Lessons from Alex Wiltschko

Alex Wiltschko is a neuroscientist and the founder of Osmo, a company building the hardware and software required to digitize the sense of smell. He is known for his previous work at Google Brain, where he led a research team that developed the Principal Odor Map—an AI model that predicts human odor perception directly from a molecule's chemical structure. This collection outlines his insights on the biology of olfaction, the application of graph neural networks to chemistry, and the future of scent-based disease detection.

Part 1: The Biology of Olfaction

1. On the brain's exposure: "That's the only part of the brain that leaves the skull and touches the air. Your brain is physically moving through the skull." — Source: Salon
2. On the physical nature of scent: "There's nerves that go through porous holes and become our sense of smell, and so molecules physically are touching your brain." — Source: Salon
3. On early fascinations: "Even as a 12-year-old, he was obsessed with perfumes and the invisible chemical structures that dictate our perception of the world." — Source: What's Your Problem?
4. On evolutionary importance: "Smell is an ancient sense, fundamentally wired into our brain's architecture long before language or advanced reasoning evolved." — Source: Brave New World
5. On olfactory mapping in animals: "Animals navigate their environments through complex odor gradients, a biological reality that we are only just beginning to model computationally." — Source: Theory and Practice
6. On metabolic processes: "The way we perceive scents is deeply connected to how our bodies metabolize biological compounds across different species." — Source: Mind & Matter
7. On the divide in research: "There is a historical divide between those who study the chemistry of aroma and those who study the neuroscience of olfaction." — Source: Mind & Matter
8. On memory and smell: "Because the olfactory bulb connects directly to the amygdala and hippocampus, scent provides the most immediate biological trigger for memory and emotion." — Source: Eye On A.I.
9. On the complexity of odors: "Unlike color, which can be mapped on a simple spectrum of wavelengths, natural odors are high-dimensional mixtures of thousands of different molecules." — Source: The Guardian
10. On sensory processing: "The biological process of translating a chemical binding event in the nose into a subjective perception like floral or musky remains one of the brain's great computational feats." — Source: Brave New World

Part 2: The Limits of Current Science

1. On the lack of textbooks: "I took callipers and measured the width of the paper that's used to teach vision and hearing... It's maybe 30 pages – a few millimetres – for smell." — Source: The Guardian
2. On scientific neglect: "They have these beautiful intellectual structures, these cathedrals of knowledge... that explain the visual and auditory world, shaming what we know about olfaction." — Source: Quanta Magazine
3. On data collection: "Cameras and microphones are cheap and pervasive... Collecting data for scent either involves using complex and expensive instruments or slow and arduous training processes." — Source: Osmo
4. On visual bias: "Human scientific progress has historically favored vision and hearing because they map easily to physics via measurable wavelengths and frequencies." — Source: Data Masters
5. On chemical complexity: "A single molecule's shape, size, and flexibility all interact simultaneously to produce a smell, defying simple linear rules." — Source: Science
6. On the vocabulary gap: "We lack a robust, standardized language for describing smells, often relying on comparisons like 'smells like a rose' rather than fundamental properties." — Source: Quanta Magazine
7. On the limits of sensors: "Hardware sensors for scent have historically been brittle, expensive, and narrow, failing to capture the broad spectrum of human olfaction." — Source: The Neuron
8. On historical stagnation: "For decades, the perfume and chemical industries relied mostly on trial, error, and human intuition rather than predictive computational models." — Source: Glossy
9. On the difficulty of digitization: "Digitizing smell is considered exponentially harder than digitizing sight or sound because it involves capturing and synthesizing physical matter rather than just capturing energy." — Source: Hard Fork
10. On data silos: "Much of the existing data mapping molecules to scents has been locked away in proprietary databases of legacy flavor and fragrance houses." — Source: Invest Like the Best

Part 3: Graph Neural Networks and Smell

1. On representing molecules: "Graph neural networks are the natural architecture for molecular chemistry because they perfectly represent atoms as nodes and chemical bonds as edges." — Source: Distill
2. On machine intuition: "By passing messages between nodes in a graph, GNNs can learn the implicit rules of chemistry that dictate how a molecule interacts with a human receptor." — Source: Google Research
3. On superhuman accuracy: "The GNN models trained at Google Brain were eventually able to predict the odor profile of a molecule more reliably than the average individual human panelist." — Source: Science
4. On model interpretability: "Understanding attribution in GNNs—knowing exactly which carbon ring or bond triggers a specific prediction—is crucial for designing new molecules safely." — Source: Distill
5. On scaling data: "Feeding the Leffingwell dataset and other large chemical libraries into GNNs provided the necessary volume to successfully map the latent space of odor." — Source: Google Research
6. On structure-odor relationships: "The models proved that the quantitative structure-odor relationship is not random; it has an underlying geometric structure that artificial intelligence can decode." — Source: eLife
7. On continuous learning: "As more molecules are synthesized and tested, the graph networks continually refine their understanding of the boundaries between descriptions like fruity and sulfurous." — Source: Science
8. On algorithmic architecture: "Message-passing frameworks allow the network to evaluate a molecule not just by its individual parts, but by the holistic shape it takes in three dimensions." — Source: Distill
9. On cross-disciplinary leverage: "The same GNN techniques used for mapping human olfaction can be adapted for broader tasks in drug discovery and materials science." — Source: Theory and Practice
10. On validation: "The ultimate test of the GNN was generating completely novel molecular structures in software and verifying that human testers agreed with the AI's predicted scent." — Source: Science

Part 4: The Principal Odor Map

1. On creating a compass: "The Principal Odor Map places molecules in a high-dimensional space where physical distance correlates precisely with perceptual similarity." — Source: Osmo
2. On the RGB of smell: "The map attempts to find the fundamental coordinates of olfaction, akin to how RGB values define the entire space of visible light." — Source: Invest Like the Best
3. On predicting the unknown: "The true power of the Principal Odor Map is its ability to take a molecule that has never been synthesized and accurately predict how it will smell." — Source: Science
4. On biological alignment: "The structure of the map naturally aligns with evolutionary biology, grouping smells in ways that reflect metabolic pathways found in nature." — Source: Quanta Magazine
5. On mapping complexity: "The model successfully disentangles molecules that look chemically identical but smell wildly different due to chirality or slight structural tweaks." — Source: Science
6. On continuous space: "Smells do not exist in discrete categories; the map shows that concepts like garlic and onion occupy a continuous region of molecular space." — Source: Osmo
7. On industry adoption: "The map provides a foundational tool for chemists to navigate the universe of possible scents mathematically rather than relying strictly on trial and error." — Source: Glossy
8. On open questions: "Even with a highly accurate map, the exact biophysics of how receptors inside the human nose trigger these mapped perceptions is still being uncovered." — Source: Mind & Matter
9. On its legacy: "The Principal Odor Map represents the first time human odor perception has been comprehensively quantified and digitized at scale." — Source: Google Research

Part 5: Digitizing and Teleporting Scent

1. On teleportation: "Osmo has demonstrated the ability to teleport a scent by capturing the molecular signature of a fresh plum and reproducing it remotely." — Source: The Neuron
2. On the reading-writing loop: "Digitizing scent requires two distinct technologies: sensors capable of reading the molecules in the air, and printers capable of writing or synthesizing them on demand." — Source: Invest Like the Best
3. On digital capture: "A molecular sensor acts as a camera for smell, analyzing the chemical composition of the air in real-time to create a digital fingerprint." — Source: Brave New World
4. On the molecular printer: "The output device for digital olfaction isn't a speaker or a screen, but a micro-fluidic system that mixes base chemicals to recreate exact odors." — Source: The Neuron
5. On consumer hardware: "The long-term vision involves shrinking molecular sensors down to the size of a microchip so they can be integrated into smartphones and wearables." — Source: Hard Fork
6. On archiving smells: "Digitization allows us to permanently archive the scent of endangered flowers, historical artifacts, and changing environments before they disappear." — Source: Invest Like the Best
7. On data transmission: "Once a smell is converted into coordinates on the Principal Odor Map, it becomes a lightweight data file that can be sent via email or text message." — Source: Axios
8. On environmental context: "Teleporting a scent perfectly requires capturing not just the primary odorant, but the ambient background molecules that contextualize the experience." — Source: Brave New World
9. On the ultimate goal: "The objective is to make sending a scent over the internet as mundane and reliable as sending a digital photograph." — Source: Osmo

Part 6: Transforming the Fragrance Industry

1. On expanding the palette: "There've only been about 100,000 fragrances ever made. I want there to be millions." — Source: Reddit
2. On AI augmentation: "We see AI augmenting the role of synthetic chemists and master perfumers, rather than replacing the human art of scent creation." — Source: Cosmetics Business
3. On democratizing creation: "Through AI-driven initiatives, independent brands can now create custom fragrances without the massive minimum chemical volume orders demanded by legacy houses." — Source: Glossy
4. On sustainability: "Digital olfaction allows chemists to discover sustainable, lab-grown alternatives to rare, endangered, or environmentally damaging natural scent ingredients." — Source: Perfumer & Flavorist
5. On safe chemistry: "AI models can filter out molecules that are toxic, allergenic, or harmful to the ozone layer before a chemist ever attempts to synthesize them." — Source: Osmo
6. On speed to market: "By predicting how a molecule will smell before it is made, artificial intelligence dramatically accelerates the research and development cycle for new consumer goods." — Source: Invest Like the Best
7. On proprietary ingredients: "Osmo has successfully developed entirely new scent molecules that did not previously exist in nature, opening new categories of fragrance." — Source: TrendHunter
8. On subjective art: "Even with perfect chemical prediction, the composition of a successful perfume still requires the subjective, emotional intuition of a human perfumer." — Source: Brave New World
9. On market disruption: "AI is breaking the oligopoly of the traditional fragrance industry by bringing technological transparency to a historically secretive trade." — Source: Independent Beauty

Part 7: AI in Molecular Chemistry

1. On the drug discovery parallel: "We see AI in drug discovery as a precedent for ways that AI will revolutionize olfaction." — Source: Cosmetics Business
2. On traversing chemical space: "The universe of possible molecules is larger than the number of atoms in the solar system; AI is the only tool capable of navigating it effectively." — Source: Theory and Practice
3. On generative design: "Machine learning models can now work backwards, taking a desired scent profile and generating the exact molecular structure needed to produce it." — Source: Science
4. On synthesis constraints: "It is not enough to design a novel molecule in software; the AI must also ensure that the molecule can actually be synthesized physically in a lab." — Source: Osmo
5. On edge cases: "Neural networks excel at identifying the exceptions to traditional chemical rules, finding rare molecules that smell completely different than their structural neighbors." — Source: Quanta Magazine
6. On computational efficiency: "Evaluating a million compounds via a graph neural network takes a fraction of a second, replacing years of manual bench chemistry." — Source: Google Research
7. On data feedback loops: "Every physical molecule synthesized is tested and fed back into the model, creating a compounding advantage in predictive accuracy over time." — Source: Invest Like the Best
8. On bridging disciplines: "The most significant breakthroughs in digital olfaction occur when deep learning engineers sit in the same room as wet-lab synthetic chemists." — Source: Data Masters
9. On fundamental physics: "Ultimately, predicting a smell requires the AI to implicitly learn the quantum and thermodynamic properties that govern molecular binding." — Source: Distill

Part 8: The Future of Health and Disease Detection

1. On diagnostic potential: "The long-term moonshot for digital olfaction is developing hardware sensors capable of detecting diseases through microscopic changes in human body odor." — Source: The Neuron
2. On canine inspiration: "Dogs can already smell Parkinson's, cancer, and malaria; the goal is to build digital sensors that match or exceed this biological capability." — Source: Apple Podcasts
3. On non-invasive testing: "A breathalyzer or skin-swab that uses AI to detect disease signatures would transform diagnostics by making clinical testing entirely non-invasive." — Source: Tamr
4. On biological markers: "Diseases alter human metabolism, which in turn alters the specific volatile organic compounds we emit into the air." — Source: Mind & Matter
5. On continuous monitoring: "If molecular sensors become small and cheap enough, they could be embedded in our homes to monitor health changes continuously over time." — Source: Invest Like the Best
6. On early intervention: "Scent signatures often change long before clinical symptoms appear, offering a vital window for much earlier therapeutic intervention." — Source: Brave New World
7. On signal extraction: "The primary challenge in disease detection is using AI to separate the faint signal of a disease biomarker from the noisy background of everyday environmental smells." — Source: Theory and Practice
8. On global health: "Cheap, deployable scent sensors could provide rapid screening for infectious diseases like tuberculosis in regions without robust clinical infrastructure." — Source: Axios
9. On the ultimate vision: "Digitizing smell is not just about making better perfumes; it is about building a new sensory layer for computers to understand and protect human health." — Source: Osmo