Startup Osmo Taught AI to Smell at Human Level Using 10 Million Scent Descriptions
Startup Osmo, founded by a former Google Brain researcher, says its AI smells at human level: it distinguishes scents like an average person and composes fragrances like a junior perfumer. The model was trained on more than 10 million descriptions from live sniffers and 800 million "smell tokens" — records of molecules and their proportions in a specific fragrance. For medicine, food safety, and sensors, the data is still not enough, so its real customers today are perfume brands.
AI-processed from MIT Technology Review; edited by Hamidun News
Computers have long had sight, hearing, and speech, but not smell, because there was no proper data for scents. Osmo, a startup founded by former Google Brain researcher Alex Wiltschko, set out to solve this problem using the recipe other AI companies have followed: assemble a huge dataset and see what happens. The result, according to Wiltschko, is "human-level olfactory intelligence."
By that the company means concrete things: the model distinguishes smells roughly as an average person does and creates fragrances roughly as a junior perfumer would. It was trained on more than 10 million descriptions from live sniffers ("nutty," "peachy," "marine") and on 800 million "smell tokens" — Osmo's term for records of which molecules, and in what concentrations, make up a particular smell.
How to Teach AI to Smell
The central challenge of smell is a lack of data. By Wiltschko's count, before Osmo all existing databases together contained fewer than 200,000 human descriptions of scents. Chemical sensors for collecting data are expensive, and the perfume industry keeps its know-how secret.
So Osmo built its dataset itself. In a lab near Newark Airport, about 30 experienced sniffers describe fragrances in words, and their notes are matched against chemical analysis. Around 70 trained annotators in the US and other countries do the same using the standard perfumers' vocabulary. In parallel, spectrometers and other sensors break down everything they can get their hands on around the clock: fruit, flowers, pantry foods, even organs.
The analogy with language is direct. Just as rearranging words changes the meaning of a sentence, changing the proportions of molecules changes the perceived smell. Wiltschko gives an unexpected example: the smell of a newborn baby and freshly cooked Barilla spaghetti share a large overlap of tokens.
What the Model Can Do Now
According to Osmo, last month its AI outperformed in-house perfumers at one of the largest consumer goods companies. The company does not name it, and there is no independent verification of the result.
For the customer, the experience looks like a chat. A reporter from MIT Technology Review asked it to generate a scent for Portuguese pastel de nata pastry aimed at men aged 18–23. Within a few minutes, the system produced three profile options (creamy, coconut, fruity) and a list of molecules for each. Samples are physically printed by a robot the size of a school bus, which mixes ingredients according to the formulas. The order would be ready in about a week.
The technology has already been used for consumer fragrances, including for Procter & Gamble. The Gates Foundation initially invested $5 million when Osmo launched, then awarded a $3.5 million grant to search for substances that repel disease-carrying insects. Osmo found eight compounds that repel more effectively than DEET, a widely used repellent. In addition, together with the platform StockX, the company tested whether counterfeit sneakers can be told apart by smell: the chemical composition of materials differs between genuine and fake pairs.
Where the Skepticism Lies
Not everyone buys the loud claims. Scent expert Avery Gilbert writes on his Substack that some of Osmo's marketing claims are unsubstantiated. The company's patent applications, he says, still rely on panels of sniffers and perfumers rather than on AI. For a company built around AI, he calls this strange.
Professor Michael Schmuker, who previously consulted for Osmo, considers the work a real achievement. Together with colleagues, he produced an open reconstruction of Osmo's 2023 model, but without the dataset the company assembled later. At the same time, in his view, the model may lack context. Camembert and used diapers partly contain the same molecules, yet people don't confuse them. In a perfumer's lab this barely matters. When checking food on a production line or searching for tumors by smell, as has been attempted with dogs, it matters a great deal. Such tasks require far better sensors and further model improvements.
Wiltschko himself acknowledges that medical diagnostics is the largest goal, but companies that pursued it directly usually failed because of cost and complexity. That is why Osmo is building a "foundation at the foot of the mountains": perfumery and repellents first, the rest later.
What This Means for the Industry
For product leads and founders, there are two lessons here. First: in niches where data is scarce, the key asset becomes a proprietary dataset rather than the model architecture. Second: a loud "human-level" claim should be read alongside its definition. Here it means the level of an average person and a junior perfumer, not an expert. So far Osmo's claims have been confirmed mainly by the company itself, but its perfume clients and the Gates Foundation grant suggest the technology has real applications.
- Nov 21, 2026Google открыла предзаказ на очки с ИИ Gemini: от $395 и перевод речи на 70+ языков
- Nov 21, 2026ИИ-помощник для кода дешевле $10: Cursor стоит $8, Xiaomi MiMo — $6 в месяц
- Nov 21, 2026Sales of AI Wearables in Russia Up 7x, AI Earbuds Up 40x
- Nov 21, 2026Anthropic goes public and raises a record $107.3 billion
Key Questions Answered
How to Teach AI to Smell?
The central challenge of smell is a lack of data. By Wiltschko's count, before Osmo all existing databases together contained fewer than 200,000 human descriptions of scents. Chemical sensors for collecting data are expensive, and the perfume industry keeps its know-how secret. So Osmo built its dataset itself. In a lab near Newark Airport, about 30 experienced sniffers describe fragrances in words, and their notes are matched against chemical analysis.
What the Model Can Do Now?
According to Osmo, last month its AI outperformed in-house perfumers at one of the largest consumer goods companies. The company does not name it, and there is no independent verification of the result. For the customer, the experience looks like a chat. A reporter from MIT Technology Review asked it to generate a scent for Portuguese pastel de nata pastry aimed at men aged 18–23. Within a few minutes, the system produced three profile options (creamy, coconut, fruity) and a list of molecules for each.
Where the Skepticism Lies?
Not everyone buys the loud claims. Scent expert Avery Gilbert writes on his Substack that some of Osmo's marketing claims are unsubstantiated. The company's patent applications, he says, still rely on panels of sniffers and perfumers rather than on AI. For a company built around AI, he calls this strange. Professor Michael Schmuker, who previously consulted for Osmo, considers the work a real achievement.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.