This smell startup says its AI nose is as good as a human’s
AI-assisted digest, checked against the original source before publishing. How we work
Osmo claims its smell foundation model now matches the average person at telling scents apart — built on a dataset the company had to manufacture from scratch.
Alex Wiltschko, who spun his startup out of Google Brain in 2023, says Osmo's model has reached "human-level olfactory intelligence" in the last few months. By the company's own definition, that means the model detects and discriminates smells about as well as an average person, and creates them roughly as well as a junior perfume designer. Wiltschko's stated ambition is to give computers a sense of smell, which he frames as Earth's original language — the only mode most species can communicate in, and one tied powerfully to memory and emotion.
Getting there meant solving a data shortage first. Wiltschko counts fewer than 200,000 human smell descriptions in existing databases before Osmo, while chemical sensor devices are expensive and the fragrance industry is secretive about its olfactory science. So the company built its own corpus: more than 10 million human labels and 800 million "scent tokens" naming the exact molecules and concentrations behind specific odors. Spectrometers and other chemical sensors run around the clock, breaking down everything from fruit and flowers to pantry staples and organs. About 30 experienced sniffers work in a lab near Newark Airport, cross-referencing notes like "nutty," "peachy," and "marine" against chemical analysis; roughly 70 trained labelers across the US and overseas write up smells in a standardized fragrance vocabulary.
Not everyone is convinced. Critics argue that Osmo's biggest claims remain undocumented and still lean heavily on human judgment. Scientists also warn that the model may miss what makes human smell uniquely powerful: context. The same molecule appears in both Camembert cheese and used diapers, yet people never confuse the two — a gap that suggests chemical composition alone doesn't capture perception.
The open questions ahead are whether Osmo will publish evidence backing its parity claims, whether scent models can learn the contextual layer that keeps human smelling from collapsing molecules into confusion, and whether that complexity is ultimately crackable. Doing so could unlock medical diagnostics, widely seen as the technology's largest potential prize.
Key facts
- • Osmo dataset: 10 million human labels plus 800 million scent tokens.
- • Fewer than 200,000 human smell descriptions existed before Osmo.
- • About 30 sniffers in a New Jersey lab; 70 labelers overall.
- • Founded 2023 by Alex Wiltschko, spun out of Google Brain.
- • Osmo says its model beat in-house perfumers at a large consumer goods company.
Why it matters
Osmo is running the familiar scaling playbook on a modality with almost no public training data, meaning its moat is proprietary sensory annotation rather than architecture. Builders should expect scent to follow other domains: early commercial wins in fragrance and detection, followed by scrutiny over undocumented parity claims. The unresolved issue is context modeling — if chemical features alon
Spotted an error? Send a correction.
Get the day's AI news in one email
The Daily Epoch · 60-second read · unsubscribe anytime