AI Learns to Smell: Can Digital Olfaction Rival the Human Nose?

An AI olfaction startup tries to digitize every smell on Earth, and together with hydropower's predicament, it reveals the real-world friction of technology deployment.
The article introduces two news items from MIT Technology Review's The Download. One is a New Jersey smell-tech startup claiming its AI nose has reached human olfactory levels and is systematically collecting and annotating Earth's smells to build a "dictionary" of smell. Olfactory digitization lags far behind vision and hearing; the difficulty lies in the complex chemical space of odorant molecules, concentration effects, and the subjective, non-standardized nature of human description. The other is hydropower's challenges: changing precipitation affects generation, aging facilities raise maintenance costs, and new projects face ecological and community resistance. The article notes both share the same undercurrent — technology moving from lab to reality must overcome multiple frictions of data, cost, and environment, and breakthroughs are not a linear narrative.
The Ambition of Digital Smell: Letting AI "Smell" Every Scent on Earth
In a lab in New Jersey, researchers are doing something that sounds almost like science fiction: going through every smell on Earth, one by one. According to the MIT Technology Review's The Download column, this smell-tech startup claims its AI nose has reached a level comparable to human olfaction, and the researchers' notes are filled with nuanced sensory descriptions like "nutty" and "peachy."
This sits on a long-neglected AI track. The digitization of vision and hearing is already quite mature — image recognition and speech transcription are practically standard capabilities — but the digitization of smell lags far behind. The reason isn't hard to grasp: the chemical space of odorant molecules is extraordinarily complex, the same substance at different concentrations may present completely different smells, and human language for describing smells is itself highly subjective and lacks a unified standard. To make a machine "smell" accurately, the first problem to solve is how to turn smell into computable data.
This company's approach is essentially to build a "dictionary" of smell — through systematic collection, annotation, and modeling, turning olfactory experience into signals that machines can learn from. If this path works, food quality inspection, medical diagnosis, environmental monitoring, and even fragrance formulation development could all be redefined.
The Technical Path of Olfactory Digitization
There are currently two mainstream machine-olfaction approaches. One is the "electronic nose" based on gas-sensitive sensor arrays (such as metal oxide semiconductors and conductive polymers), which identifies smells by forming "fingerprints" from the response patterns of different sensors to gas mixtures; its advantages are real-time operation and low cost, but its resolution is limited and it is susceptible to temperature and humidity interference. The other is high-precision molecular detection based on analytical instruments such as mass spectrometry and gas chromatography, which has emerged in recent years and can give specific chemical compositions, but the equipment is expensive and bulky.
A more cutting-edge direction is to turn the smell problem into a machine learning problem: first collect smell samples with sensors or spectra, then train models to establish a mapping from "molecular structure to odor perception." This kind of research relies on large-scale, standardized, annotated smell datasets — and such data is precisely what is extremely scarce.
The "Many-to-Many" Problem Between Smell and Molecules
The root of why smell recognition is difficult lies in the fact that chemical structure and sensory experience are not one-to-one. The same molecule at different concentrations may go from floral to a pungent odor; molecules with similar structures may smell radically different, while molecules with very different structures may smell similar. In addition, humans have about 400 types of olfactory receptors, the combinatorial encoding is extremely complex, and there are significant genetic differences between individuals, which makes "what it smells like" lack an objective yardstick. This is also why building a smell "dictionary" must first solve the problem of annotation consistency.
The Predicament of Hydropower: Why Does Clean Energy Also "Drop the Ball"?
The same issue of The Download also looked at the challenges facing the hydropower industry. As one of the most mature renewable energy sources, hydropower has long been regarded as a stabilizer of the power grid, but against the backdrop of climate change and frequent extreme weather, its reliability is being re-examined.
Changes in precipitation patterns directly affect reservoir inflows, and power generation capacity plummets in drought years; on the other hand, maintenance costs for aging dams are rising, while new projects face resistance at the ecological and community levels. The awkwardness of hydropower lies in this: it is both an important pillar of the low-carbon transition and highly dependent on natural conditions, and this dual identity makes its position in energy planning subtle.
For frontier explorations like AI olfaction, the situation of hydropower actually provides a reference — no matter how advanced the technology, it must be embedded in real resource and infrastructure constraints.
The Regulation Value of Hydropower in the Power Grid
The reason hydropower has long been regarded as a grid stabilizer is that its output can be adjusted quickly within minutes, making it suitable for peak shaving, frequency regulation, and emergency reserve, and it can also "charge" through pumped storage. But reservoir inflows are climate-driven and show obvious seasonal and interannual fluctuations, and power generation capacity may drop significantly during dry periods. As the share of volatile power sources such as wind and solar increases, the regulation role of hydropower becomes more critical, and at the same time it is more likely to conflict with power supply security needs in extreme drought years — this is the deeper reason its position is "subtle."
The Common Undercurrent of Two News Threads: The Real-World Friction of Technology Deployment
Looking at AI olfaction and hydropower together, you find they share a theme: when technology moves from the lab to the real world, it always has to face layer upon layer of friction involving data, cost, and environment. For the AI nose to "smell accurately" in real scenarios, it needs massive annotated data and cross-scenario generalization ability; for hydropower to "stand firm" in the energy transition, it needs to cope with climate uncertainty and aging infrastructure.
The value of this kind of news lies precisely in reminding us that so-called technological breakthroughs are rarely a linear narrative achieved overnight. It is more like a long run that requires sustained investment and constant calibration. For those following AI deployment, this obscure track of smell is worth continued tracking — it may be the next sensory domain to be "lit up" by large models.
One-line Summary
An AI olfaction startup tries to digitize every smell on Earth, and together with hydropower's predicament, it reveals the real-world friction of technology deployment.
Paragraph Summary
The article introduces two news items covered by MIT Technology Review's The Download column. The first is a smell-tech startup in New Jersey that claims its AI nose has reached human olfactory levels and is systematically collecting and annotating Earth's smells in an attempt to build a "dictionary" of smell. Olfactory digitization lags far behind vision and hearing; the difficulty lies in the complex chemical space of odorant molecules, the effect of concentration on perception, and the subjective, standardized-language-deficient nature of human description. The second is the challenge facing the hydropower industry: changing precipitation patterns affect generation capacity, maintenance costs for aging facilities are rising, and new projects encounter ecological and community resistance. The article points out that both news items share the same undercurrent — technology moving from lab to reality must cross multiple layers of friction involving data, cost, and environment, and technological breakthroughs are not a linear narrative.
Related articles

How the Grok Team Hits 11,000 PRs a Month: Replacing Manual Review with Deterministic Constraints
Inside the Grok team's formula for 11,000 monthly PRs at a 0.5% rollback rate: Lint rules, type systems, and formal verification as deterministic constraints.

Is Spacetime Made of Something? Six Possible Answers from Quantum Gravity
Is spacetime made of something? This article surveys the main candidate answers from quantum gravity: spacetime may be fundamental, geometry quantized, causal relations prior to distance, spacetime an emergent state of matter, geometry derived from string theory, or no spacetime atoms needed at all.

After GTA 6, Where Does the Game Industry Go?
The Vergecast talks with Game File's Stephen Totilo about GTA 6's impact on gaming: billion-dollar budgets, console market struggles, cloud gaming tests, and the future of microtransactions—and whether it marks the end of an era.