Episode Summary
Executive Summary: Alex Wilchko explains how Osmo is building olfactory intelligence: AI that can represent, predict, and design scent by creating the missing data, maps, and infrastructure for smell. The conversation covers why smell is a hard high-dimensional sensing problem, how Osmo trained models to predict odor from molecular structure, and how those models now power fragrance design, safety prediction, and future applications like disease detection and consumer sensors.
Main Topics: Why smell is a hard AI problem (Priority: 5/5): Wilchko frames scent as a physical-world modality that has lacked the digitization pipeline used for text, images, and audio. The core challenge is building a usable map between molecular structure and perceptual smell. The structure-odor map and principal odor map (Priority: 5/5): He describes training graph neural networks on molecule-odor pairs, then extracting a ~300-dimensional embedding that became the 'principal odor map,' revealing meaningful scent neighborhoods and hierarchies. Data collection as the real moat (Priority: 5/5): Osmo’s strategy depends on large, purpose-built olfactory datasets gathered from scratch through human smelling, chemical sensors, product samples, and molecular enumeration rather than scraped public data. Fragrance as the first commercial vertical (Priority: 4/5): The company uses its models to design, blend, and optimize real fragrance products for customers, working with perfumers and a factory to turn requests into market-ready scents. Safety, regulation, and manufacturability (Priority: 4/5): Beyond smell prediction, Osmo models toxicity, skin/eye/inhalation safety, environmental concerns, and manufacturability, because every compound must pass regulatory review before use. Toward broader olfactory foundation models (Priority: 4/5): Wilchko argues future olfactory models could support disease detection, spoiled-food sensing, robot kitchens, and eventually consumer devices, but only after massive data collection and miniaturized sensors. Smell, emotion, and a Copernican view of intelligence (Priority: 3/5): He suggests scent can meaningfully affect mood, focus, and perception, and argues AI should learn from non-human forms of intelligence encoded in chemistry, not just human text and images.
Key Arguments: Smell has been ignored by computing because there was no practical way to read, map, and write it back out at scale; Osmo is building the missing map. Human olfaction is more complex than commonly believed: the nose has over 300 receptor types and can detect extremely tiny quantities of some compounds. A graph neural network can learn the structure-to-odor relationship from molecular graphs, and Osmo’s model reached human-quality odor predictions in a blind comparison. The embedding space is not just predictive but semantically organized, with regions corresponding to scent families and nested subfamilies like floral > jasmine/rose/violet. The largest barrier is not modeling alone but building massive, purpose-built datasets and infrastructure for scent generation, measurement, and labeling. Commercial fragrance is the best initial use case because it provides real customers, repeatable feedback, and data that improves the models. Predicting scent alone is insufficient; practical deployment also requires separate outputs for safety, regulation, and manufacturability. Future olfactory intelligence should include detection tasks such as disease or spoilage, but these will require broad foundational datasets rather than narrow medical collections. Smell likely has a stronger connection to emotion and behavior than current mainstream AI systems account for, but measuring that responsibly remains challenging. Osmo’s long-term goal is to build a true olfactory foundation model, analogous to vision or language models, but grounded in chemistry and non-human intelligence.
Data Points: Olfactory receptor types: Over 300 - Wilchko compares the nose’s receptor diversity to the eye’s limited color channels. Human visual channels: Roughly 3 color channels (RGB) plus related structure - Used as a contrast to the much higher-dimensional olfactory system. Embedding size: Around 300 dimensions - The principal odor map embedding space that emerged from the model. Initial labeled molecules: About 5,000 - The first dataset used to train and visualize the principal odor map. Olfactory dataset size: 5.43 million sniffs - Osmo’s human-labeled olfactory dataset used for AI training. Digitized molecules: 6 billion molecules - Osmo has enumerated and modeled a massive space of possible molecules. Typical odor-active molecule size: Less than 20 atoms, more than 3 or 4 - Wilchko explains why molecules that smell are usually small enough to be airborne and fit receptor binding pockets. Training time for odor labeling: About 8 hours - He says people can be trained fairly quickly to label smells reasonably well. Natural gas odorant sensitivity: Parts per billion or trillion - Used to show how sensitive human smell can be to specific compounds like mercaptans. Sensor device size: About two shoeboxes - Osmo’s deployable chemical sensor is roughly human-nose scale but still far from phone-sized. Scent printer size: The size of a school bus - Used to illustrate the current scale of Osmo’s fragrance production infrastructure. Fragrance production cadence: A new fragrance every 100 seconds - The large robot in the factory can manufacture fragrances rapidly.
Pivotal Quotes: "99% of species on this planet can only speak with chemistry" — Alex Wilchko: He explains why AI should learn from non-human intelligence encoded in molecules. "The thing we focused on at Google Brain was the missing piece, which is for scent is the map." — Alex Wilchko: He defines the core technical challenge as creating a representational space for odor. "We have digitized 5.43 million sniffs" — Alex Wilchko: He describes the scale of Osmo’s internally generated olfactory training data.
Implications: Osmo is turning smell into a computable medium, which could reshape fragrance, sensing, diagnostics, and consumer products. The bigger bet is that AI will expand beyond human-centric data into chemistry-based intelligence.