Why AI Needs a Sense of Smell
Vision, language and sound each had a digital format waiting for them. Smell has none - no wheel, no frequency, no map - and an intelligence built only from the senses that were easy to digitise is missing the oldest one.
Read the original ↗An essay about a sense no model has, and about why that is not an oversight. Sight, language and hearing each arrived at AI already written down as numbers. Smell never has been, so the only thing a model can learn from is the words people use about it, and the essay's hardest claim is that those words are caricatures. The claim is contested by a result the essay does not mention.
Included with a plan
This discussion is included with a plan. The first three lessons of every track are free, with their boards and videos. A plan unlocks every lesson, board, video, lab and paper for one price.
Related on chalk
The same idea as a board, lab, video or lesson.
- BoardTokens and Embeddingsan embedding space is a map made from data - the essay's question is what you do for a sense that has no labels to learn from
- BoardTeaching It To Seethe sense that was solved first, and why: pixels were already a grid
- BoardWhy the Sky Is Bluecolour has a wavelength, which is why a camera exists - the contrast the essay is built on
- BoardEvery Sound Is Sine Wavessound has a frequency, so a microphone was always going to work - smell has no such axis
- BoardHow Memory Workswhy a smell brings a memory back whole, which is the essay's case for why the sense matters
- BoardConsciousness and the Hard Problemwhat it is like to smell something, and whether a map of molecules could ever contain that
- LabTokens and embeddingsrun a sentence through the map language got - then ask what the equivalent input would be for a molecule
- VideoVoice, image and video: diffusion and the other modalitiesthe senses a model has today, and how each one gets turned into tokens
- VideoCNNs: how a machine learned to seeconvolution, the trick that made vision fall - built on a property smell does not have
- VideoFourier: every sound you have ever heard is sine waves stackedthe frequency axis that makes sound a digitisable thing
- LessonTokens & EmbeddingsUses the board Tokens and Embeddings
- LessonLooking Inside the ModelUses the board Tokens and Embeddings
- LessonThe MLP Language ModelUses the board Tokens and Embeddings
- LessonMultimodality (teaching an LLM to see)Uses the board Teaching It To See
- LessonConvolutional NetworksUses the board Teaching It To See
- LessonWhy the Sky Is BlueUses the board Why the Sky Is Blue
- LessonEvery Sound Is Sine WavesUses the board Every Sound Is Sine Waves
- LessonHow Memory WorksUses the board How Memory Works
- LessonThe Wet HardwareUses the board How Memory Works
- LessonThe Words You Think InUses the board How Memory Works
- LessonConsciousness & the Hard ProblemUses the board Consciousness and the Hard Problem
Further reading
Read the original for the full argument and its evidence.
- Why AI Needs a Sense of Smell ↗Philip Maughan in Noema, 16 April 2026. The reported scenes are enjoyable and the argument is in one section, There Is No Map. Read that one twice.
- A Principal Odor Map Unifies Diverse Tasks in Olfactory Perception ↗Lee et al., Science, September 2023, free full text. The strongest evidence against the essay's central claim, and the paper it never mentions. Check the 53% figure in the results before repeating the abstract's summary of it.
- From Molecules to Mixtures ↗Tom et al., January 2025, with the odor map's lead author among them. Says in its own opening that mixture representations remain under-explored, which is the essay's best paragraph confirmed by the other side.
- Osmo announces a $70M Series B ↗February 2026, two months before the essay ran. Worth reading beside Wiltschko's quoted mission to see how much of digitising smell is now a fragrance supply chain.