← Home
TWIML · July 8, 2026 · 59m

How AI Learns to Smell with Alex Wiltschko

Alex Wiltschko
Show Notes

Alex Wiltschko, founder and CEO of Osmo, discusses building artificial olfactory intelligence—teaching computers to smell by mapping the relationship between molecular structure and human odor perception. The episode covers the neuroscience of smell, graph neural networks for encoding scent in multi-dimensional embedding spaces, Osmo's proprietary olfactory dataset, and near-term applications in fragrance, disease detection, and emotion sensing.

This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.

Novel

•
Molecular Structure Doesn't Predict Smell — Perception Does
Two molecules with identical chemical formulas can smell completely different because odor is determined by how olfactory receptors bind to molecules, not by molecular structure alone.
ClaudeChatGPTGemini
•
Olfactory AI Could Detect Disease Biomarkers Before Symptoms Appear
Volatile organic compounds (VOCs) in breath and body odor can reveal disease states—diabetes, cancer, infections—and Osmo's models could teach machines to detect these biomarkers earlier and more reliably than human noses or current diagnostics.
ClaudeChatGPTGemini

Highlights

•
Graph Neural Networks Capture Molecular Relational Structure
Graph neural networks represent molecules as networks of atoms and bonds, allowing AI to learn how molecular topology—not just individual atoms—influences perceived odor.
ClaudeChatGPTGemini
•
Olfactory Embedding Spaces Create Perceptual Neighborhoods
Osmo's models learn multi-dimensional embedding spaces where molecules that smell similar cluster together, even if they're chemically different—creating a mathematical map of human smell perception.
ClaudeChatGPTGemini
•
Proprietary Olfactory Datasets Are a Massive Competitive Moat
Osmo built the largest proprietary olfactory dataset from scratch—a multi-year effort to systematically collect human odor judgments paired with molecular data—creating a defensible advantage that no open dataset can replicate.
ClaudeChatGPTGemini

Misc

✧Wiltschko's background bridges neuroscience and machine learning—he studied how the brain processes smell before building AI to replicate it.
✧Osmo built the largest proprietary olfactory dataset from scratch, a significant engineering challenge in a field with few public benchmarks.
✧The fundamental insight: molecules don't smell the way their chemical formulas predict. AI must learn perceptual neighborhoods, not just chemical similarity.
✧Graph neural networks encode molecular structure as relational data, capturing how atoms and bonds influence perceived odor.
✧Olfactory AI could detect disease biomarkers in breath or body odor before visible symptoms appear.
Was this useful?