Before the argument, a test.

Five questions. No chemistry required, and being a chemist will not help you much. Each one is a real, settled fact about a real molecule. The point of this is not to catch you out; it is that the essay is much more convincing after you have failed it.

If you scored well, you have read a lot of perfumery. You still did not predict anything: you remembered it. That distinction is the whole essay.

We do not even agree how many smells there are.

For most of the twentieth century the received figure was that a human being can distinguish around ten thousand odours. It appeared in textbooks, in Nobel documentation, in every popular article on the subject. Nobody could quite say where it came from.

In 2014 a group at Rockefeller tested it properly, using mixtures of thirty components, and calculated that the real number is at least one trillion. The figure went around the world. Then two separate critiques argued the mathematics did not support the claim, that the same method applied to vision or hearing would produce equally absurd numbers, and that the honest answer is that the question as posed may not have a determinate answer at all.

Hold that in mind, because it sets the tone for everything below. This is a field where the most basic quantitative question, how many things can this sense tell apart, is still argued over in the pages of major journals. Compare that to vision, where nobody is unsure how many colours there are.

Lock and key, and where the key stops working.

The orthodox account is shape. An odorant molecule reaches a receptor protein in the nose, and if the shape and chemistry fit, the receptor fires. This is the same lock-and-key logic that runs most of biochemistry, and it is broadly correct.

It is also, as a predictive tool, weak. The evidence is a set of well-known embarrassments that every perfumery chemist can recite.

Nothing has its own receptor.

In 1991 Linda Buck and Richard Axel found the gene family behind olfaction, work that took the Nobel Prize in 2004. Roughly a thousand genes, about three percent of all our genes, of which around four hundred are still functional in humans. The rest have decayed, which is a quiet statement about what our species stopped needing.

The important finding was not the count. It was the logic. Each olfactory neuron expresses one receptor type. Each odorant activates several different receptors, weakly or strongly. Each receptor responds to many different odorants. So no molecule has a dedicated detector, and identity is carried by the pattern across the whole array, the way a word is carried by letters that individually mean nothing.

This explains a great deal. It explains why smells blend rather than stack, since two odorants that share receptors produce a pattern belonging to neither. It explains why concentration changes identity, because at higher doses more receptors recruit and the pattern is literally a different pattern. And it explains why prediction is so hard: to predict a smell from a structure you would need to predict roughly four hundred binding affinities at once, and then know how the brain reads that array. We can barely do the first half for one receptor.

The man who said the nose is a spectrometer.

In the 1990s the biophysicist Luca Turin, who is also one of the most gifted perfume writers alive, proposed something else. Not shape, or not only shape: vibration. Chemical bonds vibrate at characteristic frequencies, and Turin argued the receptor works by inelastic electron tunnelling, effectively reading a molecule's vibrational spectrum. It would explain the convergence problem elegantly, since unrelated structures sharing a vibrational signature would share a smell.

The theory made a testable prediction, which is to its credit. Replace hydrogen with deuterium and you change the vibrational frequencies while leaving the shape essentially untouched. If vibration matters, the deuterated version should smell different.

In 2004 Andreas Keller and Leslie Vosshall at Rockefeller ran that test on untrained subjects with acetophenone and its deuterated twin. The subjects could not reliably tell them apart. In 2013 Turin and colleagues reported that people could distinguish deuterated musks. In 2015 a group led by Eric Block tested the mechanism directly at the receptor level, on a human musk receptor and a mouse thiol receptor, found no vibrational discrimination, and argued the earlier result was confounded by an impurity. Turin's group replied. The exchange continues, and the mainstream position is that vibration theory is implausible.

And yet, in 2023, something started predicting anyway.

Here is where the title of this essay has to be qualified, honestly and at length, because something significant happened recently.

A team led by Alexander Wiltschko, working initially at Google and later at the company Osmo, trained a message-passing graph neural network on a combined industry dataset of a few thousand molecules labelled with professional odour descriptors. Rather than predicting labels directly, the model learned an internal representation, which they called a principal odor map: a space in which nearness means smelling alike.

Then they did the thing that makes it real science. They took four hundred molecules the model had never seen, had a trained human panel describe them, and compared. Published in Science in 2023, the result was that the model's odour profile matched the panel mean more closely than the median individual panelist did. They went on to map the predicted odour of around half a million molecules that no human has ever smelled.

So it is now fair to say that a machine can predict how a trained panel will describe a novel molecule, better than a typical member of that panel. That is a genuine and important result, and anybody writing about this subject who leaves it out is writing propaganda.