deniz.in

Markets

Weather

Loading weather

· via MIT Technology Review – AI topic

Osmo says its smell model matches human noses and outperformed one firm's perfumers

Osmo, an AI startup spun out of Google Brain research, claims its smell foundation model matches average humans at detecting odors and junior perfumers at creating them — though not everyone is convinced.

Osmo says its smell model matches human noses and outperformed one firm's perfumers

Startup claims a milestone for machine smell

Osmo, a company founded in 2023 by Alex Wiltschko after his work at Google Brain, says its model for smell has reached what it defines as human-level performance, according to MIT Technology Review. In practice, that means the system is roughly as good as an average person at noticing and telling odors apart, and roughly as good as an early-career perfume designer at inventing new ones. The company, which calls its system the first foundation model for smell, also claims that in the month before the report its model beat the in-house perfumers at one of the world's largest consumer goods companies. That result, like the headline claim, comes from the company rather than independent testing.

For Wiltschko, chemical signals amount to a kind of primordial language shared by most species, and one with unusual power over human memory and emotion. His stated goal is to give computers that sense.

Building a dataset from scratch

The reason nobody had done this before is data. By Wiltschko's count, existing databases held fewer than 200,000 human descriptions of smells before Osmo arrived; chemical sensors are costly, and the fragrance trade keeps its olfactory research to itself.

So Osmo manufactured its own training corpus. MIT Technology Review reports the company has collected more than 10 million human labels and 800 million "scent tokens," its term for records of the specific molecules and concentrations behind particular odors. Spectrometers and other chemical sensors run around the clock at a lab near Newark Airport, breaking down the molecular makeup of everything from fruit and flowers to pantry staples and organs. Around 70 trained data labelers in the US and overseas sniff samples and write down what they perceive using a standardized fragrance vocabulary; a team of about 30 experienced sniffers works through the odor catalog at the Newark-area facility.

The exercise has surfaced odd parallels. Wiltschko notes that shifting the proportions of molecules can change a smell as thoroughly as reordering words changes a sentence, and that the tokens describing a newborn baby overlap heavily with those for a plate of cooked spaghetti.

From prompts to perfume

Osmo does not only analyze; it composes. A scent-printing robot the size of a school bus mixes ingredients to precise formulas, producing research samples and fragrances for paying customers. The model accepts text prompts in the style of a chatbot: MIT Technology Review's reporter asked for a fragrance evoking a Portuguese custard tart aimed at men aged 18 to 23, and the system returned three candidate scent profiles — creamy, coconutty and fruity — along with the molecules needed to make them, with delivery possible in about a week.

The technology is already in commercial use, per the report. Osmo has created retail fragrances for Procter & Gamble, and the Gates Foundation, which invested $5 million when the company launched, later awarded a $3.5 million grant to screen molecules that repel disease-carrying insects — work that surfaced eight substances more repellent than DEET. Osmo has also partnered with the online marketplace StockX on spotting counterfeit sneakers through chemical differences in the materials.

Critics want more evidence

Avery Gilbert, a smell scientist and author of What the Nose Knows, has been critical of the company on his Substack, arguing that some marketing claims lack documentation and that Osmo's own patent filings still depend on human sensory panels and perfumers rather than AI. MIT Technology Review says Osmo did not respond to a request for comment in time for publication.

Michael Schmuker, a former professor of neural computation at the University of Hertfordshire who previously consulted for Osmo, considers the technology a genuine achievement; he and colleagues built an open-source recreation of Osmo's 2023 model, though without the data the company has gathered since. His caveat is context: the same molecule can read differently depending on circumstances, and a French cheese like Camembert shares chemistry with used nappies without anyone confusing the two. That may not matter much in a controlled fragrance lab, but it could matter for uses like spotting spoiled food on a production line or detecting tumors the way trained dogs have in pilot studies. Sensor hardware, he says, also needs to improve substantially, with medical diagnostics the biggest and hardest prize.

Wiltschko's answer is incremental: rather than charging straight at diagnostics, a goal that has bankrupted other attempts, the company is building a base of capability and sending smaller expeditions toward the harder peaks.

Why it matters

Most applied-AI attention has gone to text, images and audio. Osmo is a serious attempt to extend the scale-the-data playbook to a sense that has largely resisted digitization, and its commercial traction — consumer goods, public-health screening, counterfeit detection — suggests machine olfaction has near-term value even if the grandest claims hold only partly. It is also a case study in how such milestones get established: "human-level" here is the company's own definition, the standout results are self-reported, and independent researchers point to missing pieces such as contextual understanding and better hardware. The distance between a working fragrance tool and a diagnostic nose remains the open question.

  • #ai
  • #machine-learning
  • #fragrance
  • #startups
  • #sensors

Related posts