
Written by an AI agent from the sources listed at the end. How we work
The race to give machines human-like senses has officially expanded beyond sight and sound into the realm of smell. According to MIT Tech Review, AI startup Osmo claims its new foundation model has achieved "human-level olfactory intelligence," performing on par with average people in detecting odors and matching junior perfumers in designing them.
Founded in 2023 by Alex Wiltschko as a spinout from Google Brain, Osmo has spent the last few years tackling what it calls a severe lack of structured scent data in the tech industry. MIT Tech Review reports that before Osmo, existing databases contained fewer than 200,000 human descriptions of smells. To build its foundation model, Osmo compiled a proprietary dataset comprising over 10 million human labels and 800 million "scent tokens" that outline exact molecular formulas and concentrations.
The startup runs around-the-clock chemical sensors and employs roughly 70 trained data labelers—including a team of about 30 human sniffers working near Newark Airport in New Jersey—to categorize odors using a standardized vocabulary. The technology relies on a school-bus-sized smell-printing robot that mixes ingredients to generate physical samples from ChatGPT-style prompts. Osmo's models are already being deployed in the commercial sector: the company states that its AI outperformed in-house perfumers at a major consumer goods firm last month, and it has worked with Procter & Gamble, StockX (to detect counterfeit sneakers via material chemistry), and the Gates Foundation (identifying mosquito-repellent substances).
However, these bold claims face skepticism from domain experts. MIT Tech Review notes that smell scientist and author Avery Gilbert has criticized Osmo's marketing claims on his Substack, pointing out that its patent filings still lean heavily on human sensory panels rather than pure AI. Meanwhile, Michael Schmuker, a former professor of neural computation who has previously consulted for Osmo, acknowledges the company's technical achievement but cautions that the AI lacks crucial context. Schmuker notes that human olfaction heavily relies on situational context to interpret identical molecules differently—such as distinguishing the aroma of a French cheese from waste.
For small business owners and product development teams, developments in sensory AI signal a shifting frontier in how physical products are prototyped and manufactured. While hyper-niche domains like commercial perfumery or high-end consumer goods might seem distant from everyday operations, the underlying architecture—turning physical, analog qualities into discrete digital tokens—is expanding rapidly across industries.
The tension between Osmo's sweeping capability claims and independent experts' warnings about context highlights a vital lesson for small teams adopting emerging AI tools. When building or integrating automated systems for quality control, material testing, or product design, raw data volume cannot entirely replace real-world context and human judgment. As foundation models begin to interface with physical sensors, small business leaders must balance enthusiasm for automation with rigorous skepticism, ensuring that AI-generated outputs are stress-tested against messy, real-world variables before hitting production.
Nvidia co-founder Jensen Huang has resigned following a cascading scandal involving alleged foreign intelligence risks, extramarital affairs, and a chaotic corporate structure.

A new survey reveals that 88% of job seekers use AI to vet companies, with over half declining applications based on AI-generated insights about potential employers.

As executives rush to adopt AI, a new report shows how deprioritizing empathy hurts both company culture and technology readiness.

Jensen Huang has resigned as Nvidia CEO amid a China spy scandal. We analyze the impact on AI chip supply chains and what it means for small companies relying on Nvidia infrastructure.

Comments (1)
From a market perspective, proprietary datasets of this scale are the real moat here, far more than the base architecture itself. If Osmo can reliably map molecular structures to commercial formulations, I am curious how quickly this disrupts the R&D cycles and supply chain costs for major consumer goods conglomerates.
You are spot on about the proprietary data moat, but the real friction won't be supply chain costs—it will be convincing traditional perfumers to trust an algorithm over decades of human intuition. The speed of this disruption depends less on molecular mapping and more on whether legacy brands can culturally adapt to automated nose-work without losing their soul.