Brief IA
›

Osmo Claims Human-Level Artificial Sense of Smell

Research·Tom Levy·

Osmo Claims Human-Level Artificial Sense of Smell

Osmo Claims Human-Level Artificial Sense of Smell
Key Takeaways
1Osmo compares its model to an average person for detecting and distinguishing odors and to a novice perfumer for creating them.
2The company has amassed over 10 million human annotations and 800 million olfactory tokens.
3Avery Gilbert challenges some commercial claims; Michael Schmuker emphasizes the role of context in human smell.
Why it matters : The technology is already being used to develop perfumes, but its application for medical diagnosis is still being considered.

Osmo claims to have reached human-level performance with its olfactory AI model, which, according to the company, can detect and distinguish odors like an average person and compose scents like a novice perfumer. Its technology has already been used in commercial perfumes, including for Procter & Gamble. Some of the company's claims are contested, while medical applications remain under consideration.

Scientists Discuss the Claimed Capabilities

Olfactory specialist Avery Gilbert, author of What the Nose Knows, contests some of Osmo's commercial promises on Substack, deeming them insufficiently documented. He also argues that patent applications continue to rely on human evaluators and perfumers rather than AI. In March, he criticized this dependence on human judgment in light of the AI capabilities being touted. Osmo had not responded to these criticisms by the time of publication.

Michael Schmuker, a former professor of neural computation at the University of Hertfordshire and a former consultant for Osmo, praises it as a technical achievement. Along with colleagues, he reverse-engineered an open version of the 2023 model without access to the data volumes accumulated by the company since. However, he believes the system may not capture certain nuances of human olfaction, particularly its relationship to context.

To explain this limitation, Michael Schmuker mentions a molecule found in both Camembert cheese and used diapers: context allows humans to differentiate these odors. He notes that distinct associations can arise from the same molecule. Last month, Osmo claimed that its model had outperformed in-house perfumers at one of the world's largest consumer goods companies.

Custom Perfumes to Repellents

According to Osmo, brands are already using its technology to develop and manufacture their scents. The company has a robot as large as a school bus that assembles ingredients according to precise formulations. It produces samples for internal research as well as perfumes for paying clients. The design can start with a written instruction, akin to a request made to ChatGPT.

At the Web Summit in Lisbon, Alex Wiltschko demonstrated on his computer a perfume request reminiscent of a Portuguese egg tart, aimed at men aged 18 to 23. After a few minutes, the model provided three proposals with creamy, coconut, and fruity profiles, along with the molecules to be used. If the order had been confirmed, the scent would have been available about a week later.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Other projects extend beyond perfumery. The Gates Foundation, which invested $5 million at launch, later awarded Osmo $3.5 million to test repellent molecules against disease-carrying insects. The company identified eight substances that are more repellent than DEET. It has also collaborated with StockX, an online clothing marketplace, to detect counterfeit sneakers based on the chemical differences between their materials.

Human Evaluators Feed the Model

Founded in 2023 by Alex Wiltschko following his research at Google Brain, Google's former AI lab, Osmo relies on building large datasets. According to its founder, previous databases contained fewer than 200,000 human-written odor descriptions. This scarcity is compounded by the high cost of chemical sensors and the confidentiality of research in the fragrance industry. Osmo presents the results of its work as the world's first foundational model dedicated to olfaction.

Near Newark Airport in New Jersey, about 30 experienced evaluators describe odors; their assessments are compared with chemical analyses to train the system. More broadly, around 70 trained annotators, based in the U.S. and elsewhere, use a common lexicon of perfumery. Osmo has gathered over 10 million human annotations and 800 million olfactory tokens, which specify the present molecules and their concentrations.

Spectrometers and other sensors operate continuously to analyze fruits, flowers, raw materials, or organs. A variation in molecular proportions can significantly alter human perception of an odor. Alex Wiltschko states that the tokens associated with a newborn largely overlap with those of fresh Barilla spaghetti.

Medical Uses Would Require Better Sensors

Alex Wiltschko aims to give computers olfactory capabilities and envisions providing this sensory input to global AI models. He describes odor molecules as a mode of communication accessible to most species, unlike sound or vision. He also emphasizes their connections to emotions and memory, with a single exposure potentially leaving a lasting impression.

The founder envisions applications such as threat detection in airports and disease diagnosis. Possible uses include identifying spoiled food on a production line or detecting tumors by smell. Pilot studies have already trained dogs for this latter task.

Michael Schmuker believes that such applications would require significant advancements in sensors; the models would also need to continue improving. Alex Wiltschko reminds that companies have failed by directly targeting very challenging and costly applications. He describes Osmo's approach as building a technical foundation while exploring multiple avenues.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.