TWIML AI Podcast

Alex Wiltschko on Smell, AI's Missing Modality

Alex Wiltschko· Founder and CEO of Osmo at Osmo
·~59 min·English·TWIML AI
AI CompanyMultimodalTrainingReasoning
TL;DR

Osmo's Alex Wiltschko explains how a graph neural network finally gave computers a map for smell — turning molecular structure into odor, proving it against a human panel, and using the fragrance business to fund a chemical foundation model.

01Core Mental Model

A Copernican View of Intelligence

<strong>Human intelligence isn't the center of the AI universe</strong> — almost every living thing communicates in molecules, an intelligence AI has barely modeled.

99% of species on this planet can only speak with chemistry, right?

— Alex Wiltschko, TWIML AI Podcast
Key Insight
Today's frontier models are superhuman precisely because they're distilled from human output — text and images we produced. Wiltschko's bet inverts that: to capture a non-human intelligence, train on non-human output, and the largest such corpus is the chemical signaling that runs almost all of biology.
02The Framework

Read, Map, Write

<strong>Digitizing any sense takes three moves — read, map, write</strong> — and smell lagged because it lacked the shared map color and sound already had.

You got to read the world. So, like turn atoms into bits and information. you have to map it, like understand it.

— Alex Wiltschko, TWIML AI Podcast
Key Insight
Images and audio went digital decades ago because the map came first — you can't compress, transmit, or generate a signal you can't encode. Smell stalled for a century not for lack of sensors or sprayers, but because nobody had the coordinate system in the middle.
03Why Smell Is Hard

Three Channels vs. Three Hundred

<strong>The eye reads color on ~3 channels while the nose reads smell on over 300 receptor types</strong> — a dimensional gap that made scent far harder to map.

it's still a mystery exactly what they code for, but it's certainly much higher dimensional, at least in terms of channel count.

— Alex Wiltschko, TWIML AI Podcast
Key Insight
Higher dimensionality isn't just harder to sense — it's harder to represent. Tellingly, when Osmo's learned embedding was tuned, it settled at roughly 300 dimensions to work well, suspiciously close to the biology it was never told about.
04The Breakthrough

The Odor Turing Test

<strong>A graph neural network cracked a 100-year-old problem: predicting smell from molecular structure at the quality of a trained human panelist.</strong>

that problem called the structure odor relation problem had been unsolved for a hundred years and actually in some cases people thought it was unsolvable.

— Alex Wiltschko, TWIML AI Podcast
Key Insight
Framing it as a Turing test is the sharp move: smell isn't just chemistry, it's perception, so the only fair yardstick is a trained human. Reaching an individual panelist's quality means the model isn't approximating a formula — it's reproducing human judgment about what a never-before-smelled molecule will smell like.
05Emergent Structure

A Map That Looks Like Life

<strong>Nobody told the model about biology, yet its map nested jasmine and rose inside floral and drew the fermented region shaped like a bottle.</strong>

fermented and alcoholic was actually shaped like a bottle during our first model train and we haven't touched it since because it's so funny

— Alex Wiltschko, TWIML AI Podcast
Key Insight
When an embedding recovers structure it was never given — floral splitting cleanly into jasmine and rose — it's a signal the representation captured something real. Here the hidden variable is biological origin: scents made by the same organisms (yeasts, flowers) cluster, so the map is quietly encoding the tree of life.
06The Data Moat

No Mechanical Turk for Smell

<strong>You can't scrape smell off the internet</strong> — so Osmo hand-collected 5.43 million sniffs and digitized 6 billion molecules to build a corpus no scraper can copy.

there's no scale AI or there's no mechanical Turk for smell. like we've had to make that internally

— Alex Wiltschko, TWIML AI Podcast
Key Insight
In text AI, data is assumed abundant or purchasable with a credit card. Wiltschko's edge is the opposite premise: when the data doesn't exist, the moat isn't the model — it's owning the only machine that manufactures the training set, and the rate at which it runs.
07The Wedge

Fragrance First, Everything Else Later

<strong>Fragrance is Osmo's first commercial wedge, not its final mission</strong> — it funds the company while teaching the team how to operationalize smell.

the last thing that we do is not going to be fragrance design, but it is certainly the first. And so, it's teaching us and it's also funding us.

— Alex Wiltschko, TWIML AI Podcast
Key Insight
This is the classic deep-tech wedge, executed cleanly: fragrance exercises the very sensing, mapping, and production infrastructure the detection mission will need, so commercializing now advances the roadmap instead of distracting from it.
08What's Next

The Data-Center Era of Smell

<strong>Chemical sensing today is where computing was in the 1970s — mainframe-sized and waiting to shrink</strong>, from Osmo's school-bus printer and two-shoe-box sensor toward a chip.

we're still in the data center era of chemical sensing

— Alex Wiltschko, TWIML AI Podcast
Key Insight
Calling it the data-center era reframes part of the problem as an engineering timeline: the physics already works — a two-shoe-box sensor matches a human nose — so shrinking it is a question of cost, the way every computing capability was. But applications like cancer detection still need task-specific data, validated models, and viable economics, not just a smaller sensor.