my life.
33:04
Speaker 2: But I think it's going to teach us things that
33:05
Speaker 2: will allow us to do even more audacious work in
33:09
Speaker 2: the future.
33:13
Speaker 1: We'll be back in a minute with the lightning round.
33:18
Speaker 1: M let's finish with the lightning round. I'm gonna ask you.
33:27
Speaker 2: A bunch of questions.
33:28
Speaker 1: Now, what seemingly pleasant scent do you never want to
33:32
Speaker 1: smell again.
33:34
Speaker 2: Seemingly pleasant scent that I never want to smell again.
33:37
Speaker 2: Artificial cherry. It was the cough syrup that I was
33:40
Speaker 2: forced to drink as a kid, and I'm super sensitive
33:42
Speaker 2: to it. The molecules ethyl moltole do not like, are
33:46
Speaker 2: you wearing fragrance right now? And if so, what is it?
33:49
Speaker 2: I am not. I stopped wearing as soon as I
33:50
Speaker 2: started the company because I needed to smell, of course,
33:55
Speaker 2: but like, what's your what's your? Well, give me give
33:58
Speaker 2: me a pick. Name some fragrance that's that you love
34:01
Speaker 2: for some reason. So I really like this is kind
34:05
Speaker 2: of a basic choice from folks inside the industry. I
34:09
Speaker 2: love Terrator Maz. It's like the RMEZ flagship men's fragrance.
34:13
Speaker 2: It's by a perfumer, Jean Claude Eleena. I really love
34:15
Speaker 2: his work.
34:15
Speaker 1: Basic Is that like basic in the way of saying
34:18
Speaker 1: it's like if I asked you for a watch and
34:20
Speaker 1: you said a Rolex Submarine or something.
34:21
Speaker 2: It's just like exactly or saying, like what pop music
34:24
Speaker 2: do you like? You said Taylor Swift. People like it
34:26
Speaker 2: because it's great, Huh, Taylor Swift is great? A Rolex
34:30
Speaker 2: watch is a great watch. Terrtormnz is a great fragrance,
34:33
Speaker 2: but it's very popular. What is it about it that
34:35
Speaker 2: you love? I love it's minimalism and I just happen
34:38
Speaker 2: to like the notes, right, So it's really heavy on
34:40
Speaker 2: a molecule. I like iso be super. I think it's
34:42
Speaker 2: a great highlight of that ingredient and it just wears
34:45
Speaker 2: really well on my skin. So that was what I
34:47
Speaker 2: used to wear almost every day before I stopped. What's
34:51
Speaker 2: your second favorite sense? My second favorite sense is probably
34:57
Speaker 2: gonna be It's a hard between vision and hearing because
35:00
Speaker 2: I love music, but I like looking at stuff too,
35:03
Speaker 2: Like the World World of Beautiful are more expensive. Perfumes
35:09
Speaker 2: actually better sometimes, right, So I think there's just like
35:13
Speaker 2: anything like bicycles or art, as you start to pay more,
35:17
Speaker 2: everything gets better.
35:18
Speaker 1: And then at Plato's right, what's the worst thing you
35:20
Speaker 1: ever smelled?
35:22
Speaker 2: I have a memory. I picked a mushroom that I
35:26
Speaker 2: thought looked cool and wanted to show it to my
35:28
Speaker 2: dad when I was young, and I forgot about it
35:31
Speaker 2: and it was just turned completely gross.
35:35
Speaker 1: I had a version of that, of bringing shells home
35:37
Speaker 1: from the beach that were alive.
35:39
Speaker 2: It turned out I found out when they were dead.
35:41
Speaker 2: It's like great intentions, but didn't really have wherewithal to
35:45
Speaker 2: thick that through or understand the consequences.
35:54
Speaker 1: Alex Wiltscow is the co founder and CEO of OSMO.
35:58
Speaker 1: Today's show was produced by Gabriel Hunter Chang. It was
36:01
Speaker 1: edited by Lyddya jen Kott and engineered by Sarah Bruguer.
36:05
Speaker 1: You can email us at problem at Pushkin dot FM.
36:09
Speaker 1: I'm Jacob Goldstein, and we'll be back next week with
36:11
Speaker 1: another episode of What's Your Problem.