Alex Wiltschko - Giving Computers A Sense Of Smell - [Invest Like the Best, EP.415]
Summary
Osmo says it crossed smell’s zero-to-one by digitally reading a fresh summer plum and reconstituting it as “actually a freaking plum.” Its system pairs chemical sensors with an approximately 300-dimensional scent map and a writer that mixes molecules back into an experience. A successful read-write round trip can create an active-learning loop in which deliberately selected new smells make later experiments more informative.
The defensible asset is a vertically integrated data factory for a modality lacking AI-compatible data. Twice-daily ranked human panels label odors, GC-MS machines dissect samples 24/7, chemists synthesize OI-designed molecules, and Osmo stores 10,000-20,000 AI-designed compounds with physical-to-digital twins. Wiltschko’s framing: “Most of the iceberg” is creating and linking the data.
Generation turns that scientific platform into a near-term fragrance business by collapsing a 12-18-month custom-development process toward minutes. Traditional buyers often wait weeks or months only to receive an existing library scent; each revision can take three months, while raw fragrance prices range from roughly $10 to many hundreds of dollars per kilo. Osmo wants every result customized through conversational briefs, its scent map, AI, and human perfumers.
Fragrance offers unusually attractive incumbent economics, but the market’s structure pushes Osmo toward vertical integration rather than pure software. Fragrance appears in 90% of household products, category shifts can make suppliers relatively recession-resistant, repurchase rates are “well above 50%,” and secret formulation know-how creates “manufacturing businesses with non-manufacturing margins.” Osmo tried selling software, but with relatively few fragrance houses and buyers, Wiltschko concluded it may have to compete directly.
Machine smell already has a concrete authentication wedge: an Osmo sensor can classify StockX shoes as real or fake within 20 seconds. Counterfeiters can reproduce visual identity, but a shoe’s odor is “the fingerprint of everything that ever happened to make this”; each new SKU requires some real and fake examples, though Wiltschko acknowledges an emerging arms race.
The larger option value lies in portable chemical sensing for provenance, border safety, and eventually human health. Wiltschko argues that substances in blood and organs emerge through breath and sweat, creating a largely untapped signal analogous to what dogs detect. The larger reader must shrink another 4-8× to become genuinely portable, while the scent printer remains roughly half the size of the table.
Osmo’s existential risk is that hardware limits, cost, or an unknown fundamental constraint interrupts the roadmap before the system becomes personal, portable, and affordable. Rather than attack that summit directly, Wiltschko wants “a shallow enough grade where at some points we can stop and build a business,” making the company harder to kill with each commercial foothold. His time horizon is explicit: “The exit strategy is death.”
Deep dive
1. Smell becomes computable only when machines can both read and write it
Osmo’s data engine begins with twice-daily sensory panels whose members literally sniff and label samples. The company maintains “literal rankings,” reserving its “top dogs” for work requiring the most accurate judgments—an olfactory equivalent of human image-labeling infrastructure that Wiltschko said could not simply be purchased.
Its GC-MS machines act as “a camera for the molecular world.” A robotic loader operates 24/7; a scent travels through a 50-meter column that separates light and heavy molecules like runners finishing a race, before an electron gun fragments them and the system infers their identities from fragment weights and travel times.
That final interpretation traditionally combines software and human judgment. Osmo says its olfactory-intelligence system performs it entirely through software: Wiltschko’s strategy is not to replace hardware containing “about 12 Nobel Prizes’ worth” of advances, but to “rip out the brains” and supply the missing software maps between instruments.
The underlying alphabet remains unsolved: nobody knows the equivalent of primary colors or a periodic table for odor. Scientists may know maple syrup’s molecular contents, but producing “maple syrup, but with a little bit more cherry”—or replacing one molecule for safety—requires tradecraft that Osmo is trying to automate.
2. The read-write loop turns every experiment into training data
Patrick distilled the architecture into reading and writing, with writing serving as the test of whether reading was correct. Wiltschko agreed: once new smells can be created, measured, and selected to maximize what tomorrow’s experiments teach, the system enters an active-learning loop and “get[s] smart really fast.”
The higher-end reader pulls scent directly from flowers, fruit, vegetables, people, or other objects. Airborne molecules collect on a film that absorbs scent “like Kodak film absorbs light”; heat then “develops the film,” releasing the sample into the spectrometer for AI analysis.
A gas chromatograph olfactometer can also pause a complex scent and let researchers sniff its 30 or 100 component molecules individually. Wiltschko called it “a debugger for software”—a tool for building human intuition alongside the machine’s molecular decomposition.
Fidelity ultimately returns to perception: people compare the real object with the recreation and answer, effectively, “We good?” Osmo uses de-biasing methods to generate machine-readable judgments, but Wiltschko’s concession was categorical: for a familiar memory, the person smelling it remains “the arbiter.”
3. A summer plum proved the round trip was no longer theoretical
Osmo’s first fully teleported scent began with a fresh purple summer plum—the kind with “a snap when you bite into it.” The team sliced it, analyzed its headspace, and reprinted the smell elsewhere in the laboratory; the resulting vial probably contains thousands or tens of thousands of sniffs of that moment.
Patrick closed his eyes, smelled Plum 1.0, and recognized it. For Wiltschko, that vial holds “the source code” of a scent memory and marks the company’s defining moment: the reconstruction was “beautiful and almost hyperreal,” proving the project had gone “from zero to one.”
Earlier theater concepts such as AromaRama failed because they emitted a few complete, pre-programmed scene smells. Wiltschko compared that approach to a slideshow; a true scent display must mix a broad ingredient palette on demand so it can “show anything,” not merely replay eight stored experiences.
4. Wiltschko’s path joined perfume obsession to graph neural networks
Wiltschko began programming around age eight or nine and collecting perfume at 12, fascinated that an invisible substance could change how people were treated within a small radius. Bvlgari Black revealed fragrance as art: over roughly 45 minutes, screeching tires and rubber gave way to vanilla on leather, then smoky tobacco—a “movie” deliberately composed to unfold in three acts.
Neuroscience led him to a century-old question: given a molecule’s structure, can anyone predict whether it smells like apple, cinnamon, or anise? After academia, two AI startups, Twitter’s deep-learning team, and Google Brain, he revisited the problem when graph neural networks made arbitrary molecular structures tractable—“chocolate and peanut butter for AI and chemistry.”
His Google team predicted the smell of hundreds of thousands of molecules, selected 400 that were very different-looking from anything it had seen before, and kept the answers hidden while a trained panel rated the real compounds. The model performed better than the average panelist; Wiltschko said that, if adding one more evaluator, he would prefer the software’s prediction to training another human nose—“passing an odor Turing test.”
After the scientific milestones, Josh Wolfe’s decade-long interest in digitizing olfaction connected with the project through an introduction at GV. Wolfe was integral to extracting the intellectual property from Google Brain and building the company; he led the round, Krishna Yeshwant at GV co-led it, and Osmo was formed.
5. Generation compresses fragrance development from months to minutes
Traditional development begins with a free-form brief describing the brand, desired smell, benchmark fragrance, expected volume, and price—anywhere from roughly $10 to many hundreds of dollars per kilo. The fragrance house receives no upfront payment and first decides whether the opportunity is worth pursuing.
Wiltschko estimates that 90% of the time the house initially sends something already created for another client. A buyer insisting on true customization can face a 12-18-month process, with three months between rounds of feedback, followed by application testing to ensure the scent does not discolor a cream or change inside the final formulation.
Generation combines Osmo’s technology with perfumers and professional noses to modernize each stage. The brief should become an immediately available ChatGPT-style conversation; software can map the concept, while people bring the emotional and aesthetic judgment required to make something “straight up beautiful.”
In Patrick’s live Colossus brief, themes of life’s work, possibility, open air, and redwood forests became “Ambition Trail.” Osmo embedded the language into an approximately 300-dimensional scent map—not three-dimensional RGB—located it relative to 100 mass-market hits, decoded it into a scent profile and source code, and condensed what Wiltschko called a multi-month task into minutes.
6. Industry structure makes vertical integration more plausible than SaaS
Patrick pressed on whether Osmo should remain a platform for developers or own applications such as Generation. Wiltschko’s answer depended on buyer readiness: vibrant markets can support enabling platforms, but fragrance has relatively few houses and buyers, and Osmo’s attempts to sell them software did not produce an obvious path.
His analogy was Metropolis: software could make parking lots more efficient, but operators were not ready buyers, so the company became a parking-lot business. Osmo likewise sees an opportunity to enter and transform an industry whose workflow, Wiltschko argued, has not substantially changed for 300 years.
The incumbent economics are compelling. Because fragrance appears in 90% of household products, weakness in luxury perfume may be offset by hand soap or another category; proprietary formulation know-how produces “manufacturing businesses with non-manufacturing margins,” and industry repurchase rates sit well above 50%.
Wiltschko does not want to discard practices that have survived for centuries, but he sees unmet demand for speed, safety, transparency, and access. Hotels such as the Ritz-Carlton—and Patrick’s memory of the Gramercy Park Hotel—show how a proprietary scent can create instant familiarity and bind a physical place to a brand.
7. Reading scent opens authentication first and health later
For StockX, Patrick compared visually identical real and counterfeit shoes but could not reliably distinguish them by smell. Wiltschko explained that visual appearance can be copied while odor records materials and manufacturing history: “the smell of the shoe is basically the fingerprint of everything that ever happened to make this.”
Osmo reduced its laboratory sensing stack to roughly the volume of two shoeboxes. A shoebox thumb hole connects to the sniffer, which returns real or fake within 20 seconds after training on examples; each new SKU requires collecting fresh authentic and counterfeit samples, sometimes more and sometimes fewer.
Patrick imagined counterfeiters adding their own perfumes, but Wiltschko said the arms race already exists and scent will become its next frontier. The broader provenance thesis includes other counterfeit goods, harmful substances at borders, and eventually computer systems that assist or replace tasks now performed by sniffer dogs.
Human health is the “holy grail,” though Wiltschko kept the claim prospective. Blood and organ chemistry eventually reaches breath and sweat, sometimes allowing people to notice that a loved one is getting sick before they do; dogs can detect related signals. Moreover, the molecular overlap among fruits, flowers, products, and human scent means progress in fragrance may transfer into health sensing.
8. Proprietary physical data is the fuel—and nature remains the constraint
Patrick contrasted Osmo with AI applications wrapped around language models, noting ChatGPT’s stated 400 million monthly users, roughly 5% of humanity. Wiltschko argued text is near the right side of its S-curve—“there’s one internet, and we’ve trained on it”—while scent remains near the far-left takeoff point.
Osmo uses LLMs and other AI tools wherever appropriate, but refuses to force every problem into one architecture. AI touches nearly all its code, raises productivity, and probably makes the company smaller; ChatGPT can supply an 80/20 review of an NDA, though critical questions still receive an informed human opinion.
The harder work began when licensed scent data proved incompatible with AI. Osmo therefore rebuilt the image-AI ecosystem for smell: people labeling all day, sensors dissecting samples 24/7, robots producing odors, chemists making novel molecules, and 10,000-20,000 AI-designed compounds, each with a digital twin and a known physical location.
Wiltschko’s persistent fear is that “Mother Nature” blocks the next advance in 2025, 2026, or 2027. The larger reader still needs to shrink 4-8× for portability, and the writer is roughly half the size of the table; his response is to climb toward personal, affordable digitized smell by routes where Osmo can repeatedly “stop and build a business” and become harder to kill.