Osmo is giving computers a sense of smell — and already selling AI-designed fragrances at Target

Speaker 1: Have a good one. Thanks, guys. We'll talk to you soon. Up next, we have Alex from Osmo building olfactory intelligence. We've talked about this before. Can AI smell? That's our current benchmark for AGI. We say if if, you know, we talk about white collar work, we see sommeliers as white collar workers. Unless you can smell, it's not AGI and artificial intelligence falling short. But it's your first time on the show. I would love an introduction on yourself and the company because I'm fascinated by this topic.

Speaker 7: Let's talk about it. Your Somalia comment and also what you talked about with Max Koenig has been on my mind.

Speaker 1: Amazing. My name

Speaker 7: is Alex Wiltschko. I'm founder and CEO of Osmo. We're giving computers a sense of smell. I've been Hi, Mike. Working on this problem for twenty exactly twenty years or so.

Speaker 2: Twenty years.

Speaker 7: First as an academic. Tonight's success. I did my PhD in olfactory neuroscience at Harvard and trained under Bob Datto who trained with Richard Axel, who got the Nobel Prize for discovering the receptors of smell. Wow. And my AI mentor trained with Jeff Hinton who got the Nobel Prize for deep learning. Yeah. And I'm the one weirdo that's like

Speaker 1: waiting for you to join for the entire history of the show.

Speaker 2: There's been great prophecies of your arrival

Speaker 1: for hundreds of years. I'm so excited.

Speaker 7: Okay. I'm so pumped to be here.

Speaker 1: So so should we start with maybe like a olfactory science one zero one? Can you set the ground on like how does smell even work? What are the important sort of like building blocks that we should know and then we can build up to the next generation and how AI is being applied?

Speaker 7: It it a 100%. So the chemical slice of reality, all the stuff that's data in the air

Speaker 2: Mhmm.

Speaker 7: We can detect that. Our sense of smell is literally our brain leaving our skull. So when you smell a molecule, whether it's a tree or it's a meal or a drink, like the the physical pieces of that thing enter into your nose and touch a piece of tissue about the size of a postage stamp. Mhmm. And that's your brain. Right? So like you're in physical communion with that thing. Yeah. That information gets turned into neural data which actually skips all of the normal way stations for the other senses Yeah. And goes right to your centers of memory, the hippocampus, and emotion, the amygdala. So our sense of smell is very primal in that regard. So it's it's like it's the reason why when you smell something, you get dragged into a memory and you cannot stop it. Yeah. You're just like back in high school or you're back as a kid. It's because it's we're physically wired for that.

Speaker 1: So that's real. I've always heard that. I've always heard that phrase smell is the sense that's most tied to memory, but I didn't know if it was just something you saw in like a t shirt or something.

Speaker 7: No. Literally neuroanatomically tricky.

Speaker 1: Target wall art. Yeah. Yeah. It feels like Target wall art. I don't know. It's just one of those things that you repeat

Speaker 2: They say

Speaker 1: a spell

Speaker 2: pop sires. A thousand words.

Speaker 1: Okay. So that sounds like something that's extremely hard to reverse engineer. Do we have do we have sensors? Because you know LLMs it was so obvious that we had text that was already encoded into data into ones and zeros and so transforming that and encoding it, I mean it was an incredible breakthrough but it felt like the the text was the data was already in the computer and I feel like that's not true for olfactory data, for smell data, but how are we do we need to digitize this before we do anything with it? How does digitization of smell work?

Speaker 7: Yeah. Great question. So I was very fortunate to have those guys as my colleagues. I actually spun Osmo out of Google Brain. Oh. And so I was there when all that stuff got invented and I ran the digital of action team at Google Brain for about six years before we decided to make it a company through Lux and through GV. And you have it exactly right. Like, the Internet had been been accumulating for a while, so we had all this text data. Yeah. So we could basically slurp that down Yeah. And start building models. Got it. We have chemical sensors. They're called mass spectrometers. There's other kinds of chemical sensors, mock sensors. There's like a dozen. The history of sensors that can turn chemistry into data is about a 100 years old. Sure. Maybe more. I mean, a lot of it was pushed forward in the Manhattan Project actually. Wow. But what we've been missing is a map. Okay. Right? So for sound, low to high frequency is a map, which lets us build m p three and speakers and microphones and Spotify, etcetera. And for color, RGB is a three-dimensional map of color. Yeah. And that lets us build CMOS, CCD, you know, cameras, etcetera. We haven't had the map for snow. And that's not crazy because there's three channels of color information in our eye, but we know there's over 300 channels of information in our nose. Wow. So in a way, we actually did need to wait for artificial intelligence to mature in order to have the ability to extract a 300 dimensional map from data. And that's exactly what we did starting with our first work at Google Brain. Yeah. So you gotta go get a crap ton of information, right? A bunch of molecules, what they smell like. We've since collected the largest AI dataset for scent in the world. That's what drives olfactory intelligence. We have 5,000,000 sniffs digitized, over a quarter million physical samples created. We've digitized about 6,000,000,000 fragrance molecules. So all this is like inside of the company because there's literally nothing on the internet. The fragrance industry has done a phenomenal job keeping everything secret. So we built it all ourselves.

Speaker 2: Remarkable. Jurewicz? How are you gonna make money

Speaker 4: on this?

Speaker 1: It's a good question.

Speaker 7: So if if if you go actually, let's talk. How can we make money on this?

Speaker 1: Yeah.

Speaker 7: So does TBPN have this scent?

Speaker 1: Yes. It's terrible. Terrible. There's rubber smell in the studio. There's in the studio

Speaker 2: there's like thousands of cords Cables. And cables. The cables And we we do a good job hiding them. But we have so much gear going everywhere.

Speaker 1: It's lot of rubber, a lot of to

Speaker 2: get these Racetracks. Racetracks, they're called to cover all the cables. And it turns out these things smell terrible.

Speaker 7: A lot. Those off gas. Yeah.

Speaker 1: They're off gas.

Speaker 2: So we wanted to give our viewers

Speaker 1: Yes.

Speaker 2: The the full

Speaker 1: experience. I think we actually do not. It would

Speaker 2: be like a can that sits on Yeah. Their desk Yeah. Yeah. Aerosol and it would spray a rubber smell into the room Yeah. So they could experience what we experienced.

Speaker 7: We capture it but I think we should fix it.

Speaker 1: So Yes.

Speaker 7: Really concretely, we raised our series b. We put an additional 70,000,000 in the bank Kinda. With two sigma leading luxe. Love that it got the gong. Congratulations. That that was to underwrite building a fragrance factory. Okay. So we have a robot that's the size of a school bus that makes a new fragrance every hundred seconds. And what we do is we design and manufacture fragrances for brands.

Speaker 1: Oh, yeah. That makes sense.

Speaker 7: And so we we use olfactory intelligence to So design super fast, data driven, basically, you know, perfect fit for the brand and for and for the consumer of that brand. And then we actually physically make it and what leaves our factory is a steel drum that fragrance oil and we build them for it. Yeah. We also will do end to end. So like if you want to actually make a physical bottle, we'll actually put the fragrance in the bottle for you Mhmm. So the full product comes out. So if you guys want to launch a TBPN Cologne or something like that, we could design it for you. I mean, like, if you tell me

Speaker 2: the prompt right now smell like burnt rubber.

Speaker 7: No. No. No.

Speaker 1: We're not doing burnt rubber.

Speaker 7: Burnt rubber. There's some it smells like disagreement.

Speaker 1: No. It needs to smell like like like old $20 bills. Okay. From the 19 That is insane. Mahogany. The official wood of business. Need to smell like mahogany That's with alright. With old $20 bills, the smell of money.

Speaker 7: That's Okay. Cool. We've done this we've done the smell of money one which we demoed actually on the New York Stock Exchange floor which

Speaker 1: is That's pretty amazing. That's amazing. But

Speaker 7: no, I'll I'll send you I'll we'll make something. I'll send it Talk to

Speaker 1: about sensor miniaturization. Phone has three cameras and no smelling sensor. Can we swap one of these out? Like I when you say mass spec, imagine like a device the size of a living room. I imagine that they are getting dishwasher.

Speaker 7: The size of the dishwasher?

Speaker 1: Is there a path to actually shrinking that down to something that's more Sure. Portable?

Speaker 7: So, yeah. In the same there's like many kinds of cameras, right?

Speaker 1: So the

Speaker 7: one in the Hubble telescope not getting smaller. So if you need resolution, it's got it's gonna be big. But you can make trade offs. And like when the thing that's reading the data instead of it being a person, it's an algorithm, you can actually make really intelligent trade offs, which is what we've done. So we actually have a sensor right now. It's the size of two shoe boxes. Okay. And I kind of use that metric aptly because we've actually used it to smell fake shoes. Oh. So if you're buying a pair of, like, $500 Air Jordans, the reels smell different from the fakes. We can actually pick that up. That's crazy. Real from the we can

Speaker 1: Yeah.

Speaker 7: Yep. The the the counterfeiters use cheaper glues, turns out. Interesting. And the the other thing that's interesting is we can actually tell the factory of origin of the shoe 93% of the time. Mhmm. So there it smells a fingerprint. So we're already miniaturizing these devices. Look, the path to get from two shoe boxes to one shoe box is pretty clear. Yeah. We're working on that. To go to something that's like the size of the AirPods case, there's gonna be some like, part four engineering required. To have it be a component that fits in your phone, there's some breakthroughs like, can't quite see through the fog yet, but there's nothing like, look, our our noses do it. So there's nothing that mother nature is saying is like impossible. Yeah. But we just got a lot of work to do.

Speaker 1: Yeah. That makes a lot of sense. What about taste? How closely is taste linked? Talk me through the sommelier example.

Speaker 7: Yeah. So flavor is everything that happens in your mouth, you know, that's that's a sensory experience of food. Taste is not as like 10% of that. It's like what happens on your tongue.

Speaker 3: Like Yeah.

Speaker 7: You ever eat a jelly bean and like plug your nose?

Speaker 1: Yep.

Speaker 7: And you you just actually can detect very little of what's going on there? Yeah. It's because 90% of what you experience is actually called retro nasal olfaction, where when you're biting or biting on something, there's a chimney effect in the kind of the almost the steam of what you're eating goes back through your nose and you

Speaker 1: smell it. Oh. Excuse me.

Speaker 7: And then there's also the texture and everything in your mouth. So we've done tests and our OI models, this is from a while ago, we haven't revisited it. We're really focused on fragrance right now. But our OI models actually work on flavor surprisingly well. And so the whole world of flavor is there for us with Meredi, but we're we're really focused on on this particular business. Yeah.

Speaker 1: I've seen a couple of these sort of I don't wanna call them niche, but, like, vertical AI projects that are not fully generalizable. There's a DNA model also from Google or D Mind. And it feels like they're starting to get on scaling curves, on scaling laws. Are you at a point where you feel like, oh, if I 10 x the computer, 100 x the compute that goes into some I believe

Speaker 2: Alex is ready for a one gigawatt data center.

Speaker 1: He can be trusted with

Speaker 7: that.

Speaker 2: I I I

Speaker 1: would trust you. But but Appreciate it. How universal do you think scaling laws are? Is there a scaling law here? Is it data based? Is it compute based? Both? How are you thinking about it?

Speaker 7: The better lesson's real. The better lesson's super real. I always think about technology as s curves. Right? And like, what's driving you up that s curve and then how can you hop on the next one? Our current s curve is data, which is why we're maniacally focused on like generating a ton of data. Like, have a giant fragrance robot that spits out a ton of fragrances. We have mass specs running twenty four seven. We have sensory panels, both domestically. Have a building of people that just smell all day abroad. Wow. And we ship them crates of stuff to smell. How we get to 5,000,000 sniffs. Right?

Speaker 1: Sure.

Speaker 7: So data, data, data. The the the size of the models is not the limiting factor right now.

Speaker 1: Yeah.

Speaker 7: And it will be at some point, and then switch to the other s curve.

Speaker 1: Yeah. Yeah. Because you don't just have, like, the open Internet to scrape because there's not an existing data set. Makes sense.

Speaker 7: Totally. Double edged sword. Right? So we've had to make it all. Right? Which is really hard. But also, nobody else has it because we had to make it all and had to learn a ton of stuff in order to do that at scale and efficiently and all that stuff.

Speaker 1: Yeah. Yep. Where is the where is the business today? I mean, you've raised money. It seems like there's, you know, monetization opportunities for sure. Yep. Are you fully in commercialization? Are you still in research? Is it half and half? Like, how do you think about raising more money over time and and just growing the business?

Speaker 7: Yeah. So we're we're always going like, we we started with like this curiosity driven drive to figure out how to digitize Snell, which is like a pretty wacky thing to do. So that we're always going to be trying to push the edge here.

Speaker 9: Yeah.

Speaker 7: But look, we have a factory. We manufacture fragrance for brands. We did this commercial kind of r and d to commercial transition last summer. Sure.

Speaker 4: And we're

Speaker 7: kind of almost at the end of that, and we built a manufacturing organization. We built a sales organization. We have some really amazing partnerships with some big brands. We're making fragrances for brands. You can go into Target and buy a product that has our fragrance in it today.

Speaker 1: No

Speaker 7: way. And so we're scaling this part of our business. We're still placing bets on the future though. Right? So I think we've got really the tiger by the tail in this it's a whole other conversation. Sometimes you should come to the factory in in New Jersey and see how it operates. But like the fragrance industry is wild. Yeah. We've got a lot of work to do there, lot of opportunities, so we're focused on that.

Speaker 1: Amazing. Well, and thank you for the work that you do. We think it's so important.

Speaker 2: And One of the most interesting companies we've ever learned about on the show. Yeah.

Speaker 1: True science fiction. Awesome. I love it.

Speaker 7: And And we're trying to make science fiction into science fact but like open invitation to come see how it all gets made. It's crazy in person. So come to the Willy Wonka chocolate factory where all this stuff happens.

Speaker 1: I would love to. Thanks so much. It's so great