Great writeup. This focuses heavily on the fragrance/perception side of digital olfaction, but there's a whole other dimension here: detection. Think of it as speaker vs microphone — most of the hype (Osmo, Givaudan) is on generating and designing scents, but detecting and identifying chemical signatures in real-world environments is an equally hard problem.
The article nails the core challenge: the structure-odor relationship is messy and there's no "RGB of smell." We're hitting the same wall from the detection side. Getting to parts-per-trillion sensitivity is one thing; making sense of complex, noisy molecular signatures in the wild is another. The DREAM challenge benchmarks are encouraging, but the gap between lab conditions and real-world deployment is enormous.
Curious if anyone else is working on the detection and safety rather than fragrance. Feels underrepresented relative to the urgency.
LLMs are trained on text about the world, not the world itself. Olfaction is an interesting test case because it's one of the most ancient and direct sensory modalities. no symbolic abstraction layer, just molecular binding triggering pattern recognition.
What's compelling about pairing e-nose hardware with transformer architectures is you get that grounded perception loop you're describing. The sensor array produces high-dimensional response patterns from real physical interactions, and the model learns to classify and reason over patterns it's never been explicitly trained on—genuine novelty detection rather than interpolation over training data.
The "this is outside what I know" capability is critical for real-world deployment. A model that hallucinates a scent classification is potentially dangerous (think: fentanyl detection in law enforcement). You need calibrated uncertainty, not just a softmax score.
This is exactly the problem I've been obsessing over. The challenge is that olfaction isn't like vision. you're not detecting photons at discrete wavelengths, you're dealing with ~400 olfactory receptor types responding to millions of possible volatile molecules in combinatorial ways.
MOX sensors (like the SnO2 in this paper) have been around for decades but hit a fundamental ceiling—they require specific coatings to bind to specific VOCs. Want to detect a new substance? You're changing hardware.
The more promising path, IMO, is carbon nanotube (CNT) sensors that actually mimic how our nose works. Instead of measuring bulk resistance changes, you functionalize CNT arrays to respond to specific molecular binding events—much closer to how olfactory receptors operate. detection of new substances becomes a software/ML problem rather than a hardware redesign. That's how biology does it—your nose doesn't grow new receptors, your brain learns new patterns.
Full disclosure: I'm building in this space (https://nosy.network) Nosy is using CNT paired with transformer models to create what we call a "Large Essence Model" (LEM). LEM "GPT for smell" processes scent information similar to how LLMs process text.
We have created an AI assistant to answer tax questions. It does a better job (for the most part) compared to a system developed by IRS called Interactive Tax Assistant (ITA). I appreciate any feedback for improvement.
We are building a Social Ebook Reader for Tablets - http://www.Appitude.com. Applied for OAF at Seattle and had a chance to talk to Jason Krute. Launched private beta in november and have signed up interesting authors like Mark Victor Hansen (Chicken Soup for the soul), Jay Conrad Levinson (Guerrilla Marketing) etc.