👃Osmo Trains AI on 10 Million Human Scent Labels to Give Computers a Sense of Smell
Your computer is about to get a nose
TL;DR
Osmo has built an AI model it says reaches human-level olfactory intelligence, trained on 10 million human labels and 800 million scent tokens. It already beat in-house perfumers at a major consumer goods company. For developers, this points to a new data pipeline where molecular chemistry, sensors, and generative models meet.
A startup called Osmo says its AI model has reached what it calls human-level olfactory intelligence, and it is already being used to design real fragrances. The model reportedly outperformed in-house perfumers at one of the world's largest consumer goods companies, and Procter & Gamble is named among the brands that have used the technology to create retail scents. If you work on generative models, this is a clean case study in building a domain dataset from scratch. Around 30 human sniffers and roughly 70 trained data labelers write down what they smell using a standardized fragrance vocabulary, and those notes get cross-referenced with spectrometer and chemical sensor readings. The result is a pipeline where human judgment and instrument data are treated as paired training signal, which is a pattern many teams will want to steal for other sensory or hard-to-measure domains. The numbers: over 10 million human labels, 800 million scent tokens describing exact molecules and concentrations, and a smell-printing robot that mixes ingredients to precise formulas. The model can take a ChatGPT-style prompt and print a scent to order. Osmo has also identified eight substances more repellent than DEET, backed by a $3.5 million Gates Foundation grant, and worked with StockX on spotting counterfeit sneakers by chemical composition. The catch? Critics say some marketing claims are undocumented and that patent filings still depend on human sensory panels. Medical diagnostics will need better sensors before this leaves the lab. Worth watching whether the claims hold up under independent evaluation.

Key Points
Osmo employs roughly 30 human sniffers plus about 70 trained data labelers who record odor descriptions using a standardized fragrance vocabulary, creating the human side of the training dataset.
Keep reading — it's free
Enter your email to keep reading — plus the best of AI & tech, daily. Free, forever.
Already a member? Sign in
Comments
Be the first to comment