Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

Osmo says its smell model has reached 'human-level olfactory intelligence'

MIT Technology Review reports that startup Osmo claims its model, which the company calls the world's first foundation model for smell, has reached human-level olfactory intelligence. The technology is already used in the fragrance industry, but researchers say sensors and context modeling must improve before it can serve medical diagnostics or security screening.

Published
Osmo称其气味模型达到“人类水平嗅觉智能”
Image source: technologyreview.com

In a lab near Newark Airport in New Jersey, a team of around 30 experienced human sniffers works slowly through odor after odor. They jot down notes like “nutty,” “peachy,” and “marine,” then cross-reference them against chemical analysis to build training data for what the AI startup Osmo calls the world's first foundation model for smell.

Osmo founder Alex Wiltschko puts the goal plainly: to give computers a sense of smell. He spun the company out of Google Brain, the search giant's former AI research lab, in 2023, and is running the same playbook used by AI startups in other domains — throw a lot of data at a very hard problem and see what happens.

The hard part starts with data. By Wiltschko's count, fewer than 200,000 human descriptions of smells existed in databases before Osmo; the chemical sensors needed to collect more are expensive, and the fragrance industry is secretive about its olfactory science. Osmo built its own dataset of more than 10 million human labels and 800 million “scent tokens,” which describe the exact molecules and concentrations that produce a given odor. Spectrometers and other sensors run around the clock, breaking down everything from fruit and flowers to pantry staples and organs, while roughly 70 trained labelers in the US and overseas write down what they smell.

Osmo defines “human-level olfactory intelligence” as being about as good as the average person at detecting and discriminating between smells, and about as good as a junior perfume designer at creating them. The company says the model reached that level in recent months and is already being used in the fragrance industry to help brands design and manufacture new scents; last month, it says, the model outperformed in-house perfumers at one of the world's largest consumer goods companies.

The commercial footprint extends beyond perfume. Osmo's technology has been used to create retail fragrances for giants such as Procter & Gamble. The Gates Foundation invested $5 million when the company launched and later gave it a $3.5 million grant to assess molecules that repel disease-carrying insects, work that produced eight substances the company says are more repellent than DEET. Osmo has also worked with the sneaker marketplace StockX on detecting counterfeits through differences in the chemical composition of materials.

Not everyone is convinced. Smell scientist and author Avery Gilbert has been sharply critical of Osmo on his Substack, arguing that some of its marketing claims are undocumented and that its patent filings still rely on human sensory panels and perfumers rather than AI. A representative for Osmo did not respond in time for publication.

The technical caveats are specific. Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who has consulted for Osmo, calls the work an achievement but notes that smell depends heavily on context: Camembert and used diapers share some of the same molecules, and people do not confuse them because the setting differs. In a controlled fragrance lab that may matter little; for sniffing out spoilage on a food line, or detecting tumors the way trained dogs have in pilot studies, both sensors and models will need to get much better.

Wiltschko does not gloss over the difficulty. Many companies have gone straight at the mountain peak of medical diagnostics and failed, he says, because it is expensive and hard. Osmo is instead building a foundation at the foothills and sending expeditions up the mountains — “basically smelling the whole world.”

If smell can be broken into computable tokens the way language can, olfaction could become another channel through which AI enters the physical world, after text, images, and audio. How far that channel reaches will depend on data, sensors, and independent verification, not on the company's own framing.

Why it matters

Smell has long been a gap for AI, and Osmo's attempt to turn the molecule-to-odor relationship into a foundation model opens a route into fragrance, food quality control, and eventually diagnostics. The “human-level” claim still needs outside validation, and the near-term business remains fragrance and consumer goods.

OsmoFoundation ModelOlfactory AI
Back to AI Daily

Nearby Updates

All