How AI Learns to Smell with Alex Wiltschko

**Sam Charrington** (0:00)
AI has advanced primarily by learning from the digital world, text, images, audio, and increasingly video. But many of the problems people want AI to solve live outside these modalities in the physical world. Smell is one of the most interesting examples. It's how animals detect disease, identify food, navigate environments, and communicate through chemistry.
Yet, scent has remained largely outside the reach of computing because, unlike language or images, there has never been a practical way to digitize it at scale. Alex Wiltschko, founder and CEO of Osmo, and a former Google DeepMind researcher, is working to change this. His team is building what they call olfactory intelligence, AI systems that can model, predict, and design scents while creating the datasets and infrastructure needed to bring smell to the digital world.
In this conversation, we explore what it takes to give computers a sense of smell, why scent is such a difficult AI problem, and what it teaches us about the next generation of foundation models. Here's Alex.

**Alex Wiltschko** (1:02)
99% of species on this planet can only speak with chemistry. Thinking of bacteria and fungi and plants and insects, they just only can talk with molecules.
And I think that it's really worth adding other alien forms of intelligence to our AI models. And the way to do that is to train it on the intellectual output of those other intellects, which is that's chemistry. That's the sense that's in the air. They're produced by living things for reasons to talk to each other.

**Sam Charrington** (1:30)
I'm Sam Charrington, and this is The TWIML AI Podcast. For over a decade, I've been exploring the ideas and innovations shaping the future of AI through conversations like this one that help you understand what's real, what's next and what matters. Let's jump in.
But I think about giving computers a sense of smell. There's kind of two angles to this. One is, you know, there's some scent out in the world, and I want my computer to be able to recognize it the same way I do. And the other, which is, I think, more along the lines of what you're working on at Osmo, at least initially, is to have the computer kind of grok the idea of scent so that it can create new ones.

**Alex Wiltschko** (2:17)
Any scent that's been given to computers, there's three kind of broad steps. You got to read the world, so turn atoms into bits and information. You have to map it, understand it, so be able to manipulate it, digitally encode it, send it. And that's like JPEG and RGB, right? And then you have to be able to write it back out again, right? So a printer or a display or a speaker.
And so the thing we focused on at Google Brain was the missing piece, which is for scent is the map. So color has had a map for it.

**Sam Charrington** (2:48)
So it's a kind of representation of, well, what to what, though?

**Alex Wiltschko** (2:53)
Exactly, exactly. So let's approach it from the side. Like how did this work for vision? How did this work for hearing, right? We've had maps for a long time, right? So the map for sound is just one dimension. It's low to high frequency, really simple to say. And then for color, it's three numbers. It's RGB, right? Or whatever your preferred color space is. But those three numbers tell you how to deal with color.
There's three channels of color information in our eye.

**Sam Charrington** (3:18)
But of course, you're simplifying a lot, because for each of those other modalities, there's lots of different maps. Those are just examples of simplified examples.

**Alex Wiltschko** (3:26)
And they can kind of be translated into each other. But I'm like, I'm papering over centuries of psychophysics here. And anybody who knows anything about those things is going to come screaming at me. But you have to forgive the simplifications. I'm going to simplify sense stuff too. And if people talk the way that I'm going to talk, I would come after them too.
So yeah, there's CMYK, there's LAB, there's HSV, there's many different maps. And then there's more complex maps.

**Sam Charrington** (3:48)
I was scarred by a DSP class in grass. It all came back in.

**Alex Wiltschko** (3:52)
Exactly, you know. Like wow. There's filter sets, there's Gabor filter sets, there's all kinds of ways of representing images. And I'm super simplifying it, right? But maps, for certainly, they're there. And we know them, we've known them for a while. And the notion that we can map color has been instrumental in building like CCDs and CMOS and therefore like digital imaging. Exactly. And then also the printers, right? So like the ink in the inkjet cartridges, we know we can combine them to make millions of colors.