What the Monkey Couldn’t Type

Yesterday, researchers at Lund University¹ reported in Science Advances² that an ancient virus may help explain a difference between human and chimpanzee brains. We share many genes with chimpanzees, inherited from common ancestors who lived five to seven million years ago. In human brain tissue grown in the lab, one of those genes, LINC00662, is »switched on«. A gene is switched on when a cell reads it, off when it doesn’t, and this particular one seems to help young nerve cells grow their connecting fibres. In chimpanzee tissue grown the same way, that gene is switched off.

»A difference doesn’t have to be something that’s present. It can be something that was prevented«

The reason seems to be a virus. After our lineages parted, a virus called PTERV1 spread through the chimpanzees’ ancestors, writing its genetic code into their DNA. When that happens in egg or sperm cells, the code is inherited, and PTERV1 left around 150 copies. One landed inside the chimpanzee version of LINC00662, but the human version has no such copy. When the team used gene editing to cut that copy out of chimpanzee cells, the gene switched back on. In other words, the difference between our brains may have been shaped by something that happened to one lineage and not the other.

Explaining a difference by an off switch sounds a bit like explaining a bagel by its hole. And yet in 1975, geneticist Mary-Claire King and biologist Allan Wilson found that human and chimpanzee proteins are strikingly alike, and proposed that much of what separates us lies in how genes are regulated.³ A difference doesn’t have to be something that’s present. It can be something that was prevented.

The virus is long dead. What it left behind is a damaged copy of its DNA, which can probably no longer make new viruses. So how can a dead copy silence a gene? The answer seems to be the cell’s own defences. Cells guard against viral DNA by attaching chemical tags that stop it being read. The authors suggest the tags spread from the viral copy to the gene’s on-switch, at the front of the gene, and jammed it, though they haven’t yet shown exactly how. Why humans were spared the virus is still open: the ancestors of humans may have lived elsewhere, or been able to resist it. Either way, the damaged copy is still inherited, and each generation’s cells silence it again, and the gene with it. The virus is gone, but the restraint it left on the gene remains.

In Thursday’s essay, we borrowed anthropologist Gregory Bateson’s definition of information: a difference that makes a difference, meaning a signal is information only if it changes something for its recipient.⁴ Bateson’s 1967 essay »Cybernetic Explanation« turns to explanation.⁵ Usually, he writes, we explain by naming what hit what: one billiard ball moves because another struck it. But Bateson turns this around: list everything that could have happened, then ask what kept all the other possibilities from happening. His example is a monkey hitting the keys of a typewriter and producing a page of meaningful text. Perhaps the monkey can’t strike the wrong keys, or the machine jams if it does, or wrong letters are wiped from the page; somewhere, he says, there must be a circuit that spots errors and deletes them. Bateson calls these restraints, and says that’s where the explanation lives. A restraint doesn’t need to outright forbid anything: whatever makes some outcomes likelier than others counts. In the physical world, a wall forbids you from going somewhere while a slope makes one direction easier than the other, and both are restraints.

But most things that don’t happen also don’t explain anything. What makes a restraint count is that it can be measured, and Bateson measures it. In an English message, a letter might have been any of twenty-six, so the letter K rules out twenty-five alternatives. The more a thing rules out, the more information it carries. And English isn’t flat: some letters are more common than others, and certain pairs turn up together more often. That’s a slope, not a wall – and it can be measured.

Bateson also writes about a jigsaw puzzle: each piece’s shape, colour and edges restrain which piece can go where. To the person solving the puzzle, these are clues, information that guides the choice. To an observer asking why the puzzle came out as it did, they are restraints. The facts are the same; what they are depends on where you sit.

The same structure runs inside the large language models behind chatbots. Those models score every possible next word, and then equations turn the scores into odds: a slope that tilts them towards likelier words. To make sure the output is valid, engineers add a mask that sets forbidden words to zero chance before the model’s preferences operate on those that remain. A wall with a slope inside it. Bateson’s monkey needed something that spots errors and deletes them; in a language model, one version of that something is a mask.

So what does it mean to explain something by what didn’t happen? It means finding what held the other possibilities back. In Lund, the researchers could go further: they lifted the restraint, and the gene came back. The chimpanzee’s gene is silent because something happened to it; ours may work because something didn’t.

Few restraints can be lifted in a lab. Bateson’s recipe starts with a list of everything that could have happened, and someone has to draw up that list. In politics it is called the agenda: the political scientist E. E. Schattschneider wrote in 1960 that »organization is the mobilization of bias«, with some issues organised into politics and others organised out.⁶ Some restraints are so old that we mistake them for nature, others are so ordinary that we mistake them for choice.

References