Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarMIT Technology Review→ original

Startup Osmo Claims Its AI Can Distinguish Smells at Human Level

Osmo, a startup from New Jersey, is teaching AI to tell smells apart the way a human does. Experts sniff substances and write down notes like "nutty" or "marine," and those notes are matched against the chemical composition of each substance. Founder Alex Wilczko says the model has reached "human-level olfactory performance" and is helping brands create fragrances. There is no independent verification yet, so this is the company's claim, not a proven result.

AI-processed from MIT Technology Review; edited by Hamidun News
Startup Osmo Claims Its AI Can Distinguish Smells at Human Level
Source: MIT Technology Review. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceMIT Technology Review ↗
Editor-in-Chief
The goals go beyond perfumerydetecting counterfeit goods and diagnosing diseases
Key metric #2Training uses pairs of "expert description + chemical composition of the substance"
Key metric #3Not everyone is sure the technology can live up to these promises

Neural networks have long been able to recognize faces, understand speech, and describe pictures, but computers handle smell differently: the sense of smell has neither a familiar "camera" nor a shared vocabulary for description. Osmo, a startup from New Jersey, is trying to change that. The company is training a neural network to determine what a substance smells like, and it claims that its "AI nose" has reached "human-level olfactory intelligence." This is Osmo's own claim; MIT Technology Review's report includes no independent confirmation. Let's look at what lies behind the bold wording.

How the neural network is taught to smell

In Osmo's lab, researchers are, in the outlet's words, trying to "smell every odor on Earth." Experts inhale a substance and record what it resembles: "nutty," "peachy," "marine." These verbal notes are matched against the chemical analysis of the same substance, that is, data on which molecules it is made of.

The principle is familiar from other neural networks. If you show a model many photos with captions, it learns to guess what is in a new photo. Here, the chemical composition plays the role of the photos, and human descriptions of the smell play the role of the captions. A trained model should predict from a substance's composition how people will perceive its scent. Osmo calls this system the world's first foundation model for smell, meaning a general-purpose base that can be applied to a range of tasks. The source material contains no details about the architecture, the volume of data, or how quality is verified.

What the founder promises

Osmo founder Alex Wilczko states the goal briefly: "to give computers a sense of smell." According to him, the model is already helping perfume brands develop new fragrances. For the industry, the use case is clear: instead of blind trial-and-error formulation, companies can more quickly filter out unsuccessful variants.

The company's ambitions reach further. Osmo talks about detecting counterfeit goods, since a fake is often given away by the smell of its materials or glue. Another direction is disease diagnostics: some conditions change the smell of breath or skin, and an electronic nose could pick up such signals. For now, these are plans, not finished products.

Where skepticism is warranted

The main reason for caution is that the phrase "human level" comes from the startup itself. The article states plainly that not everyone is convinced the technology can deliver what is promised. Smell is hard to measure. People perceive it differently, and the vocabulary for describing it is subjective: what one person calls "marine" another calls "fresh." Diagnostics and counterfeit detection require not a impressive demonstration but stable accuracy under real-world conditions. Clinical trials and comparisons with existing instruments are how such accuracy is checked, but the text says nothing about such tests.

What this means for practitioners

For founders and product managers, this is an example of how the foundation-model idea is moving from text and images into new types of data. The value of such projects lies not in the neural network itself but in the unique dataset: in Osmo's case, smells described manually by experts and matched to chemistry. Such a dataset is difficult to replicate.

If you work in perfumery, the food industry, quality control, or medicine, it makes sense to follow Osmo's publications and any independent tests. One thing is known for certain: the company is assembling a large dataset, building a model on it, and already offering it to perfume brands. Everything else, from disease diagnostics to counterfeit detection, remains a promise for the future.

Story

Key Questions Answered

How the neural network is taught to smell?

In Osmo's lab, researchers are, in the outlet's words, trying to "smell every odor on Earth." Experts inhale a substance and record what it resembles: "nutty," "peachy," "marine." These verbal notes are matched against the chemical analysis of the same substance, that is, data on which molecules it is made of. The principle is familiar from other neural networks. If you show a model many photos with captions, it learns to guess what is in a new photo.

What the founder promises?

Osmo founder Alex Wilczko states the goal briefly: "to give computers a sense of smell." According to him, the model is already helping perfume brands develop new fragrances. For the industry, the use case is clear: instead of blind trial-and-error formulation, companies can more quickly filter out unsuccessful variants. The company's ambitions reach further. Osmo talks about detecting counterfeit goods, since a fake is often given away by the smell of its materials or glue.

Where skepticism is warranted?

The main reason for caution is that the phrase "human level" comes from the startup itself. The article states plainly that not everyone is convinced the technology can deliver what is promised. Smell is hard to measure. People perceive it differently, and the vocabulary for describing it is subjective: what one person calls "marine" another calls "fresh." Diagnostics and counterfeit detection require not a impressive demonstration but stable accuracy under real-world conditions.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

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.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…