How AI Learns to Smell with Alex Wiltschko - #771

What It Covers

Alex Wiltschko, founder of Osmo and former Google DeepMind researcher, explains how his team built olfactory AI by solving the century-old structure-odor mapping problem, creating a 5.43-million-scent dataset, and deploying fragrance-design models that now generate commercially viable products for real customers.

Key Questions Answered

• •Principal Odor Map: The olfactory embedding space requires approximately 300 dimensions to model smell accurately — matching the 300+ receptor channel count in the human nose. Training a graph neural network on molecule-to-odor pairs produces a structured map where perceptually similar scents cluster as geographic neighbors, enabling arithmetic-style manipulation of fragrance properties.
• •Odor Turing Test benchmark: To validate predictive accuracy, Osmo predicted the smell of never-before-synthesized molecules, sealed those predictions, then had trained human panelists evaluate the physical samples blind. The model matched or exceeded any single panelist's accuracy, establishing a concrete, reproducible benchmark for olfactory AI performance evaluation.
• •Data moat over model architecture: Osmo's 5.43-million-scent dataset — built entirely in-house because no external labeling vendor exists for smell — outpaces century-old fragrance companies whose legacy data sits in disconnected spreadsheets. Prioritizing data collection infrastructure over algorithmic novelty produces compounding advantages as every new customer formulation generates additional training signal.
• •Fleet-of-models architecture: Olfactory intelligence at Osmo runs as dozens of specialized models — covering perceptual prediction, regulatory safety, manufacturability, and consumer preference — rather than one unified foundation model. Regulatory requirements mandate discrete safety outputs, making multi-head or fully unified architectures impractical; the fleet connects along a shared embedding spine similar to autonomous vehicle stacks.
• •Fragrance industry as funding mechanism: Osmo operates a factory robot capable of producing a new fragrance formula every 100 seconds. Customers submit text, image, or audio descriptions; models convert those inputs into formulas blended from pre-approved ingredients. Each commercial order funds further data collection, creating a self-sustaining loop between revenue generation and olfactory foundation model development.