Alex Wiltschko - 让计算机拥有嗅觉 - [Invest Like the Best, EP.415]

We smell a lot of stuff. Twice a day, we run sensory panels where we just sniff stuff and label stuff. So, just like Scale AI has people labeling images all over the world, it turns out we couldn't just buy that service from anybody. We had to build it from the ground up.

This is where things get tested and low-latency work happens. When we scale it, that happens elsewhere.

So they're literally going through—

Yeah.

—smelling stuff. Are they notably talented smellers?

We actually have literal rankings. When we need really accurate data of a certain kind, we'll call on our top dogs.

1. Designing New Scent Molecules

In there is synthetic chemistry. Just like a drug company can create a new molecule to affect human health, we create new molecules that affect human perception.

Mm.

We design those on spec for customers. A large customer might say, “All right, we're having this problem making our detergent smell this way,” or, “The regulatory landscape is changing. We can't use this molecule. Can you please help us?”

We take all those requirements, go back into the lab, and use AI. When it's applied to olfaction, we call it OI, or olfactory intelligence. We use OI to design new molecules, which we then synthesize with our synthetic chemistry team, smell them, and, if they work, launch them.

The smell is so remarkable, and it wafts and changes.

Yeah.

It's so cool.

Exactly. It's a little bit maple-syrupy today. I don't know what they're making. I guess they're going through a lot of commercial fragrances right now, so you're smelling the sum total of our work.

Behind you is what's called the perfumer's organ. Each bottle there is a key on a piano, and instead of 88 keys, it has about 1,200 keys. A perfumer can pull any of these ingredients together and mix them in the right ratios to recreate your scent memory.

The smell of fresh laundry was made by a person, right, from some of these ingredients. The smell of a clean kitchen was created from some of these ingredients. Ninety percent of the products in your home have a fragrance, and every single one of those fragrances was crafted by an individual and made by combining these ingredients.

So what we're doing at Osmo is teaching AI about these ingredients and how to work with them in a safe way, to do it super fast, to do it super affordably, and to be able to launch new, beautiful scents that were impossible before.

What is the first set of building blocks for doing that? You've got the individual, isolated smells, or whatever.

Yep.

How do you create the digital footprint for each one of those things?

2. Smells Become Molecular Data

This is a machine called a GC-MS. This is basically a camera for the molecular world. I'll show you how it works.

This is a robotic autoloader, and each one of these has a smell that we want to analyze at the molecular level. This thing can run 24/7, so we load it up and let it run.

What happens is, you suck up a little bit of the smell as a liquid, inject it, and it goes into this half of the device, which is basically an oven with a 50-meter-long, very thin cable. You're shoving the smell through that cable.

What you're trying to make the smell do is act like runners in a marathon. Every molecule in that scent is clumped together, and you experience that as one unified sensation—a smell. You have to separate them to analyze them.

First, you run them through a race. The light molecules make it through the race first, so they can be analyzed one by one here. Then the heavy molecules come out later and later and later. This basically separates the scent into each individual molecule that's in the smell.

Then this side weighs them. The molecules enter the mass spectrometer after being separated, and you hit them with an electron gun that shatters the molecule into pieces. You very carefully weigh those pieces.

Then you play a kind of Sudoku puzzle to figure out: Given the weights of these fragments, and given how long it took to run this race, what was that molecule? Typically, this interpretation is done partly by software and partly by people. What we've done at Osmo is make that happen entirely by software. That's a part of our OI system.

Mm. So how much does the individual atomic unit of smell differ from the combinations? If I think about color—primary colors or something—these are primary smells, is how I'm thinking about it.

Is it pretty reliable, how you can combine those things into some new set of things? What is the periodic-table equivalent?

Nobody knows, but we're teaching the machines to figure that out. That has been the core issue of why scent hasn't been digitized, because exactly what you're saying is—

You don't know what maple syrup breaks down into as primary smells.

Exactly. People have been analyzing the molecular content of these smells for a long time. You can look up in some textbook what the molecules in maple syrup are.

But the ability to say, “Okay, I want maple syrup, but with a little bit more cherry,” or, “I want maple syrup, but don't use that molecule because we know it's not safe. Use this other molecule”—that requires tons of tradecraft. That is what we're automating.

How much will this machine change in the next 5 years if you're successful? Will you be building your own version of this?

I'll show you.

I see GenTech on there. It's not an Osmo machine yet.

These machines are great, and what we're not going to do is change the hardware, because there's about 12 Nobel Prizes' worth of advances inside these machines. They're fantastic.

What we've done is rip out the brains and replace them with our own brains. A lot of what we've noticed is that the hardware is already pretty good in the realm of scent and chemistry, but the software, or the maps that link the different pieces of hardware, has been completely missing.

That's what we built.

This is an inner sanctum here. This is where we keep every AI-designed molecule that we've made, which is probably a significant fraction of all AI-designed molecules ever.

This is just one slice of it. In this room are 10,000 to 20,000 molecules that have all been designed by AI, and we have a digital twin of each. If we need to go back and access one, we know it's in fridge 1, shelf 3, row 2, column 4.

And the sum total of it kind of smells like a bready radish or something like that.

Radish. Yeah.

Yeah.

Do you yourself have an abnormally good sense of smell?

We've brought a lot of people into Osmo who have truly world-class noses, and I can say definitively that I'm not world-class. This is kind of the Rolls-Royce machine. It does the same thing as the other one, except there are 2 more things that are interesting.

One is, you don't have to inject a liquid into this. You can put anything into these vials, and it will suck the smell out of whatever you put in the vials and analyze it.

By turning it into a liquid first, or just directly?

Directly. What it does is pump air into these vials with a needle syringe. The sample gets dropped in here, and a needle is pushed into it. It basically sucks the air and concentrates it onto—you know how Kodak film absorbs light? We have film that absorbs scent.

Yeah.

They basically concentrate the smell on that thin piece of film, and then you move the needle and inject it into the spectrometer. It uses a flash of heat to remove all those molecules. It kind of develops the film, and then the normal machine runs. We analyze the data with AI, and then we can pull back out what the scent actually was.

This means that we can analyze flowers, vegetables, people, and fruits. The first scent that we fully teleported digitally was a fresh summer plum. It was the purple plum. You know, the really good ones have a snap when you bite into them? It was one of those fresh ones.

We sliced it, put it into one of these vials, and analyzed the smell. Then we actually reprinted the smell on the other side of the lab, which I'll show you. The other thing you can do with this machine, which is really cool, is pause the smell at any point in time and sniff molecule by molecule.

A scent will be 30 molecules, 100 molecules, all blended together—different types. You can smell them one by one by putting your nose on here.

Wow.

It's kind of like a debugger for software.

Wow.

This is called a GC-O, or gas chromatograph-olfactometer. When we really want to understand the smell and build our intuition while we're building new protocols, we'll sit here and sniff stuff that comes off the machine.

I understand the strategy behind all this. You've got the ability to read.

Yep.

Then you've got the ability to write.

Yep.

In so doing, those are the first steps to giving computers a sense of smell. We'll talk more later about all the applications that you could then build on top of that capability. Is that how you thought about it—that to give computers the capability in the first place, it's read and write, and write is especially important because it confirms whether or not it's being read correctly?

Exactly. If you can read and write, then you can create this virtuous cycle where you're creating data at every run of the loop.

Right.

If you can actually create new smells and then turn those smells into data readings of some kind, you're training AI, right? If you can tilt that process so the next smells that you create the next day teach the system even more, that's what's called active learning. That's how you get AI systems to get smart really fast, and that's what we do.

Can you talk about the measurement of the fidelity gap between read and write?

Mm.

If I give you the plum—

Mm-hmm.

—and you stick it through your machine and read it in—

Yeah, yeah.

—and then you give me the essential oil—

Yeah.

—on the other side, how do you measure the gap between the smell of one versus the other—

Yeah—

—and how close you are?

I'm simplifying, but I'm going to hand you the real one and the recreated one, and you're going to tell me, “Are we good?”

Yeah, yeah. Close, not close.

There's more nuance to how we do that to create data that can actually be fed into a machine-learning system, but that's effectively it: Do these things match? There are a few tricks that you use to help debias people and get reliable data, but you're the arbiter, right? If it's a smell you're familiar with and I'm trying to recreate a memory that you have, we either did it or we didn't.

We were just with a very famous Hollywood person who said that in the '80s they tried to do this in theaters, where they would have something that puffed out smells.

AromaRama.

Yeah.

Yeah.

And it just didn't work.

They only had certain smells that were whole scenes.

Yeah.

They didn't have primary odors. They didn't have the ability to create any smell. Here on this robot, you're obviously not putting it behind a couch cushion yet, but we're going to make this smaller. The idea is that you need to have all the ingredients together that can be mixed on the fly to create any experience, not just 8 preprogrammed experiences. That's like a slideshow, right? We want an actual display that can show anything.

What is the most surprising thing about primary smells?

We're kind of at the scientific frontier, so everything that we discover every week, every month, pushes back what's known about smell and how to construct it. One thing that I've found in doing science and machine learning and combining these things is that problems that people sometimes think are totally intractable, once you just get started, you're like, “Oh, we're actually making progress.”

The idea of creating scents with AI and creating those scents in partnership with people, fusing human and machine to work in this very emotional world of scent—you just don't think of it. It's like, “Oh, that's crazy.”

It seems crazy—

Exactly.

—and then you start making progress.

Then you start making progress bit by bit. The first scents are dumb. They don't work, or they don't smell right. Then you come back 4 weeks later, and you're like, “That one was really good. Holy crap. I think it's working.”

Then they all start to work, and you start to talk to customers. They start to accept some of your scents for products, and then it really starts to roll. It's bit by bit. You just keep going.

3. Smell Fights Counterfeits

You know what StockX is? They have a problem with fakes from Temu. I want you to hold this in your right hand and smell inside of it, and I want you to hold this in your left hand and smell inside of it. I want you to look at them. Can you tell the difference between these?

Not really.

They're the same, right? They're constructed to be perfectly the same.

This one smells like a new shoe.

That's the fake. You can't work at StockX. But it's more pungent, right?

I'm not a nose guy.

It's more pungent, right? The difference is that the counterfeiters are really good at visual identity, but the smell of the shoe is basically the fingerprint of everything that ever happened to make it.

What we've been able to do with StockX is show that we can take those really big sensors that are in the lab—

Yeah.

—cut the right corners, and make them smaller. The sensors are about the size of these 2 shoeboxes together. That's one right there.

Yeah.

What we do is take the thumb hole of the shoebox and basically insert it into a sniffer—

It will smell—

—and within 20 seconds, it'll tell you whether it's real or fake.

Hmm.

We just have to show it a few real ones and a few fakes. Sometimes it's more, sometimes it's less. Whenever a new SKU shows up, we grab some of the fakes, grab some of the real ones, train it on the new SKU, and then there's a device that can tell them apart.

It's funny to imagine a future where there's an arms race and the counterfeiters are injecting their own—

Yep.

—sexual perfumes.

There already is an arms race. I'm sure this is going to be the next frontier—

Yeah.

—because we're about to really stem the tide.

God, the mind races with applications.

The first investor prospectus that we made, the double-edged sword was, “Okay, we have this entire space to ourselves.” The risk is, how do you focus?

This is kind of a piece of history here.

Oh.

Plum 1.0. This was the first scent that was teleported—the first scent to be digitized and then reconstituted in another place. What we've done with this, and I will show you the vial here, is capture the essence of a fresh summer plum.

In this vial, this clear liquid, is probably thousands or tens of thousands of sniffs' worth of this one moment of biting into a fresh summer plum. What we're showing here is literally everything—the source code of that—

Hmm.

memory of that scent experience. So I want you to smell it. I've already dipped some in the vial, so just put the blotter into the vial. Close your eyes and think of biting into a fresh summer plum.

Hmm. Crazy. That's wild.

Right? Did we get it?

Yeah.

So this is, I think, a piece of history. We've only made 100 of these, but I'd like you to have one.

Oh, wow. Amazing.

It's important to me that what we've done—I think it's just a plum today, but it's a lot more that we're doing in the future.

Hmm.

4. Generation Opens Scent Design

So one thing that we're doing is we're launching a new kind of fragrance house called Generation. The idea is to take all the technology that we've built, but also all the humanity—the people, the perfumers, the noses—and combine that in a new way of designing scents for people who might not have had access to designing a new scent or who haven't been able to do it fast enough. That's called Generation, and that's something that we're just now launching.

What we're doing with this scent is that this is a scent that we custom-designed for a creator, someone who has a very large, active audience on Instagram, has a really great rapport and a brand, frankly, but what she doesn't have access to is a way of creating her own product that is resonant with her values but also just straight-up beautiful. What we're able to do is take all of our technology, take our perfumery, and design her a fragrance in very short order. We're going to help her launch it.

Hmm.

Part of the value proposition of Generation is: Do you want to launch a fragrance? Do you have a place to put it in front of people, but you're missing all the pieces? We can now automate lots of this, and we can bring humanity to the rest of it. Let's work together. Let's build you a fragrance.

If I wanted to go through that process and said, “I wanna create my own”—what are the building blocks of that process? I could start to describe, “I like plums, and I like fennel, and I like this.” Is it that simple? And then—

It can be. Let me tell you how it's done today and then how we're changing it.

Yeah.

The way that you get a fragrance design today—and I'm not even talking about launching the full product, literally just the smell; there's more that you have to do to launch the full fragrance—you submit what's called a brief. In other industries, a brief would be called a request for proposal, or an RFP, and it can be anything. It's actually very free-form in this industry. That itself could be revolutionized.

Let's say you write a Word document and describe your brand. You're gonna describe what you want it to smell like, who you are, what you want the brand to be resonant of, and then at the bottom you'll usually specify 2 pretty important things. What's the benchmark? Is there a scent that you wanna beat? Usually that means, “I kind of want my thing to smell like this, but make these changes.” And then what's the price you wanna pay?

How many dollars per kilo? That can be as low as $10 per kilo. For fine fragrance, it can go up into the many hundreds of dollars per kilo. It just depends on what you wanna launch.

You submit that brief to a fragrance house, and, again, this is the traditional way. Somebody receives that and, in their weekly meeting, they read the brief, the volume that you wanna make, and the price, and they decide whether or not they wanna work on it. Here, you haven't paid the fragrance house anything. They look at the brief and then say, “Okay, we wanna work on this.”

Usually—90% of the time—they go to their library and say, “Oh, we've already made something for somebody else. Let's send Patrick this scent that is from our library.” Then you'll get that, and maybe you like it, maybe you're done. But 90% of the time, sometimes more, you aren't getting a new custom scent. By the way, that process may take weeks or months.

Let's say you push back and say, “Actually, I want something that's really custom.” You're a sophisticated buyer. You say, “Don't give me a library sample.” Now you're looking at a 12- to 18-month process of going back and forth. Every time you submit notes, it might take 3 months for them to get back to you. It's a super-long process.

By the way, once you get the fragrance, you still don't know if it actually works in the application that you're going to put it in, right? You have to do application testing. Let's say you wanna launch a skin cream and you want it to be slightly fragranced. The fragrance can't make the skin cream turn the wrong color, and it can't make the scent change too much, so you have to do application testing as well. Now you've added more time.

That whole process is handled in most fragrance houses with something that looks very close to pencil and paper and a lot of guesswork. What we're doing with Generation is taking each one of those pieces and applying modern methodology—in some cases AI, in some cases just efficiencies—to make that faster and to make sure that you get something custom every time that's actually tailored to your brand, right?

When you submit the brief, that should be a ChatGPT interface. You should have a conversation. That should be available to start instantly. So we have a tool. Do you wanna make a Colossus fragrance?

Sure.

Great. All right.

Why not?

A fragrance for Colossus. What do you want it to evoke?

Well, I'll describe what our mission is.

Let's do it, yeah.

And maybe it'll come out of that. Our hope would be that, by finding, frankly, people like you who are in pursuit of what we would call their life's work—for you, it might be giving computers a sense of smell and all the applications that are born from that—by showing people these great examples of people really doggedly on the hunt to build their thing, they'll wonder what their thing is and start building it.

I'd say we're very focused on business and investing. Those are the forms of art that I like. But I would hope that our work encourages more people to chase their thing because we're showing them such great examples of people like you chasing theirs, and really just give them permission to do so.

So what do I hope it evokes? Possibility. I think of open air. Maybe that would be a smell I think about: the Sequoia parks or the redwood parks in San Francisco. When you get to the top, there's a very specific, crisp smell.

Mm-hmm.

So what's happening behind the scenes is that we're tapping into all the tools we've built, the olfactory intelligence we've built over time. We'll take what we've put in and embed it into our map of scent.

Mm.

Our map is not 3-dimensional like RGB; it's about 300-dimensional, which I think is why scent had to wait for artificial intelligence to be digitized, because it's just more complicated. We'll then decode that coordinate into a scent profile. We'll show you where the scent lives in the map of the 100 top mass-market hits. Then I'll show you the source code of the fragrance. If we have something similar to it, we can actually go grab it and smell it.

Crazy.

This is, by the way, a multi-month process that we're condensing down into minutes.

How long do you think it will be until there's literally something sitting here where the feedback loop is more or less instantaneous?

Yeah. The mountain peak is something you can hold in your hand, like your AirPods and your phone, that can read the chemical slice of reality, capture scent moments, tell you if you need to go to the doctor, tell you what vitamins to take—to really read scent—and then another device that can recreate it.

So that we can fill this room with something that smells like walking through the redwoods. That forest-bathing smell is one of my favorite smells of all time. We're getting there.

You saw the big reader. You saw the one that we made smaller. We have to take that down by a factor of 4 to 8 in order for this to be something you would really say is portable. Right now, we can move it around and deploy it. I don't think you can say it's portable today.

Then you saw the scent printer, which is half the size of this table, right? We've got a lot of work to do to make that thing smaller, but if there's one thing we've known about technology, it's that making things smaller is something we can do.

Yeah.

So let's take a look. We've got an image.

Ambition Trail.

Ambition Trail. All right. The marketing copy is, “A fragrance inspired by a timeless determination in the pursuit of one's mission, capturing the essence of ambition and discovery.”

It seems, based on all this, that it's not that long from now until you enable things that traditionally have scents: candles—

Yep.

Detergent products—

Yep.

Whatever, infusers...

because they're all based on the essential oil that you're delivering—

Yep.

—that people will be able to design scents very soon.

Absolutely. So, we are rolling that out.

And Generation—is that—

Generation is our business to open up scent design to more people, right, and to do that faster.

If this company is going to be the thing that sort of has your name written all over it—

Yeah.

—and that you're the most proud of having created—

Yeah.

—what are the earliest seeds of that story—

Yeah.

—in your life?

5. From Perfume Nerd To Founder

David Senra's in your world. He has a few phrases that really resonate, one of which is, “The exit strategy is death,” right? This is the last thing I want to do.

Yeah.

I grew up in a town called College Station in Texas. Not too big, not too small. I got bit by the computer bug pretty early, so I started programming computers when I was 8 or 9. I was a full-on computer nerd by 12, when I started collecting perfume.

I started collecting perfume because I noticed that when people put this invisible thing on them, they would all of a sudden be treated differently by everybody around them, but within this little radius, right? It was this magic-potion spell combination that, when you just say it like that, is almost unbelievable: Can you spray an invisible thing on you that changes how people see you and treat you, for better or for worse?

I just couldn't understand it. I had already felt a little bit like a social outsider, and I was trying to decode that. Why are they popular, and I feel like I'm on the outside? I looked at the clothes, but it was the fragrance that really confused me to no end at first. Then it fascinated me.

I started looking into fragrance, and I found out what these kids my age were buying. It was Polo Blue and Abercrombie & Fitch Fierce, both fragrances that our master perfumer, who we just passed, designed—

Oh, wow.

—many years ago. So it's completely full circle now.

Then I realized there weren't just 2 or 3 fragrances; there were tens of thousands. It was like discovering that movies exist, and you can go into the movie theater and there’s nobody watching the good films. Everybody's watching the popular films.

For me, this whole world of fragrance opened up. I remember the first fragrance that really taught me that this is an art. It was Bvlgari Black, which is frankly not a very long-lasting or particularly performant fragrance. It comes in a bottle shaped like a hockey puck, and you spray it on and it lasts 45 minutes.

But what it does is unfold in 3 acts, right? The first smell is like the smell of screeching tires and rubber. Then, within 5 or 10 minutes, it cools down to this vanilla rubbed on a leather chair. Then, after another 15 or 20 minutes, there's this smoky tobacco, leather-chair, smoking-room vibe.

I remember the first time I experienced it. “Whoa, this fragrance changed. Has it gone bad?” Then I sprayed it again and again and again and watched this movie play out for an afternoon. I was like, “No, no, somebody made this.”

Yeah.

This is the whole thing. The whole fragrance unfolds over time. That was kind of the end of it for me. I got completely hooked.

Mm.

The way my brain works, I wanted to understand where it came from, how it was made, and how the brain processes it the way that it does. I think if I were born in southern France, maybe I'd be a perfumer, but I was born to 2 academics, and so I became a scientist.

I went to school for neuroscience at the University of Michigan, and then realized that there is a subspecialty of neuroscience called olfactory neuroscience—the study of how the brain processes smell. Most of the people who study that are at Harvard, so I went to Harvard.

After many years of doing science there, I realized that we don't really know how smell works at all. We're making progress and learning things, but a simple question like, “Let me draw a molecule on the whiteboard like we're in chemistry class. Can you look at that molecule and tell me what it's going to smell like? Will it smell like apple or cinnamon or anise or what?”—it turns out that's a 100-year-old problem that nobody had been able to solve.

That really stuck in my craw, and I thought, “Why don't we know how to do this?”

I ended up leaving academia. I started and sold 2 AI companies: 1 in the biotech space and 1 in the pure-ML-as-a-service space. That ML company was bought by Twitter. I helped start their deep-learning team with my co-founders and with another company that we were combined with.

That's where I really learned internet-scale artificial-intelligence applications. We applied AI to their ads platform and made them a lot of money, and applied AI to their data centers and saved them a lot of money.

Then I was recruited away to Google Brain, which is now called Google DeepMind. It's kind of their Xerox PARC or Bell Labs. I worked on some internal projects for a bit, but after a year or so, I said, “You know what? Let's take a crack at this smell problem again.”

It turned out that, in the time between when I left academia and got into tech and entrepreneurship, and when I arrived at Google Brain, some breakthroughs had happened. AI researchers figured out how to make artificial intelligence work on chemistry.

That may not sound too crazy, but up until then, AI systems really liked their inputs to be rectangles, right? Images are grids of pixels, and text is a long, thin string of words. But molecules can have any shape. There can be any number of atoms, and the bonds can be rearranged.

There had been a technique that had been really improved and figured out, made to work better, called graph neural networks. That turned out to be like chocolate and peanut butter for AI and chemistry.

We didn't figure that out, but a lot of my colleagues at Brain, whom I was very fortunate to work with, figured that out for the world of drug discovery. So the intellectual arbitrage that we did was to say, “Let's take those techniques and apply them to the realm of scent.”

I had spent a long time thinking about scent and traveling in that world, so I knew where to get the data sets, where to buy them, where to license them, and how to treat them. We fused those 2 things together, and I was fortunate enough to work with an incredibly talented team of folks at Google Brain. We made this happen together.

What we're able to do is solve this 100-year-old problem. It sounds so simple, but why does this molecule with this shape smell the way that it does?

We validated it in a really stringent way. We basically did a double-blind trial where we predicted the smell of hundreds of thousands of molecules. We picked 400 that were very different-looking from anything we'd seen before, and we kept our predictions secret. We bought or made the molecules, and some of these had never been made before.

We sent them to our collaborator at Monell. Professor Mainland was running this, and he trained a panel of people. This is kind of what we do now, but initially it was at a smaller scale. We trained people to smell something and say, “Okay, this smells fruity and mineral.” That's it. I'd give it a 3 out of 5 for fruit, a 1 out of 5 for mineral, and zeros for the rest.

That's called rate-all-that-apply. It's just how we label data. Then we said, “Okay, we have our predictions. People have their double-blind ratings. Where do our predictions fit within the people?” The best is the average of the panel. That's how you get really high-quality data for AI.

Were our predictions worse than the worst person, or were they somehow in the pack? It turned out that our AI predictions of what these smells were going to be were better than the average panelist.

That means if you're going to add 1 more person to this panel, you'd actually prefer to ask our software what it smells like—it doesn't have access to the physical molecule—than to train up another person to physically smell it. That's kind of like passing an odor Turing test.

When that happened, it was very clear that Mother Nature was not going to stand in the way of continuing on this journey of actually digitizing scents.

If you can solve that 1 problem, it means you can start to ask, “Okay, great. Now what happens if, instead of feeding this AI algorithm a pre-digitized molecule, I feed it a reading from a sensor? Or the data off of a camera, right, if we were talking about images?”

Then what if I ask it to recreate that smell, with the ability to mix together different molecules to create a new scent?

If you can actually round-trip a smell—take a physical smell, put it in 1 system, round-trip it through the reader and this map that we built, this graph-neural-network-based map, and then write it back out again—and it actually smells like the thing that you put in, it means that you have actually digitized a human sense.

We hit all of our scientific milestones at Google, and we asked ourselves, “What's the right way to scale this idea?” That's where Josh Wolfe comes in.

So we were thinking internally at Google: maybe this should be a company. I was working with Krishna Yeshwant at GV, who is a very close friend. We’d worked together for 5 years.

Someone at GV, another investor named Izzy Rosen, was having lunch with Josh, who is the founding and managing partner at Lux Capital. Apparently, Josh had this 10-year-long thesis about digitizing olfaction, and we’d never met. Izzy was listening to Josh give this pitch again, and he said, “Hey, have you talked to this guy Alex? He’s kind of into smell.”

Then Josh and I met, and it was just an instant connection. Josh was integral in pulling this IP out of Google Brain and building it into a completely new company. Josh led the round, Krishna at GV co-led, and we built Osmo. We’re on our way.

6. Building A Business Around Smell

When you think about building Osmo the business, how do you think about the trade-off between creating a pure-play platform—

Yep.

—defined as: you enable—I’ll call them developers—to build any sort of application they want on top of Osmo’s root-level capabilities—

Yeah.

—all the things we’ve talked about—

Yeah.

—and you charge them a platform fee—

Yeah.

—and there are lots of platforms out there that are wonderful businesses—versus, okay, we have the platform, but we’re also going to create the generative, vertical application companies—

Yeah.

—on top of our own raw tech capabilities. What are the trade-offs of one approach versus the other? Are they mutually exclusive?

I think, in the limit, we’re going to be able to explore that design space more fully, but it really depends on what market you’re entering with that platform capability. A lot of successful platforms are entering markets where there’s already a ton of vibrant activity, and they’re helping to grease business and make that happen more fluidly.

There aren’t a lot of fragrance companies out there, right? So there aren’t that many buyers. The question is: do you become a software provider for the incumbents, or do you take your capabilities and compete in that market? I think there have been examples on both sides of this. There are plenty where you are an input or a service provider. I think a recent example where folks decided to just enter and compete would be Metropolis, if you’ve heard of that example.

Yeah, sure.

They’re making software for managing parking lots. The parking lot industry just wasn’t ready to buy that software, but it actually worked. It made parking lots more efficient, so they became a parking lot company, right?

We went through a similar journey. If I could have sold the software here—and believe me, we’ve tried—we’d be selling software. We might not be talking. But there aren’t that many fragrance houses, period.

I think that we have the opportunity not just to sell into an industry that is very large and very secretive, but with this software and these tools, we have the chance to really transform it for the better. The way that business is done here hasn’t transformed for 300 years, and a lot of the way that things are done should stay the same, right? They’ve stood the test of time.

But the world’s getting faster. People are asking for more transparency. People want to make sure that the fragrances they’re using are safe. There are a lot of people who still don’t even know how to get a fragrance made, right?

Look, every company has visual branding. They have a feeling that they’re trying to create for the people who interact with the company, for the people who are in the company, but there is no modality, there is no sense that is more emotional and has deeper ties and associations it can build than scent.

So there are a lot of businesses out there that need to have a smell. And it’s already happening, right? The Ritz-Carlton has a scent.

I remember the Gramercy Park Hotel so distinctly.

Right? And you walk in, and what do you feel when you walk in?

Yeah. Familiarity.

Yeah, familiarity. It’s elegant, right? If you smell it anywhere else, you’re going to think of exactly that hotel.

I think one thing that we’re realizing is that there is an appetite to add scent to more layers of our economy, to more businesses, to more markets, to more products. It’s just inaccessible, and so what we’re trying to do is bring more people to scent and create more scents for people.

That’s what Generation is all about. If you want to make a scent, and if you want to do it quickly, and if you want to do it safely, we’re here. We figured out how to fuse AI with the human aspect to create really beautiful smells.

Teach us just a little bit about smell itself. What’s its history? Why is it so important? Why is it—

Yeah.

—so emotional? Why are scents so memorable?

This sounds hyperbolic, it sounds extreme, but it’s the first sense, right? If you think of us as little microbes a billion years ago, we survived by eating things, and we got better at surviving by eating things by detecting if the thing we want to eat is nearby.

That’s what smell is, right? Smell is like sipping little amounts of the chemical environment around us to figure out where there is more of that thing or less of that thing. So it’s a super-old sense. You can even see it in the brain.

If you smell something, first of all, that is the physical world touching your brain, right? Your brain sends neurons out of your skull into the top of your nose, and your brain is literally touching the world when you smell something.

The information gets to your centers of memory and your centers of emotion faster than any other sense. We’re anatomically wired to have scent project directly to our memory and our centers of emotion. Those areas are called the hippocampus and the amygdala. We’re wired to associate smell with emotion.

It’s evolutionarily super old. There are still a lot of mysteries that remain about smell, and there are amazing researchers who are pushing back the frontiers of what we know, figuring out why things smell the way that they do and engineering better smells.

It’s still the most mysterious sense. We were talking a little bit earlier, and it feels weird that we haven’t figured out scent, right? Computers can see. Computers can hear. They can touch, right? We have touchscreens, and we have the work that the former CTRL-Labs team is doing with these wristbands, haptics.

Yeah, sure.

But computers can’t smell, and that’s weird because it’s such a fundamental thing. It feels almost free or easy for us to smell things. Why can’t we teach computers to do this?

There’s this concept called Moravec’s paradox. Have you heard of it?

Yeah.

Yeah. Just really briefly, the idea is that the easier it is for a person to do, the harder it is for a computer. If it’s easy for a person, it means evolution has spent a ton of time making it easy for us.

But if something is really hard, like proving a math theorem or something, it turns out we’ve taught computers to do that stuff, right? It was weird that those were the first problems to fall.

Walking has been hard, and we’re just now getting good at that by making robots walk. Smelling has been extremely hard, and we’re just now cracking the code there.

So if I think of Generation as the creation and manufacture of any smell that I want—

Yep.

—with no limits on the possibilities—

Yeah.

—of how I could use that scent in a product, in a space, in a—

Right.

—showroom, in a whatever. What are the next 2, 3, 4 things where there’s a big stack of potential utility—

Yeah.

—that would be unlocked because computers can smell? I think of dogs in an airport or something like that.

Totally. That’s what we’re doing with StockX. Again, our main thing is Generation. We think that’s going to have a massive impact. That’s creating smell.

But there are all these applications for detecting smell. I don’t know if you’ve ever experienced this, but a lot of people can sometimes smell if their loved one or their partner is getting sick, or if something’s a little bit off, and then 2 days later they actually get sick. That’s real, right?

What’s on the inside of us gets to the outside. What’s in our blood and in our organs is eventually excreted through our breath, our sweat, whatever, and we know that dogs can pick up on that stuff.

We’ve proved out that we can use the scent of a product to tell whether it’s real or fake, basically tell its provenance, and that’s something we’ve got deployed at StockX, and that’s going quite well.

We think there are other counterfeit-detection and truth-and-safety applications for scent, but I think it goes deeper than that. We could be detecting harmful substances at the border and stopping them. Whatever sniffer dogs are doing, I think eventually a computer will be able to either help with or do entirely.

But a holy grail for us is human health and wellness, right? I think the signal inside the scent that we emit is completely untapped.

What’s weird is it turns out that by getting really good at designing the scents of fruits, flowers, and vegetables, you actually, for free, get good at these other scent problems, like human scent or product scent, because the overlap of the actual molecules that you see is pretty high.

Mm.

There’s not an infinite number of molecules out there. There’s a lot, but if you get really good at one domain of scent, it turns out to help you in other adjacent domains.

Last time you and I spoke—

Yeah.

—we were looking up together on our computers the market caps of the fragrance houses.

Right.

They’re quite huge companies.

Yeah. No, they’re big.

Maybe you can describe why they’ve been such good businesses, the sort of margin profile of these things. What is it that makes fragrance, just in the publicly traded ones that you can go check out, relatively big, good businesses?

Yeah.

7. The Economics And Risks

There are a few pieces here. One is that, at first blush, they’re recession-proof. If people aren’t buying luxury fragrances, they’re buying hand soap, right? Fragrance is in 90% of the products in your household, and there’s a very small number of companies that provide all of that. If one category is going down, another is typically going up, so it has a really, really great long-term profile.

The margin profiles are also very great. Fundamentally, these are manufacturing businesses with non-manufacturing margins because there’s a ton of know-how that goes into producing the finished blended product. Although it’s just ingredients mixed together in a jar that’s then sent to a customer who puts it in their packaging, how you get the exact right blend of those molecules is typically a deeply held secret.

What we’ve done is study the industry very, very deeply and figured out that a lot of what is being provided in the industry can be augmented by artificial intelligence, and we can do this faster. The other piece here is that customers are typically quite sticky. If you’re running a beauty or a CPG business and you run out of stock, your first inclination is to reorder from your past supplier, not to bid out again.

Typically, if you’ve won the business, industry standards for repurchasing are well above 50%. If you build this very wide book of business, with different parts that fluctuate based on the macro, you have a really resilient business there, and the margin profile is typically quite good.

Hmm. If you think about the things that could go wrong on this journey of yours, what do you think gets in the way?

Which I always do every day.

I’m sure you do. What do you think gets in the way?

I’m always paranoid that Mother Nature is going to show up and say, “You’re done,” right? Like, “No, in 2025 or 2026 or 2027, this is not the year for you to peel back another mystery of how this human sense works, and so you’re blocked from taking the next step.”

That could manifest in any number of places. We could fail to make these sensors small enough to be held in your hand at an appropriate price. There could be something fundamental we don’t understand about the world. This is an existential risk that we can never really remove, but we continue into the darkness and into the fog regardless.

What we’re trying to do is never lose sight of the mountaintop, which is that we’ve fully digitized a human sense, and it’s personal, portable, and affordable. Our philosophy for doing that is not to climb up the sheer face of the mountain to that single goal, but to find a route up that mountain with a shallow enough grade where, at some points, we can stop and build a business.

The philosophy here—and I’ve seen other startups fail to do this—is to build along a responsible path that makes you harder to kill over time, as opposed to making your likelihood of success even riskier over time. Because I want to do this for my whole life, I don’t want to just flip this company and sell it. I really want this to survive. It has to survive.

We’re building into our strategy a way to make that much more likely than not.

8. AI Climbs The Scent Curve

One of the things that’s so interesting to me about Osmo is that it is an AI company in a very strict sense.

Oh, yeah. If you look at the org chart, you’re like, “Oh, it’s an AI company that married a chemistry company.”

But it’s quite distinctive in the sense that most AI companies, especially those building what I’ll call applications, are remixes of a lot of the same stuff. There’s a lot of traditional code involved.

Yeah.

There’s software involved.

Yeah.

There’s putting things around the incredibly powerful reasoning models that now exist. The whole world is kind of—I saw the other day that ChatGPT now has 400 million monthly active users.

Mm.

About 5% of the world’s population is using ChatGPT. People are now familiar with these language models and these generative models.

Yeah.

This seems like you’re using the power much differently.

Yep.

I’d love you to describe some of the ins and outs of that as we think about other problems where we can apply AI that aren’t just text generation—

Yeah.

—image generation—

Yeah.

—video generation, that aren’t pure generative in that sense and pure software, and get into some other world. Maybe here it’s chemistry and AI.

Yeah.

But maybe describe how you’re using the tools—

Yeah.

—and how you think about the growth of the capability of those tools—

Yeah.

—and how it will impact what you do.

Technology usually proceeds on an S-curve, right?

Yeah.

It sucks, it sucks, it’s getting better, oh my gosh, it’s getting better super fast, and we’re pretty much done, and it levels off. I think we’re pretty close to the right side of that S-curve with text. We kind of blew past it, but we passed the Turing test, right? I’m regularly fooled and curious whether or not something was written by ChatGPT or not. So text works.

Ilya Sutskever, who is really one of the progenitors of modern AI for text, got up at the main AI conference, NeurIPS, and said, “Look, there’s one internet, and we’ve trained on it,” right? There’s no more data. Yes, there will be some remaining tricks. We’ll make it cheaper, we’ll make it better, and we’ll add reasoning. But we’re out of the raw fuel that drove a ton of the innovation in text.

I think it’s also similar for images, right? We’ve downloaded all the world’s images, and all the image models are trained on all those images. With video, we’re not done yet because it’s super expensive, so we’re not quite at the end of the curve.

Those are just 3 modalities, right? There’s drug discovery. There’s chemistry, the design of chemistry to treat diseases. There’s materials design. There are all kinds of things. What I’m concerned with at Osmo is marching up the S-curve of a human sense.

We’re on text. We’re on vision and images. Those are handled by really brilliant people. But as far as I can tell, we’re the ones who are driving AI up the S-curve for scent, and we’re really at the far left. We’re just starting to take off right now, and the thing that is the fuel here is data.

Yes, we use specific kinds of AI models. We even use LLMs in some of the work that we do. They’re super useful. Our philosophy is the right tool for the job, and so we’re not going to take a dogmatic approach and try to shove everything into an LLM, although LLMs are extremely useful for this, and I think they will get more capable over time, even for what we do.

Most of what’s under the waterline, most of the iceberg, as it were, is just creating the data. It’s having the infrastructure for accepting that data and having the operations to create the physical and the digital, all linked together. You have to have the data. That’s the fuel that drives you up that S-curve.

We’ve just realized this step by step. We started out saying, “Hey, we want to digitize smell.” Great. Where’s the data? We tried to do some licensing deals. We were successful, but the data wasn’t AI-compatible, so we said, “Okay, I guess we have to create it ourselves.”

Bit by bit, we began to build basically the entire AI ecosystem that exists for images, internally and proprietary for the sense of smell. We have a building full of people who label smells all day, every day. We have a laboratory full of sensors that 24/7 are dissecting scent down to the molecular level. We have robots that are creating smell. We’re going to get even bigger robots to create even more smells.

We’ve created the entire virtuous cycle here that allows us to build the data, train the models, and bring these capabilities to the world.

How much of the tooling do you use? Are you tapping models from the major model providers—

Oh, yeah.

—or anyone else, and if so, how?

It’s suffused into everything that we do. Again, it’s the right tool for the job. I don’t think there’s any code that we write that isn’t at least touched by AI in some way, right? It’s just like the new autocomplete.

Yeah.

Right? A ton of boilerplate stuff is just no longer relevant. It's amazing, right? So productivity is higher. Osmo's probably a smaller company as a result of all these AI tools.

There's a bunch of stuff that we just need to get up to speed on, and we can just ask ChatGPT for a reasonable 80/20 answer. “Hey, are there any problems with this NDA?” We can just do that, and then, if it's really critical, we obviously get an informed opinion. So, yeah, it's suffused into almost everything that we do.

What excites you most about the frontier that you're exploring that you have to hold yourself back from spending time on because you're focused on the things you are?

Yeah. I love everything that we do, and if it isn't clear, I really like smell. I really like making smell and experiencing it and sharing it, and so Generation, launching Generation, is kind of a dream come true. But it's not the last thing that we'll do.

What's really wonderful is that, by building all the systems that are going to allow us to create beautiful scents for folks in Generation, that platform is going to help us on our mission to understand human wellness with scent. That is one of the next mountain peaks for us: what in the smells that we're exuding contains information that helps us make better health decisions?

There's already a link there between what we do at Generation and what we will be doing on our journey for the next R&D frontiers. Scent has powerful effects on our mood, and there's already a very deep tradition of aromatherapy. The science, I think, needs to be expanded there, and I think that we'll contribute to that.

As we're turning emotions into data and scent into data, and then data into products for people to build better businesses and launch better products, all of that goes into one platform that's going to help us push back the frontier.

One of my favorite concepts behind technology platforms is that it's very hard to predict what people will use them to do.

Yep.

It seems like every time you digitize anything in technology history, crazy stuff happens—

Totally.

—that we can't predict.

Totally.

And I'm sure that's going to happen here, too.

That's the idea. We know what we need to do now, right? We know the markets where we can be valuable now, and so we're not going to waste any time on anything else other than building a business that makes people happier and makes other people's businesses run better, make more beautiful scents faster.

But it's really hard to predict the future, right? Computers haven't had a sense of smell. I can't see on the other side of that wall. What happens when our sensors actually can fit in your pocket? What happens not just to the products that we build or the partners that we have who are building on our platform, but what happens to society?

Society's different because that computer in your pocket can see and hear. I think it's largely better, right? There's more information flowing. You're remembering more things. You can hold on to moments. There's some beauty that's there that wasn't there before, and I think that there will be beauty in what comes out of what we're doing.

Osmo's young, not that old. Can you describe what you would define as the defining moment so far in the company's history?

I think, on the journey of asking whether or not what we hope to do is possible—“Hey, we've got this crazy dream of actually digitizing smell. Is that possible?”—we had to get together this crazy team that had never been assembled before and then tackle this technology that had never been built before.

When we actually teleported a plum the first time, and we smelled it and it was actually a freaking plum, and it was beautiful and almost hyperreal, I just fell out of my chair. That was a really dreamlike moment. It's like, “This works,” right? Yeah, we did it. Computers can smell now. It's in the lab. We'll make it smaller and cheaper and better, but it's no longer zero. We've gone from zero to one here.

Then there's all the stuff that came out of that, like all these capabilities and tools that were turned into the ability to design scent even better and capture other scents, train the AI models, and build the entire olfactory intelligence platform. But that moment when I smelled the plum was really special.

I think what you're building is singular, very unique. There's no company that I've really encountered quite like this one. What I find so cool about it is that it's taking advantage of the technology that everyone is so excited about, so thanks for letting us in today. It's been an incredible experience seeing it all. I can't wait to see the next iteration of the smells and of the factory—

Oh, yeah.

—and of all the machinery.

Yeah.

When I do these, I ask everyone the same traditional closing question: What is the kindest thing that anyone's ever done for you?

There were moments when people could have closed the door on me or said, “No, thanks,” or, “I don't believe you,” or, “I don't want to take a bet on you.”

I really divide the people who took a bet on me into the 3 chapters of my professional life. First, in my academic life, Josh Burk and then Bob Datta took a huge bet on a know-nothing kid to mentor and to grow as a scientist, and I'm indebted to them for taking the time and taking me in a raw form and helping mold me into a scientist.

Then I moved from academic science into industry, into entrepreneurship, and I have a number of people to thank as well. It's Brian Adams, first of all, for taking a bet on me to co-found a company with him, which we ultimately sold to Twitter. Then, when I moved to Google Brain, D. Sculley, with really no reason, believed in this idea of digitizing olfaction before anybody else did at Google Brain. And then Jeff Dean saw what we were doing and said, “You know what? This weird little thing, let's just let this flower grow. Let's see how it goes.”

Jeff was instrumental in giving us cover, or a force field, just to grow this very delicate, young idea into what it's become today.

Then, when we converted from an industrial research project and decided that the right way to scale this was as a company, a whole new cast of characters took a bet as well, when there was really not a lot of evidence that they should have.

I'm thinking of Andy Palmer, who believed, first of all, that I could do it when I wasn't really sure that I could; Krishna Yeshwant, who was there every single step of the way, both when I was inside of Google and then when it spun out the company; Josh Wolfe, for playing the instrumental role—I mean, the tagline at his fund is, “We believe before others understand,” and he lived that very, very deeply with me to bring Osmo to life.

Then, more recently, Colin Burns at Two Sigma Ventures has been following along with the story and has just been an incredible cheerleader and continues to bet on the company.

Look, this is my board. These are the people that bet on me, and they're involved in actually growing the company. I'm super grateful that those people are one and the same.

I'm leaving people out. There's so many other people to thank, but it's just these little moments where people say yes that can make all the difference.

It's a wonderful way to put it, the little moments where people say yes.

Yeah.

Great place to close.

Yeah.

Alex, thanks so much for your time.

Patrick, thank you so much. It's so fun to bring you here to the lab to show you what we do, to share this passion, and I hope we can do it again sometime soon.