Loading prices…
STKR NewsSTKR News0 of 3 free this month
AI

This smell startup says its AI nose is as good as a human’s

A deep dive into Osmo, the startup digitizing scent through chemical mapping and foundation models, and why this is a massive frontier for AI builders beyond text and vision.

Originally on MIT Technology Review →
AB

Adrian Boysel

Contributor

Nov 18, 2026

4 min read

Photo illustration / STKR News

Digitizing the Invisible Frontier

For the last decade, we have been obsessed with teaching machines to see and speak. We have poured trillions into computer vision and Large Language Models, assuming that if a bot can describe a photo or write a sonnet, it has conquered the human experience. But there is a massive hole in that logic. We are sensory creatures, and one of our most primitive, evocative inputs has remained completely analog: smell.

Osmo, a startup emerging from Google Research, is trying to fix that. They are not just building another sensor; they are building what they call a foundation model for scent. In a lab near Newark, they are doing the grunt work that every serious founder knows is the real secret to AI success: gathering high-quality, proprietary data that nobody else has.

The Newark Sniff Test

Building a model for smell is infinitely harder than building one for text. When you train a model on the internet, the data is already digitized. With smell, you have to create the map from scratch. Osmo has a team of professional sniffers—people whose noses are trained to pick out notes of peach, seawater, or sulfur—working through thousands of samples. This is the human-in-the-loop phase on steroids.

They take a chemical compound, run it through a gas chromatograph and a mass spectrometer to see the molecular breakdown, and then have the humans tag it with descriptive labels. By pairing the chemical structure with the human perception, they are teaching the AI to predict how a specific molecule will smell before a human ever takes a whiff. This is how you move from trial-and-error chemistry to predictive engineering.

Why Builders Should Care

If you are a builder looking at the AI landscape, you might feel like you are late to the party for LLMs. But the world of "Physical AI"—models that interact with the messy, biological world—is still in the first inning. Scent is a data problem masquerading as a biology problem. Once you can digitize a smell, you can reproduce it, filter it, or even use it as a diagnostic tool.

Think about the sectors this disrupts. The fragrance industry is a multi-billion dollar machine built on secretive master perfumers. If you can use a foundation model to design a scent that mimics a rare flower without needing the flower itself, you have changed the supply chain forever. More importantly, smell is a biological signal. Changes in body odor are often the first signs of disease. A digital nose in a smartphone could potentially act as a continuous health monitor, detecting issues long before a blood test would.

The Skeptical Take: Can We Really Map the Nose?

As a founder, I am always wary of the word "foundation model." It has become a marketing buzzword used to juice valuations. The challenge Osmo faces is that smell is incredibly subjective. What smells like "fresh rain" to a developer in California might smell like "damp concrete" to someone in Tokyo. The human olfactory system is deeply tied to memory and context, not just chemistry.

Osmo claims their AI nose is now as good as a human's, but that is a slippery metric. Accuracy in a lab setting is one thing; utility in the real world is another. The hardware problem is also looming. We have screens for eyes and speakers for ears, but we don't have a "smell speaker" yet. Digitizing the input is only half the battle; the output requires a whole new category of hardware that doesn't really exist in a consumer-friendly way.

The Founder's Playbook

What I love about the Osmo approach is the commitment to the "hard stuff." They aren't just wrapping a GPT API. They are building a molecular map of the world. For builders, the lesson here is about defensibility. If you want to build a lasting AI company, you have to go where the data is hard to get. You have to go into the labs, deal with the hardware, and bridge the gap between the digital and physical worlds.

  • Data Supremacy: Osmo is winning because they are creating a unique dataset. They are the only ones with this specific chemical-to-perception map.
  • Vertical Focus: They aren't trying to solve everything. They are starting with the fragrance and flavor industries, where the money is already flowing.
  • Hardware Integration: To make this work, they have to master the interface between sensors and software. This is a massive moat.

The Bottom Line

We are entering an era where AI is moving past the screen. Osmo is a signal that the next wave of unicorns won't just be chatbots; they will be companies that help machines sense the world the way we do. It is expensive, it is slow, and it is chemically complex—which is exactly why it is a massive opportunity.

The most valuable AI companies of the next decade will be the ones that digitize the senses we currently take for granted.

If you are building in AI right now, ask yourself: what is the data that everyone else is ignoring because it's too hard to collect? That is where the real value is hiding. Osmo found it in the air; you just have to find yours.


Read the original at MIT Technology Review →

The Brief

Stay Updated on Cutting-Edge Tech

A six-minute morning dispatch on the markets and the technology shaping them.

Free. No spam. Unsubscribe anytime.

Write for STKR

Become a Contributor

Earn $STKR for published stories on markets, protocols, and culture.

  • Earn $STKR for every published piece
  • Editorial support from the STKR desk
  • Byline visibility across the network
  • First look at the upcoming creator program
Apply to Write

Keep reading

All stories

Comments

24 reader responses