Osmo hired 30 noses and built a 10-million-label scent dataset for AI
Language models grew up on the internet, but scents have no internet. Before Osmo, the world's databases held fewer than 200,000 human descriptions of smells. The startup enlisted about 30 professional noses to describe samples in words like "peachy", while spectrometers break the same samples down into molecules. Each "molecules-word" pair became a label, and there are now more than 10 million of them. If your field has no data, you will have to grow it with human sensors.
AI-processed from TG: @disruptors_official (Дизраптор); edited by Hamidun NewsLanguage models grew up on the internet: humanity spent thirty years writing texts there for free, and engineers only had to collect them. Scents have no such internet. Nobody posts detailed descriptions of what a stairwell or a bowl of soup smells like.
According to Osmo's founder, fewer than 200,000 human descriptions of smells existed in all databases worldwide before them. That is too little to train a large model. So Osmo made the data itself: about 30 professional noses describe samples in words, while spectrometers break the same samples down into molecules.
Each pair of "a set of molecules and a word like 'peachy'" counts as one label. The company already has more than 10 million of them. These are the numbers from the headline: 30 people with a good nose and 10 million descriptions that the model learns from.
How scents got their own "internet"
Osmo was founded by Alex Wiltschko, a former Google Brain researcher. The company is building a foundation model for scents, meaning a large general-purpose model, but not for text, for aromas.
In a lab in New Jersey, professional noses sniff everything: fruit, flowers, pantry staples. They record their impressions in short words: "nutty", "peachy", "marine". Spectrometers run around the clock nearby. These instruments determine which molecules a substance consists of and in what quantities. Around 70 more labelers worldwide describe scents using a single perfumery vocabulary, so the wording can be compared.
Next, the model learns to match one to the other. In plain terms: this set of molecules at this concentration is what people call "smells like peach". The more such pairs there are, the better the model guesses which molecules will produce the desired sensation.
What the model can do and who is paying
It works like a text prompt, except the output is a scent formula. An MIT Technology Review journalist asked for the smell of pastel de nata, a Portuguese egg tart, that would appeal to men aged 18 to 23. The target audience can also be written into the prompt. Within a couple of minutes he received three variants: creamy, coconut and fruity. Then a robot the size of a bus mixes the chosen formula from real ingredients.
For now, the money mainly comes from perfumery, and Procter & Gamble is named among the clients. The plans are more ambitious: diagnosing diseases by smell and detecting dangerous substances at airports. A separate project came from Bill Gates, who paid to search for agents that repel mosquitoes. According to Osmo itself, the model found 8 substances that repel mosquitoes better than the active ingredient in today's sprays. The material contains no independent verification of this claim.
Along the way, some amusing findings surfaced. The molecular profile of a newborn's smell overlaps strongly with that of freshly cooked Barilla spaghetti. Camembert and a used diaper are made of similar molecules, and people tell them apart mostly because they can see or know what is in front of them.
Caveats and takeaways for practitioners
The claim about "human-level sense of smell" should be treated with caution: the company set that benchmark for itself. In essence, the model smells like an average person under neutral conditions, and it composes scents at the level of a novice perfumer. This is a notable achievement, but not a replacement for a master. Moreover, the details come from interviews and press retellings, not from a peer-reviewed study with open data.
For developers and founders, there are three lessons here.
First, if your field lacks a ready-made data corpus, you can build one. Osmo combines objective measurements (spectrometers) with subjective human assessments (noses), and the model learns from the link between them.
Second, a single labeling vocabulary is critical. Hundreds of labelers will produce noise if each names smells in their own way. A perfumery vocabulary turns personal sensations into data fit for training.
Third, the further AI moves from text into the physical world, the more often people end up serving as living sensors. A paid job as a professional nose emerged directly from neural networks' need for data. If you plan to build models for taste, touch or other "non-text" senses, budget not only for compute but also for the people who will create the labels.
- Nov 21, 2026Google DeepMind Launches Genie 4: Text Reshapes 3D Worlds With Sound On the Fly
- Nov 21, 2026The last30days AI skill gained 122 stars in one day and searches Reddit, X and Polymarket
- Nov 21, 2026Perplexity returned risky shops in 15% of shopping answers, ChatGPT in 8%
- Nov 21, 2026XYBench test: AI picks the useful advice in 71–84% of cases, but gives it itself only 35–41% of the time
Key Questions Answered
How scents got their own "internet"?
Osmo was founded by Alex Wiltschko, a former Google Brain researcher. The company is building a foundation model for scents, meaning a large general-purpose model, but not for text, for aromas. In a lab in New Jersey, professional noses sniff everything: fruit, flowers, pantry staples. They record their impressions in short words: "nutty", "peachy", "marine". Spectrometers run around the clock nearby. These instruments determine which molecules a substance consists of and in what quantities.
What the model can do and who is paying?
It works like a text prompt, except the output is a scent formula. An MIT Technology Review journalist asked for the smell of pastel de nata, a Portuguese egg tart, that would appeal to men aged 18 to 23. The target audience can also be written into the prompt. Within a couple of minutes he received three variants: creamy, coconut and fruity. Then a robot the size of a bus mixes the chosen formula from real ingredients.
What are the key details on Caveats and takeaways for practitioners?
The claim about "human-level sense of smell" should be treated with caution: the company set that benchmark for itself. In essence, the model smells like an average person under neutral conditions, and it composes scents at the level of a novice perfumer. This is a notable achievement, but not a replacement for a master. Moreover, the details come from interviews and press retellings, not from a peer-reviewed study with open data. For developers and founders, there are three lessons here.
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.