Alex Wiltschko

The through-line

Everything Wiltschko says circles one problem: smell is the last major sense that computing never learned to handle, and the missing piece was never the hardware but the representation. He frames the task in three steps, read the world, map it, write it back out, and is explicit that the map was the gap: sound has frequency, color has RGB, and scent had nothing comparable [2]. The work at Google Brain was aimed squarely at that gap, and the company built since then treats a learned odor map as the foundation for everything else, reading scents in and printing them back out [2][5]. He is candid that the full vision is decades of work, and that this is a feature rather than a problem, because "the mission that we're on is bigger than any one person and it's longer than any one life. And so that makes it worth working on" [1].

What has shifted is the centre of gravity from research to commerce. Having crossed what he considers the real threshold, a smell going into a mass spectrometer and coming out of a scent printer with no human in the loop, he describes the science as finally "mature enough to build a business", which is why Osmo spawned Generation as a fragrance house while Osmo itself keeps advancing the R&D [1].

On what digitizing smell actually means

The three steps are read, map, write, and he insists all three have to work end to end [2]. The demonstrated version is unglamorous and concrete: put a smell into a gas chromatograph mass spectrometer roughly the size of a dishwasher, let Osmo's software analyse the output with no human intervention, turn it into a coordinate on the scent map, decode that coordinate into instructions for a formulation robot, and print the smell out again, fast and repeatably [1]. He dates the first success to roughly half a year before the conversation and says it is now routine and generating data continuously [1]. The same capability has been shown as teleporting a scent across a building [4], and the longer arc is a company teaching computers to both read and write scent [5]. The endpoint he describes is miniaturisation: a handheld device that can "smell as well as a dog's nose can" or a high-end mass spectrometer, and an emitter better than a flower, since a flower produces one scent for its whole life while a device could produce a different scent every second [1].

On the principal odor map

The hard scientific core was the structure odor relation problem, predicting what a molecule smells like from its structure, which he says "had been unsolved for a hundred years and actually in some cases people thought it was unsolvable" [2]. The approach was to gather thousands of molecule-to-descriptor pairs and train a graph neural network, an architecture specialised for chemistry and closely related to the transformer, where nodes are atoms and edges are bonds [2]. Odorous molecules are conveniently small, typically under twenty atoms, because anything larger neither evaporates into the air nor fits the binding pockets of olfactory receptors [2]. Cracking the trained network open and reading its embedding layer gave the map, which needed around 300 dimensions to work well, a number he calls suspicious given the receptor count without claiming causation [2].

The structure inside that embedding is what he finds beautiful. Plotting about 5,000 molecules in a two-dimensional PCA shadow, the smell categories form neighbourhoods rather than scattering: floral occupies a large region, and jasmine, rose and violet fall out as sub-regions nested inside it, a hierarchy nobody supplied [2]. The fermented and alcoholic region came out "shaped like a bottle during our first model train and we haven't touched it since because it's so funny" [2]. His working interpretation, still unresolved, is that the map partly recapitulates biology: fermented notes cluster because yeast makes them, floral notes cluster because genetically related plants make them, so the geometry of scent traces the story of how living things produce molecules [2].

On how good the human nose really is

He pushes back hard on the assumption that humans are poor smellers. The eye has roughly four channels through rods and cones; the nose has over 300 receptor types, each differently sensitive to the chemical world, making olfaction far higher dimensional in channel count even though what those channels code for remains a mystery [2]. Olfactory sensory neurons are frontline cells that begin in the brain and poke through the skull, one of only two parts of the brain that leaves it, so when you smell something "your brain is physically touching another living thing" that has let off a piece of itself [2]. On sensitivity he cites the mercaptans added to natural gas, detectable at parts per billion or trillion, and Noam Sobel's experiment showing blindfolded people can scent-track a trail of chocolate or cinnamon like a dog, slowly but genuinely [2]. His verdict: "We're freaking amazing at smelling" [2]. He is equally willing to say what is not known, dismissing the famous trillion-odour estimate as thoroughly debunked and declining to guess at olfactory resolution [2]. He compares his own pre-surgery eyesight to explain the difference between not perceiving at all and merely failing to resolve subtle differences [2].

On chemistry as another intelligence

The broadest claim he makes for the field is that AI has learned almost entirely from digital modalities, text, images, audio, video, while the problems people want it to solve live in the physical world [2]. Scent is where that gap bites hardest, because "99% of species on this planet can only speak with chemistry", bacteria, fungi, plants and insects communicating exclusively through molecules released for reasons [2]. Training models on that chemical output is, in his framing, a way of adding alien forms of intelligence to AI [2]. The downstream applications he points to run from medicine and mood to a bridge with traditional practices like Ayurveda and aromatherapy [2], and further to counterfeit detection and health monitoring [5].

On designing molecules that do not exist

Osmo's first line of business was generating genuinely new fragrance molecules that smell good, are safe, scale in production and stay affordable [2]. The commercial logic is regulatory as much as creative: molecules get withdrawn, and replacements are needed if laundry and homes are to keep smelling the way people expect [2]. Targets are specific rather than vague, a citrus note that lasts longer, a vanilla note that is optically clear because vanilla is normally brown and brands want to control product colour [2]. Confidence in predicting novel molecules came from an odour Turing test with Joel Mainland at Monell: predict the smell of never-before-made molecules, keep predictions secret, train a panel for about eight hours on roughly fifty descriptors, and ask whether you would rather "add another person or would you rather add the predictions of a model" [2]. The model beat the average individual panellist [2].

On rebuilding a fragrance house from first principles

Generation began as a deliberately cheap experiment in October: a website under a different name, no advertising, an open invitation to send a brief [1]. The volume of briefs proved a latent market, so "we basically shut the experiment down because it was a success", followed by three or four months of standing up a factory, supply chain, account executives and operations [1]. The customer pain he heard was consistent: brands unable to reach fragrance houses at all, or receiving a library fragrance instead of something bespoke, or waiting six to eighteen months [1][5]. Generation targets small and medium businesses, roughly zero to $250 million in revenue, anywhere in the world, and has extended into turnkey manufacturing because brands asked for bottling and packaging too, which is marginal extra work given an existing factory, fragrance robot, compounding team and supply chain [1].

The mindset he brings is startup, but with a stated condition: it "requires is a very, very healthy and deep respect for the heritage of the industry", which has existed in roughly its current form for 300 years and contains a great deal worth preserving [1]. The rest has simply not been re-examined, so the guiding thought experiment is "If it's Givaudan or Firmenich or IFF were starting today, how would they do it?" [1]. His answer is to listen to customers closely and build solutions with them, and the branding follows the same instinct, fusing the natural with the modern [1].

On curiosity, hiring and leadership

Curiosity is his stated first principle, both personally and as a hiring filter: "there are few things more powerful than curiosity", the childhood insistence on asking why fused with adult capability to investigate it [1]. He grew up in a household where "there was nothing more important in my household than learning", with a small allowance for sweets but no limit on books [1]. Books remain his model of leverage, delivering "the distilled experience of a person who spent a significant chunk of their life on something" for ten or twenty dollars, and he extends the same logic to podcasts [1]. Luca Turin and Tanya Sanchez's guide he treats as a work of art whose reviews read almost as poems, and he notes he disagrees with Turin about Jean-Claude Ellena's minimalism, which he loves [1]. When curiosity fails, his trick is to redirect it inward: be curious about the boredom itself, "a judo flip on yourself" [1].

Hiring runs on a gut check phrased in chemistry: is the relationship endothermic or exothermic, and is he "leaving a conversation with more energy than I came in with" [1]. He is honest that he has never audited that instinct against data, but says ignored red and yellow flags usually trace back to it [1]. Energy is necessary and not sufficient; skills and cultural fit follow, with three ranked values: curiosity first, tenacity second, kindness third, where kindness means honesty, honour and the ability to collaborate [1]. Leadership he defines without reference to title, as "helping communities around you define and pursue and achieve a common goal", something that can happen on an apartment block, in a religious community or in an online forum, and should happen in more places [1]. On the personal cost of a hard mission his line is that "bravery isn't the absence of fear, it's functioning, through fear" [1].

Takeaways

• Giving computers a sense of smell breaks into three steps, read, map and write, and the map was the piece that did not exist before the principal odor map [2].
• A graph neural network trained on molecule-to-descriptor pairs cracked the century-old structure odor relation problem, and its roughly 300-dimensional embedding became the map used commercially since [2].
• The map's geometry appears to encode biology, with jasmine, rose and violet nesting inside a floral region and fermented notes clustering because yeast produces them [2].
• Osmo's demonstrated milestone is a smell going into a gas chromatograph mass spectrometer and out of a formulation robot with no human intervention, now run routinely to generate data [1].
• Generation was validated by an unadvertised website that collected enough briefs to prove demand, then shut down deliberately: "we basically shut the experiment down because it was a success" [1].
• The strategic question driving Generation is what Givaudan, Firmenich or IFF would build if they started in 2025 and simply listened to customers [1].
• Hiring starts with an energy test, is this conversation exothermic, then requires curiosity, tenacity and kindness in that order plus the skills for the job [1].
• Humans are far better smellers than assumed, detecting mercaptans at parts per billion and able to scent-track a trail blindfolded [2].