SDSignal Desk

Osmo says it has digitized the nose. Its own data shows how human the work still is

· Updated · Signal Desk newsroom

Analysis of reporting by MIT Technology Review. How we work

Image: MIT Technology Review

The Google Brain spinout claims “human-level olfactory intelligence” and real fragrance customers. The dataset is impressive; the claims are largely self-defined, and the hardest uses, like medical diagnosis, remain far off.

Key takeaways

  • Osmo defines human-level smell as matching the average person at detecting and telling smells apart and a junior perfumer at creating them.
  • It built a dataset of more than 10 million human labels and 800 million “scent tokens,” using about 70 trained labelers and round-the-clock chemical sensors.
  • Its technology has been used on retail fragrances, including for Procter & Gamble, and to spot counterfeit sneakers with StockX.
  • A $3.5 million Gates Foundation grant led to eight substances Osmo says repel insects better than DEET.
  • Critic Avery Gilbert calls some claims undocumented; a former consultant says context, which models lack, shapes how we perceive smell.

The claim, and its definition

Osmo founder Alex Wiltschko told MIT Technology Review that in recent months his company’s model reached “human-level olfactory intelligence.” It is worth reading the fine print, because Osmo supplies the definition: about as good as an average person at detecting and discriminating smells, and about as good as a junior perfume designer at creating them. The company also says its model recently beat in-house perfumers at one of the world’s largest consumer goods companies, which it did not name in the piece.

Those are meaningful milestones if they hold, but they are self-graded. Unlike language or image models, smell has no widely shared public benchmark that outsiders can rerun, which makes it hard to separate genuine capability from good framing. We would treat “human-level” here as a marketing summary of real progress, not a scientific finding.

Building the data from scratch

The most impressive part of Osmo’s story is the unglamorous one. Wiltschko spun the company out of Google Brain in 2023 and found a field starved of data: by his count, fewer than 200,000 human smell descriptions existed in prior databases. Osmo has since built more than 10 million human labels and 800 million scent tokens describing molecules and concentrations, with spectrometers running around the clock and roughly 70 trained labelers sniffing samples and writing standardized descriptions. A school-bus-sized robot prints scents to order from text prompts.

That approach mirrors what drove progress elsewhere in AI, and it also exposes the dependency. Smell scientist Avery Gilbert has argued that Osmo’s patent filings still lean on human panels and perfumers. Wiltschko’s response, that the model learns from what people smell the way language models learn from what people write, is reasonable. It also concedes the point that this is a system for predicting human perception, built on human labor, like the data markets quietly feeding other models.

Fragrance today, diagnosis someday

The near-term business is real. Osmo has worked on retail fragrances for Procter & Gamble, used chemical differences to help StockX flag counterfeit sneakers, and, with a $3.5 million Gates Foundation grant, identified eight substances it says repel disease-carrying insects better than DEET. Those are controlled, commercial settings where the model’s limits matter less.

The bigger ambitions, airport threat detection, food-spoilage sensing, disease diagnosis, are a different order of difficulty. Former Osmo consultant Michael Schmuker, who helped build an open-source reverse-engineered version of Osmo’s 2023 model, notes that the same molecule can mean Camembert or a dirty diaper depending on context. Sensors need to improve, too. Wiltschko himself says companies that charged straight at medical diagnostics have failed. We appreciate that candor, and it is a reason to judge Osmo by the fragrance and repellent work it can show, much as we judged robotics claims against what labs actually demonstrate.

Who it affects

Fragrance and consumer goods firms
Faster scent design and formulation could compress product cycles and reduce reliance on scarce senior perfumers.
Perfumers and sensory panelists
Their judgments train the system, and the system is now pitched as matching junior designers.
Public health
Better insect repellents are a tangible benefit if Osmo’s DEET comparison holds up in further testing.
Medical diagnostics
The long-term prize remains speculative until sensors and context modeling improve.

What to watch

  1. Independent evaluations or peer-reviewed results supporting the human-level claim.
  2. Regulatory or field testing of the insect-repellent molecules.
  3. Whether Osmo names the consumer-goods customer behind its perfumer comparison.
  4. Advances in low-cost chemical sensors needed for real-world detection.

FAQ

What does Osmo mean by human-level smell?

Its own definition: matching an average person at detecting and distinguishing smells, and a junior perfumer at designing them.

How big is Osmo’s dataset?

More than 10 million human labels and 800 million scent tokens, according to the company.

Who are Osmo’s customers?

MIT Technology Review cites fragrance work including for Procter & Gamble, a counterfeit-detection project with StockX, and Gates Foundation-funded repellent research.

Why are some scientists skeptical?

Avery Gilbert says some marketing claims are undocumented, and Michael Schmuker notes smell depends on context that current models may not capture.

Read the original

Continue at the source.

MIT Technology Review

More on AI

All AI coverage →