Summary
This episode explores the challenge of teaching computers to understand smell, a sense that is far more complex and high-dimensional than vision or hearing. Alex Wiltschko explains how Osmo approached the long-unsolved structure-to-odor problem by using graph neural networks to predict odor from molecular structure. The conversation also covers how the company built a massive proprietary olfactory dataset, including millions of smell labels and aligned chemical measurements, to train predictive models. A major theme is that learned embedding spaces can organize odors into meaningful perceptual neighborhoods, enabling both prediction and discovery. The episode closes by looking ahead to applications beyond fragrance, including disease detection, while emphasizing the need for careful validation and scientifically grounded measurement.
Key Takeaways
- 1Smell is a far richer biological signal than most people assume, with hundreds of olfactory receptor types creating a highly dimensional perceptual space.
- 2Osmo's early breakthrough was using a graph neural network to predict odor from molecular structure.
- 3The model's embedding space groups odors into meaningful perceptual neighborhoods, not just labels.
- 4A proprietary dataset was essential to progress, and Osmo built it from scratch with millions of smell labels and aligned chemical-sensor measurements.
- 5The long-term opportunity extends beyond fragrance into areas like disease screening, but those uses require much larger datasets and strict validation.
- 6Scent may have measurable effects on mood, focus, and anxiety, but the speaker argues these effects should be quantified rather than treated as vague claims.
Notable Quotes
""There are over 300 channels of all factor information in the nose.""
""The structure odor relation problem had been unsolved for a hundred years.""
""It turned out our model predictions were better than any one individual panelist on average in the panel.""
""We've digitized 5.43 million SNPs.""
""we've re've not waited into that space yet because I just I have to do it right like I can't live with myself if I if we don't do this right""
""there are some sense that do things that are positive to your mood and it's almost physiological and it's can increase your focus or can increase your you know awareness or reducing anxiety""
""aroma therapy uh has elements of really deep truth in it that same with Ayurveda like I think that these are really old traditions""
""plants are like our medicine and poison factories they make all the molecules that do good and bad things""
Episode questions
Why is smell considered a harder AI problem than vision or hearing?
Because smell has a much higher-dimensional receptor space and lacks a simple, universally accepted representation like RGB for color or frequency for sound. The transcript also notes that scent cannot be digitized at internet scale in the same way as text or images.
How did Osmo first model the relationship between molecules and odor?
They trained a graph neural network on molecule graphs, where atoms are nodes and bonds are edges. That model learned an embedding space that could predict odor from molecular structure and was strong enough to outperform individual human panelists in one evaluation.
What makes the odor embedding space useful beyond prediction?
It organizes smells into perceptual neighborhoods, where related odors cluster together and nested categories emerge naturally. This means the model can support both classification and more flexible generative or retrieval-style workflows.
What is the biggest bottleneck for expanding olfactory AI into disease detection?
Data collection is the bottleneck, because disease-related scent signals are subtle and hard to measure at scale. Alex argues that detecting conditions like cancer or malaria will require large, carefully designed datasets, not just a clever model.