Startup Osmo Trains AI on Smells Using 10 Million Human Ratings and Generates Fragrances From Text
Startup Osmo is teaching a neural network to "smell." From a molecule's composition, the model predicts how a person will rate the scent, and it assembles a fragrance from a text prompt. Fewer than 200,000 usable scent descriptions existed, so the company built its own dataset: 10 million ratings from human sniffers and 800 million "tokens" of records on molecules and their concentrations. The claimed success has not been independently verified. The takeaway: in a field without data, you have to build the dataset yourself.
AI-processed from Habr AI; edited by Hamidun News
Neural networks have long learned to understand vision and sound, but smell remained a blank spot: there are no ready-made databases and no familiar "text from the internet" to learn from. Startup Osmo, founded in 2023 by former Google Brain researcher Alex Wiltschko, tackled the problem with brute force. The company collected more than 10 million ratings from living people who sniffed samples, plus 800 million "aroma tokens," which is what it calls precise records of which molecules, and in what concentrations, make up a scent.
Osmo is using this data to build what it calls the first "foundation model" for smell. The model is meant to predict how a person will perceive a scent based on its molecular composition. The next step in the plan is to generate new fragrances from a text prompt, much as ChatGPT writes text from a prompt.
Where the data comes from
The main problem with smell, according to Wiltschko, is a lack of data. Before Osmo, the available databases held fewer than 200,000 human odor descriptions. Chemical analysis equipment is expensive, and perfume companies keep their research secret.
So Osmo took the most labor-intensive route and built its own dataset. Spectrometers and chemical sensors run around the clock, analyzing the composition of everything from fruit and flowers to food and even internal organs. At the same time, about 70 specially trained people in the US and other countries sniff samples and record their impressions using standardized terminology. These are the ratings: if chemistry tells you what a smell is made of, people tell you what it smells like.
The idea resembles working with language. Rearranging words changes the meaning of a sentence, and changing the composition and proportions of molecules changes the perceived aroma. Wiltschko offers an example: the set of molecules in a newborn's smell largely overlaps with that of fresh spaghetti.
What already works and what is still just claims
Osmo says its model has reached "human-level olfactory intelligence." On recognizing and distinguishing smells, it reportedly does not fall short of an average person, and on creating fragrances, it reportedly matches a beginning perfumer. A month ago, the company said, its AI outperformed in-house perfumers at one of the largest consumer goods corporations.
The technology is already in use. It is applied to developing commercial fragrances, including for Procter & Gamble. For the marketplace StockX, Osmo proposed finding counterfeit sneakers by differences in the chemical composition of materials. Samples based on the generated formulas are produced by a robotic system the size of a school bus: it mixes ingredients and makes test batches for research and for clients.
Further plans are more ambitious: detecting threats in airports and diagnosing diseases by smell. Wiltschko himself acknowledges that tackling this "mountain" head-on is expensive and difficult. So the company, in his words, is first laying a foundation at the base.
Criticism and honest limitations
All the key figures on the model's quality still come from the company itself. The material includes no independent verification.
Olfactologist Avery Gilbert criticized Osmo in a Substack newsletter. According to him, some of the marketing claims are unsupported, and the patent applications rely on human sensory panels and perfumers rather than on AI. He calls it odd that a company with "cutting-edge AI" depends so heavily on subjective ratings. Wiltschko responds that human perception is not a "crutch." Language models learn from what people have written, while Osmo's model learns from what people feel, and that is why it can predict how a smell will be perceived.
Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who consulted for Osmo, took a more cautious position. He considers the technology a significant achievement. Together with colleagues, he released an open version of Osmo's 2023 model, reproduced through reverse engineering, but it lacks the vast data collected since then. At the same time, Schmuker points to context: Camembert and used diapers contain similar molecules, yet people don't confuse them. In a perfume lab, such a gap may not be critical. For an "olfactory channel" in world models or for medicine, the sensors and the models themselves will need major improvements.
Practical takeaways
For founders and product managers, Osmo's story shows how to build a model in a field without ready-made data. First, you build an expensive proprietary dataset: instruments plus trained human raters. Then you find a narrow market where errors are not critical and paying customers already exist (perfumery, authenticity verification). Only after that do you move on to ambitious goals like medicine. The open question is how far the claimed "human level" will hold up outside the lab and without marketing language.
- Nov 21, 2026Anthropic mathematician says Claude helped prove the Riemann hypothesis in six weeks
- Nov 21, 2026North Korea Shows First Chatbot "Source of Wisdom": Science Only, No Politics
- Nov 21, 2026OpenAI Opens ChatGPT Spaces to Plus Subscribers: Shared Documents with AI
- Nov 21, 2026Chinese Tech Giants Double AI Spending to $139B, While Labs Recover Less Than 8%
Key Questions Answered
Where the data comes from?
The main problem with smell, according to Wiltschko, is a lack of data. Before Osmo, the available databases held fewer than 200,000 human odor descriptions. Chemical analysis equipment is expensive, and perfume companies keep their research secret. So Osmo took the most labor-intensive route and built its own dataset. Spectrometers and chemical sensors run around the clock, analyzing the composition of everything from fruit and flowers to food and even internal organs.
What already works and what is still just claims?
Osmo says its model has reached "human-level olfactory intelligence." On recognizing and distinguishing smells, it reportedly does not fall short of an average person, and on creating fragrances, it reportedly matches a beginning perfumer. A month ago, the company said, its AI outperformed in-house perfumers at one of the largest consumer goods corporations. The technology is already in use. It is applied to developing commercial fragrances, including for Procter & Gamble.
What are the key details on Criticism and honest limitations?
All the key figures on the model's quality still come from the company itself. The material includes no independent verification. Olfactologist Avery Gilbert criticized Osmo in a Substack newsletter. According to him, some of the marketing claims are unsupported, and the patent applications rely on human sensory panels and perfumers rather than on AI. He calls it odd that a company with "cutting-edge AI" depends so heavily on subjective ratings.
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.