Video · 2026-07-08 · 59m · 30 moments

Teaching Computers to Smell | Alex Wiltschko

✦ AI generated

timeline · colored by role

Claim◆

AI models should incorporate 'alien' forms of intelligence beyond human-generated text and images — specifically the chemical language used by the 99% of species (bacteria, fungi, plants, insects) that can only communicate through molecules — by training systems directly on that chemistry, echoing Terry Tao's call for a 'Copernican view of intelligence' that stops centering human cognition.

Alex argues AI should move beyond a human-centric view and incorporate 'alien' intelligences like the chemical communication used by most species on Earth, echoing mathematician Terry Tao's call for a 'Copernican view of intelligence.'

transcript

Alex Wiltschko: 99% of species on this planet can only speak with chemistry, right? Thinking of bacteria and fungi and plants and insects, they just only can talk with molecules. And I think that it's really worth adding other kinds of alien forms of intelligence to our AI models. And the way to do that is to train it on the intellectual output of those other intellects, which is chemistry.

Claim◆

AI should be trained on the chemical 'language' that 99% of species — bacteria, fungi, plants, insects — use to communicate, treating it as an alien form of intelligence worth adding to our models.

Alex Wiltschko argues that since most life on Earth communicates only through chemistry, AI should be trained on that 'alien' chemical intelligence, not just human-generated text and images.

transcript

Alex Wiltschko: 99% of species on this planet can only speak with chemistry, right? Thinking of bacteria and fungi and and plants and insects, like they just only can talk with molecules. And I think that it's really worth adding other kind of alien forms of intelligence to our AI models. And the way to do that is to train it on the intellectual output of those other intellects, which is that's chemistry.

Mechanism

Digitizing any sense requires three steps: reading it from the physical world into data, mapping that data into a structured, manipulable representation, and writing it back out; for smell, the map has been the missing piece.

Wiltschko frames sensory digitization (as done for vision and sound) as a three-part problem—read, map, write—and explains that Osmo's early work at Google Brain targeted the missing 'map' for smell.

transcript

Alex Wiltschko: Any scent that's been given to computers, there's three kind of broad steps. You got to read the world. So, like turn atoms into bits and information. you have to map it, like understand it. So, you know, be able to manipulate it digitally, encode it, send it, and that's like JPEG and RGB, right? Um, and then you have to be able to write it back out again, right? So, a printer or a display or a speaker.

Mechanism

Digitizing any sense requires three steps — reading (turning physical signal into data), mapping (a representation you can manipulate/encode), and writing (reproducing it) — and for smell, the map has historically been the missing piece.

Wiltschko frames the problem of computerizing smell using the same read/map/write framework used for vision (RGB) and sound (frequency), explaining that Osmo's original focus at Google Brain was building the missing 'map' for scent.

transcript

Mechanism

Giving computers a sense of smell requires the same three steps used to digitize vision and hearing — reading (turning atoms into bits), mapping (encoding into a manipulable representation like RGB/JPEG), and writing (outputting via a printer, display, or speaker) — and smell has always been missing the 'map' step.

Wiltschko lays out the read-map-write framework used for vision and audio and explains that Osmo's core contribution was building the missing 'map' for scent.

transcript

Mechanism

Giving computers a sense of smell requires three steps — reading (digitizing scent into data), mapping (building a representational space for it), and writing (synthesizing scent back out) — and the map has been the missing piece that other modalities like vision and sound already had.

Alex frames the olfactory-AI problem using the same read/map/write pipeline that enabled digital color and sound, arguing scent's missing piece has always been the map.

transcript

Alex Wiltschko: Any scent that's been given to computers, there's three kind of broad steps. You got to read the world. So, like turn atoms into bits and information... and then you have to be able to write it back out again, right? So, a printer or a display or a speaker. Um, and so the thing we focused on at Google Brain was the missing piece, which is for scent is the map.

gives example · 3

Definition

Giving computers a sense of smell requires three steps: reading the world (turning atoms into bits), mapping it (encoding and manipulating it digitally, the way JPEG and RGB do for images), and writing it back out (through something like a printer or speaker) — and the piece that's been missing for scent, which Osmo focused on, is the map.

Alex frames the problem of computerizing smell as analogous to color and sound digitization, breaking it into read/map/write stages, and identifies the missing 'map' as Osmo's founding technical focus.

transcript

Alex Wiltschko: Any scent that's been given to computers, there's three kind of broad steps. You got to read the world. So, like turn atoms into bits and information. You have to map it, like understand it... and then you have to be able to write it back out again... the thing we focused on at Google Brain was the missing piece, which is for scent is the map.

gives example · 1

Fact

The human eye has roughly three channels of color information, but the human nose has over 300 types of olfactory receptors, making smell a far higher-dimensional sense than vision, even though what those channels encode is still not fully understood.

Alex contrasts the ~3 color channels in human vision with the 300+ olfactory receptor types in the nose, underscoring how much higher-dimensional smell is as a sensory input.

transcript

Alex Wiltschko: There's three channels of color information roughly roughly RGB uh in the eye, but there's over 300 channels of alactory information in the nose. And it's still a mystery exactly what they code for, but it's certainly much higher dimensional, at least in terms of channel count.

Fact

The human eye has roughly three channels of color information (RGB), while the human nose has over 300 different olfactory receptor types, making smell a far higher-dimensional sensory channel than vision, even though it's still unclear exactly what each receptor codes for.

Alex explains that smell is encoded by over 300 types of olfactory receptors in the nose, compared to roughly three color channels in the eye, making it a much higher-dimensional signal.

transcript

Alex Wiltschko: There's three channels of color information roughly RGB in the eye, but there's over 300 channels of olfactory information in the nose. And it's still a mystery exactly what they code for, but it's certainly much higher dimensional, at least in terms of channel count.

Fact

The human nose has over 300 distinct olfactory receptor channels coding for smell, far more than the roughly three or four channels used for color vision.

Wiltschko notes that while color vision relies on roughly three or four channels (RGB plus grayscale), smell is coded by over 300 distinct olfactory receptor types, making it far higher-dimensional even though what each channel encodes remains unclear.

transcript

Data◆

A graph neural network trained on structure-to-odor data predicted the smell of never-before-smelled molecules more accurately on average than an individual human panelist, effectively passing an 'odor Turing test.'

In a double-blind 'odor Turing test,' Osmo's graph neural network predicted the smells of brand-new, never-before-smelled molecules and outperformed the average individual human panelist, solving a structure-odor relationship problem that had stood unsolved for a century.

transcript

Alex Wiltschko: it turned out our model predictions were better than any one individual panelist on average in the panel meaning we've passed a not during test like our model predictions were human quality which was pretty cool.

Data◆

A graph neural network trained to predict odor from molecular structure produced predictions rated as good as, or better than, individual trained human panelists in a blind 'odor Turing test' on never-before-smelled molecules.

Osmo ran a double-blind 'odor Turing test' comparing model predictions of novel molecules' smells against trained human panelists, and the model beat the average individual panelist.

transcript

Alex Wiltschko: Let's go find molecules that nobody's ever smelled before. Some have never been made before, like nature has not seen them. Let's predict what they smell like ahead of time and keep our predictions secret. Let's go get those molecules physically, send them to another location.

Data

A graph neural network trained on molecular structure predicted novel, never-smelled molecules' odor descriptions well enough to beat the average individual human panelist, passing what the team called an 'odor Turing test.'

Osmo's graph neural network, trained on structure-to-odor pairs, predicted the smell of brand-new molecules and outperformed the average individual on a trained human sensory panel.

transcript

Alex Wiltschko: the odor touring test is if you want to make your panel better would you rather add another person or would you rather add the predictions of a model right um and it turned out our model predictions were better than any one individual panelist on average in the panel meaning we've passed a not during test like our model predictions were human quality which was pretty cool.

Data◆

A graph neural network trained on thousands of molecule-to-odor-descriptor pairs predicted what novel, never-before-smelled molecules would smell like well enough that, in a blind 'odor Turing test' against trained human panelists, its predictions were rated as good as or better than any individual human panelist's average performance.

Osmo's graph neural network, trained on structure-to-odor data, passed a blind 'odor Turing test' by predicting the smell of new molecules as well as or better than individual trained human panelists.

transcript

Alex Wiltschko: The odor Turing test is if you want to make your panel better would you rather add another person or would you rather add the predictions of a model. And it turned out our model predictions were better than any one individual panelist on average in the panel, meaning we've passed a Turing test — like our model predictions were human quality.

explains mechanism · 1

Data

A graph neural network trained to predict odor descriptors from molecular structure performed an 'odor Turing test' against novel, never-smelled molecules and its predictions were more accurate on average than any individual human panelist, matching human-quality smell prediction.

Osmo's neural network predicted the smell of never-before-smelled molecules and beat the average individual human panelist, passing what Alex calls an odor Turing test.

transcript

Alex Wiltschko: The question is basically you know the odor touring test is if you want to make your panel better would you rather add another person or would you rather add the predictions of a model right um and it turned out our model predictions were better than any one individual panelist on average in the panel meaning we've passed a not during test.

explains mechanism · 2

Data

The learned 'principal odor map' embedding organizes molecules so that perceptually related scents form nested sub-clusters (e.g., jasmine, rose, and violet nested inside a broader floral region) without ever being explicitly told this hierarchical structure exists.

When Osmo visualized their learned odor embedding space, semantically related scent categories clustered into nested neighborhoods (florals containing jasmine/rose/violet) that the model was never told to form.

transcript

Alex Wiltschko: if you draw like jasmine or rose or um you know, violet, they actually ended up being sub regions inside of floral. And we didn't tell it that like we didn't tell it that there was this nested relationship of of nature there. And similarly, the the region for um for fermented and alcoholic was actually shaped like a bottle during our first model train and we haven't touched it since because it's so funny.

provides context · 1

Example◆

The learned 'principal odor map' embedding spontaneously organizes molecules into nested perceptual neighborhoods — for example, jasmine, rose, and violet form sub-regions inside a larger floral region — without ever being told this hierarchy exists.

When Osmo visualized its odor embedding space, molecules naturally clustered into coherent, nested scent neighborhoods (like sub-florals inside a floral region) that mirrored real olfactory categories nobody had labeled.

transcript

Alex Wiltschko: if we circled the floral region, it was this pretty big part on the left... this the floral region was pretty big and was on a side of the left. But then if you draw like jasmine or rose or um you know, violet, they actually ended up being sub regions inside of floral. And we didn't tell it that like we didn't tell it that there was this nested relationship of of of nature there.

Example

When the smell-prediction model's molecule embeddings were plotted in 2D, scents formed coherent, nested neighborhoods — e.g., jasmine, rose, and violet emerged as sub-clusters within a larger floral region — without ever being told such relationships existed.

Visualizing the 'principal odor map' via PCA, Osmo found scent categories self-organized into clean, nested clusters (florals containing jasmine/rose/violet sub-regions, fermented scents even shaped like a bottle) that emerged purely from the data, without explicit supervision.

transcript

Alex Wiltschko: But then if you draw like jasmine or rose or um you know, violet, they actually ended up being sub regions inside of floral. And we didn't tell it that like we didn't tell it that there was this nested relationship of of of nature there.

gives example · 2

Example◆

When Osmo's learned odor embedding space was visualized in two dimensions, molecules with the same human-labeled smell clustered together, and finer distinctions like jasmine, rose, and violet emerged as nested sub-regions inside the broader floral cluster without ever being told that hierarchy existed.

Alex describes how the principal odor map's embedding space self-organized into perceptually meaningful, nested neighborhoods (e.g., jasmine/rose/violet inside 'floral') purely from data, mirroring biological reality.

transcript

Alex Wiltschko: If we circled the floral region, it was this pretty big part on the left... But then if you draw like jasmine or rose or um you know, violet, they actually ended up being sub regions inside of floral. And we didn't tell it that... there was this nested relationship of of nature there.

Example

When Osmo's molecular smell-embedding space is visualized in two dimensions, molecules that humans labeled with the same scent form coherent neighborhoods, and those neighborhoods nest in ways that mirror nature — for example, jasmine, rose, and violet form sub-regions inside a larger floral cluster — even though the model was never told this hierarchical relationship exists.

Osmo's 'principal odor map' clusters molecules with similar smells into neighborhoods that mirror natural scent taxonomy — e.g., jasmine, rose, and violet nest inside a larger floral region — without being explicitly told that structure exists.

transcript

Alex Wiltschko: If we circled the floral region, it was this pretty big part on the left. But then if you draw jasmine or rose or violet, they actually ended up being sub regions inside of floral. And we didn't tell it that there was this nested relationship of nature there.

Data

Osmo has built what is likely the largest olfactory dataset ever assembled for AI training — 6 billion enumerated candidate molecules and 5.43 million human sniff labels — created entirely from scratch, since no equivalent of Scale AI or Mechanical Turk exists for smell.

Osmo has digitized 6 billion candidate molecules and 5.43 million human sniff labels, building from scratch what is likely the largest olfactory AI training dataset in existence, since no outsourced labeling service for smell exists.

transcript

Alex Wiltschko: We've digitized 5.43 million sniffs. So that's like the largest olfactory data set for the purposes of training AI models, I think, ever. And that's we had to make all of that from scratch, right? Like there's no Scale AI or there's no Mechanical Turk for smell.

Data

Osmo has built what is likely the largest olfactory dataset ever created — 5.43 million labeled human sniffs plus 6 billion digitized candidate molecules — constructed entirely from scratch because no crowdsourcing platform for smell exists.

Osmo has amassed 5.43 million human 'sniff' labels and digitized 6 billion possible molecules, building the data pipeline from scratch since no mechanical-turk-style service exists for smell.

transcript

Alex Wiltschko: I think I want to get the number right. It's probably shifted since I last looked this up like a week ago, but we've digitized 5.43 million sniffs. Um, so that's like the largest alactory data set for, you know, the purposes of training AI models, I think, ever. Um, and, uh, that's we had to make all of that from scratch, right? Like there's no scale AI or there's no mechanical Turk for smell.

Data

Osmo has digitized 6 billion candidate molecules and 5.43 million human sniffs, making it what Alex believes is the largest olfactory dataset ever built for training AI models.

Alex quantifies Osmo's data scale: billions of enumerated candidate molecules and millions of labeled human sniffs, calling it the largest olfactory AI dataset ever created.

transcript

Alex Wiltschko: We've digitized 5.43 million sniffs. Um, so that's like the largest alactory data set for, you know, the purposes of training AI models, I think, ever. Um, and, uh, that's we had to make all of that from scratch, right? Like there's no scale AI or there's no mechanical Turk for smell.

Data

Osmo built the largest olfactory dataset ever assembled from scratch, since no existing service like Scale AI or Mechanical Turk exists for smell data collection.

Because no crowdsourcing infrastructure exists for scent labeling, Osmo had to build its own pipeline internationally, resulting in what Wiltschko calls the largest olfactory dataset ever created for AI training.

transcript

Alex Wiltschko: we've digitized 5.43 million sniffs. Um, so that's like the largest alactory data set for, you know, the purposes of training AI models, I think, ever. Um, and, uh, that's we had to make all of that from scratch, right? Like there's no scale AI or there's no mechanical Turk for smell.

Prediction

Detecting diseases like cancer or malaria from scent will require assembling a massive, ground-up olfactory dataset by sniffing thousands of people and everyday objects, because you can never collect enough subjects to statistically learn the subtle, distributed molecular signal directly.

Wiltschko argues scent-based disease detection can't be cracked by collecting data only from sick versus healthy people, since sample sizes will never be large enough to isolate the subtle multi-molecule signal — instead it requires a massive, general-purpose olfactory data-collection effort across thousands of people and objects.

transcript

Alex Wiltschko: You can go directly after that and like collect uh scent data from people with or without those conditions and try to build models, you're never going to get enough people to build a great model, right? And like there's just it's just hard to go get that much data.

Claim

Building AI systems that can detect diseases like cancer or malaria via scent requires collecting massive, broad data (e.g., sniffing 10,000 people regardless of health status) rather than small targeted studies, because the disease signal is a subtle pattern across many molecules rather than one obvious marker.

Wiltschko argues that scent-based disease detection can't be cracked with small targeted studies; it requires massive indiscriminate data collection, akin to what animals like dogs can already do intuitively.

transcript

Alex Wiltschko: We just need to go get a ton of data, right? Like we need to like band together and build a huge effort where we just like let's go sniff 10,000 people. I don't care if they're healthy or sick. I don't care how old they are. I don't care.

Prediction

Detecting diseases like cancer or malaria through scent will require sniffing tens of thousands of people regardless of health status, because the disease signal isn't one obvious marker but subtle changes spread across hundreds or thousands of molecular signals — making it far harder data-collection work than scraping text from the internet.

Wiltschko argues that scent-based disease detection (cancer, malaria) demands massive, deliberately collected sniff datasets from thousands of people, since the signal is subtle and diffuse — much harder to obtain than scraping existing internet data.

transcript

Claim◆

AI should incorporate 'alien' forms of intelligence, like the chemical communication used by the 99% of species that can only speak through molecules, rather than centering artificial intelligence solely on human-derived output, paralleling a Copernican shift away from human-centered views of intelligence.

Invoking mathematician Terry Tao's 'Copernican view of intelligence,' Wiltschko argues AI should learn from non-human intellects like the chemical signaling used by most species on Earth, not just human-generated text and images.

transcript

Claim◆

AI models should be trained on the 'intellectual output' of non-human species — the chemistry they use to communicate — since 99% of species on Earth can only speak through chemistry, adding a genuinely alien form of intelligence to AI.

Wiltschko argues that because most life on Earth communicates only through chemistry, building AI that understands scent is a path to incorporating a genuinely non-human ('alien') intelligence into AI systems, echoing mathematician Terry Tao's call for a 'Copernican' view of intelligence.

transcript

Alex Wiltschko: 99% of species on this planet can only speak with chemistry, right? Thinking of bacteria and fungi and and plants and insects, like they just only can talk with molecules. And I I I think that it's really worth adding other kind of alien forms of intelligence to our AI models. And the way to do that is to train it on the intellectual output of those other intellects, which is that's chemistry. That's the sense that's in the air.

Claim◆

Because 99% of species on Earth communicate only through chemistry, training AI on olfactory/chemical data would add a genuinely alien, non-human form of intelligence into AI systems, and we should adopt a 'Copernican' view that removes human intelligence from the center of how we think about AI.

Alex argues that since almost all species communicate via chemistry, giving AI a sense of smell trains it on a genuinely alien form of intelligence, and invokes Terry Tao's call for a 'Copernican' view that decenters human intelligence in AI.

transcript

Highlight slides

AI Needs 'Alien' Intelligence, Not Just Human Data✦ from: AI models should incorporate 'alien' forms of intelligence beyond human-generated text and images — specifically the chemical language used by the 99% of species (bacteria, fungi, plants, insects) that can only communicate through molecules — by training systems directly on that chemistry, echoing Terry Tao's call for a 'Copernican view of intelligence' that stops centering human cognition.Most Life Speaks Only in Chemistry✦ from: AI should be trained on the chemical 'language' that 99% of species — bacteria, fungi, plants, insects — use to communicate, treating it as an alien form of intelligence worth adding to our models.Train AI on 'Alien' Chemical Intelligence✦ from: AI should be trained on the chemical 'language' that 99% of species — bacteria, fungi, plants, insects — use to communicate, treating it as an alien form of intelligence worth adding to our models.A Copernican View of Intelligence✦ from: AI models should incorporate 'alien' forms of intelligence beyond human-generated text and images — specifically the chemical language used by the 99% of species (bacteria, fungi, plants, insects) that can only communicate through molecules — by training systems directly on that chemistry, echoing Terry Tao's call for a 'Copernican view of intelligence' that stops centering human cognition.A Copernican View of Intelligence✦ from: AI models should incorporate 'alien' forms of intelligence beyond human-generated text and images — specifically the chemical language used by the 99% of species (bacteria, fungi, plants, insects) that can only communicate through molecules — by training systems directly on that chemistry, echoing Terry Tao's call for a 'Copernican view of intelligence' that stops centering human cognition.AI Predicts Smell Before Molecule Is Sniffed✦ from: A graph neural network trained to predict odor from molecular structure produced predictions rated as good as, or better than, individual trained human panelists in a blind 'odor Turing test' on never-before-smelled molecules.AI Passes the Odor Turing Test✦ from: A graph neural network trained on structure-to-odor data predicted the smell of never-before-smelled molecules more accurately on average than an individual human panelist, effectively passing an 'odor Turing test.'Why It Matters✦ from: A graph neural network trained on structure-to-odor data predicted the smell of never-before-smelled molecules more accurately on average than an individual human panelist, effectively passing an 'odor Turing test.'Blind Odor Turing Test: Model vs. Humans✦ from: A graph neural network trained to predict odor from molecular structure produced predictions rated as good as, or better than, individual trained human panelists in a blind 'odor Turing test' on never-before-smelled molecules.Osmo's AI Passes the Odor Turing Test✦ from: A graph neural network trained on thousands of molecule-to-odor-descriptor pairs predicted what novel, never-before-smelled molecules would smell like well enough that, in a blind 'odor Turing test' against trained human panelists, its predictions were rated as good as or better than any individual human panelist's average performance.Why It's a 'Turing Test' for Smell✦ from: A graph neural network trained on thousands of molecule-to-odor-descriptor pairs predicted what novel, never-before-smelled molecules would smell like well enough that, in a blind 'odor Turing test' against trained human panelists, its predictions were rated as good as or better than any individual human panelist's average performance.Odor Map Learns Hidden Scent Hierarchy✦ from: The learned 'principal odor map' embedding spontaneously organizes molecules into nested perceptual neighborhoods — for example, jasmine, rose, and violet form sub-regions inside a larger floral region — without ever being told this hierarchy exists.Nested Structure, Not Imposed✦ from: The learned 'principal odor map' embedding spontaneously organizes molecules into nested perceptual neighborhoods — for example, jasmine, rose, and violet form sub-regions inside a larger floral region — without ever being told this hierarchy exists.Odor Map Self-Organizes by Smell✦ from: When Osmo's learned odor embedding space was visualized in two dimensions, molecules with the same human-labeled smell clustered together, and finer distinctions like jasmine, rose, and violet emerged as nested sub-regions inside the broader floral cluster without ever being told that hierarchy existed.Nested Sub-Regions Inside 'Floral'✦ from: When Osmo's learned odor embedding space was visualized in two dimensions, molecules with the same human-labeled smell clustered together, and finer distinctions like jasmine, rose, and violet emerged as nested sub-regions inside the broader floral cluster without ever being told that hierarchy existed.AI Should Learn 'Alien' Intelligence, Not Just Human Output✦ from: AI should incorporate 'alien' forms of intelligence, like the chemical communication used by the 99% of species that can only speak through molecules, rather than centering artificial intelligence solely on human-derived output, paralleling a Copernican shift away from human-centered views of intelligence.Beyond Human-Centered AI✦ from: AI should incorporate 'alien' forms of intelligence, like the chemical communication used by the 99% of species that can only speak through molecules, rather than centering artificial intelligence solely on human-derived output, paralleling a Copernican shift away from human-centered views of intelligence.Chemistry: The Universal Language of Life✦ from: AI models should be trained on the 'intellectual output' of non-human species — the chemistry they use to communicate — since 99% of species on Earth can only speak through chemistry, adding a genuinely alien form of intelligence to AI.Training AI on Alien Intelligence✦ from: AI models should be trained on the 'intellectual output' of non-human species — the chemistry they use to communicate — since 99% of species on Earth can only speak through chemistry, adding a genuinely alien form of intelligence to AI.A Copernican View of Intelligence✦ from: AI models should be trained on the 'intellectual output' of non-human species — the chemistry they use to communicate — since 99% of species on Earth can only speak through chemistry, adding a genuinely alien form of intelligence to AI.Chemistry Is Life's Universal Language✦ from: Because 99% of species on Earth communicate only through chemistry, training AI on olfactory/chemical data would add a genuinely alien, non-human form of intelligence into AI systems, and we should adopt a 'Copernican' view that removes human intelligence from the center of how we think about AI.Training AI on an Alien Intelligence✦ from: Because 99% of species on Earth communicate only through chemistry, training AI on olfactory/chemical data would add a genuinely alien, non-human form of intelligence into AI systems, and we should adopt a 'Copernican' view that removes human intelligence from the center of how we think about AI.A Copernican Shift for AI✦ from: Because 99% of species on Earth communicate only through chemistry, training AI on olfactory/chemical data would add a genuinely alien, non-human form of intelligence into AI systems, and we should adopt a 'Copernican' view that removes human intelligence from the center of how we think about AI.

Related episodes

Video · a16z Podcast

Ben Horowitz on the Global Race for Tech, Power, and Influence

→

Video · Lenny's Podcast

The rise of taste, human authenticity and judgment in an AI world | Adam Mosseri (Head of IG)

→

Audio · All-In Podcast

SpaceX-Cursor Deal, SaaS Debt Bomb, New Apple CEO, SPLC Indictment, Colon Cancer Spike

→