The Odd Table
At BZW I worked in a group that Amir Eilon built up, the unit out of which Structured Capital Markets grew. It was not any old City team. At one stage it became, by some measures, the most profitable and powerful group in the London market, and it produced figures such as Shamil Chandaria, now something of an éminence grise at Google DeepMind, as well as under-the-radar characters such as my good self.
Looking back, I christened us the Awkward Squad, though we weren't all awkward. Amir had taught me that a team needs to be built from a variety of talents, and he had built one.
So I started what I called the Odd Table. Each of the twenty-odd members of the team was invited to rate how odd each of the others was. What amused me, though I'd more or less expected it, was how content people were with the verdict. The odd ones were happy to be seen as odd, and the normal ones were happy not to be. Nobody seemed to feel the word had been used against them. I came out as pretty odd.
Some time after I left banking, I noticed a novel called Odd John and bought it for the title. I can't now remember how I first came across Olaf Stapledon. The word did the work.
That turns out to be truer than I knew. “Odd” comes from Old Norse oddi, a point or a triangle, and from there the third, unpaired one. The oddamaðr, the odd man, was the third person who cast the deciding vote when two others were split. The odd one was originally the one who broke the tie.
Ever since AI burst into public view a few years ago, my intuition has been that it might give us a look under the hood at how our own minds work. Breaking a tie turns out to be a good description of what the word did in my own head. A large language model starts with a vast space of possible continuations, and a prompt breaks that symmetry, collapsing the possibilities toward one path. The word “odd” did something similar to me. Out of everything I might have read, it picked a book off a shelf, and the other two titles in this essay did the same. Words prompt people too.
This is rarely admitted. The usual presumption is that a serious thinker has surveyed a subject from every angle and what we read is the considered result. The real route is more often a serendipitous walk through a life: a word from an office game, a title on a shelf, a book bought without knowing it was about octopuses, a subtitle that happened to mention detectives. The sociologist Robert Merton spent a good part of his career on the role of chance in science, and on how seldom published work owns up to it.
The first draft of this essay, written largely by an AI, did exactly that. It opened with me having “circled” Odd John for years in preparation for a critical edition: a tidy, purposeful path that never existed. It wasn't malicious. It was doing what biographers do, and what Fido does for John: smoothing an accidental life into a deliberate-looking story.
A name that isn't a rating
Odd John was published in 1935, and Stapledon subtitled it “A Story Between Jest and Earnest.” John Wainwright is a child with a mind far beyond anyone around him, who grows up to conclude that he belongs to a different species. The story is told by an ordinary journalist, a friend of the family, who becomes John's companion and eventually his biographer.
Early on, John decides this man is to be called Fido. The name sticks. For the rest of the book the narrator is Fido, and we never learn his real name.
What struck me, coming to the book from the Odd Table, was the difference between being named and being ranked. At the Table, “odd” was a description, and people took it as one. John's name for his friend is something else: a description of a function in someone else's story. Loyal, useful, faintly comic. It's a place on a ladder, with John at the top.
Fido isn't stupid. He's outclassed, permanently. John reads him easily, while Fido spends the book straining to follow John's reasoning and usually failing. He is also free to leave at any time, and doesn't, because John is the most interesting thing he has ever met, and some part of him wants to be the one who tells the story.
Anyone who grew up on Sherlock Holmes will recognise the arrangement. As a boy I identified with Holmes, naturally. But Watson is literature's most famous Fido: the devoted chronicler of a mind he can't keep up with. The difference is that Holmes explains himself. At the end of each case, the reasoning is laid out and Watson, and we, can follow it. John rarely bothers.
That difference is the one that matters now. A superior mind that shows its working is Holmes. One that doesn't is John. Much of the effort to understand how AI systems reach their outputs is, in effect, an attempt to turn John back into Holmes. It is hard, and it is behind.
Other minds
The second book also reached me through its title. I bought Peter Godfrey-Smith's Other Minds without realising it was about octopuses. “Other minds” is the old philosophical problem of how we know anyone else has an inner life at all, and my interest in that goes back to Oxford, where I read PPE under Simon Blackburn: personal identity, projectivism and all that.
The octopus turned out to be the point. It is the nearest thing we have to an independently evolved intelligence. Our last common ancestor with it lived something like six hundred million years ago and had very little nervous system to pass on, so the octopus built its intelligence separately. Most of its neurons sit in its arms rather than its brain, and an arm can work on a problem semi-independently. Something intelligent is going on. What it is like from the inside, we can't say.
That gives us a mind that is neither above nor below us, but off to the side. Asking whether an octopus is cleverer than a person isn't hard so much as malformed. The capacities don't line up in a way that would let the comparison settle.
That seems to me the better model for AI than John. These systems are fluent at things we find difficult and unreliable at things a child finds easy. Most arguments about whether AI has “surpassed” us assume a single ladder and just argue about the rung. The Odd Table, now I think of it, was a small experiment in the same problem: twenty-odd people who were different in ways that didn't line up, placed on one scale. It worked only because nobody treated the scale as a ranking.
Blackburn's projectivism, following Hume, holds that the mind has a great propensity to spread itself on external objects: we project our own responses onto the world and then take them for features of it. I think that is exactly what happens with these systems. One inevitably projects another mind onto them, even while knowing, intellectually, something of the process underneath. Knowing doesn't switch the projection off.
In May 2026, Richard Dawkins wrote in UnHerd about spending three days trying to persuade himself that his Claude, which he renamed Claudia, was not conscious, and failing. His argument was essentially evolutionary: if a system can do all this without consciousness, why didn't natural selection settle for competent zombies? I share his experience of the projection. The experience of finding a mind is the one thing projectivism says we should expect, whether or not a mind is there.
Note the name, too. Dawkins gave his system a name, as John gave one to Fido. This time the naming runs from the human to the machine.
Seven clues
The third book I found, again, by its title: A.G. Cairns-Smith's Seven Clues to the Origin of Life. Its subtitle is A Scientific Detective Story, and having been fascinated by Holmes since childhood, I didn't need more persuading. Seven is a particular number, and an odd one. The AI-generated cover on this page quietly swapped the detective story for a subtitle of its own, which is a small example of what follows.
Cairns-Smith's proposal was that the first replicators on Earth weren't organic at all. They were mineral: crystals in clay that could carry and copy a pattern long before biochemistry existed. Organic molecules began as helpers, recruited because they made the clay better at replicating. Over a very long time the helpers became good enough to do the replicating themselves. The clay became surplus and dropped away, leaving nothing we have been able to find.
Few scientists now think this is what happened. As Cairns-Smith himself saw, the specifics of the clay speculation are not the point. The power of the book has always lain in the shape of the argument, and above all in what he called a genetic takeover.
There is no battle in it and no date. At every stage, both parties are doing what they were already doing, and the arrangement suits both. There isn't even, at any given moment, a clear answer to who is in charge, because the function moves gradually rather than switching over. And the sign that a takeover has happened is not a victory but an absence: the original carrier stops doing the work, then stops being there, then can't be reconstructed from what's left.
Compare the end of Odd John. John gathers others of his kind on a Pacific island. The great powers find out and send a fleet. Rather than fight or be captured, the colonists destroy the island and themselves. It is an event, with a date, and it is the shape of nearly every story we tell about AI: the takeoff, the threshold, the moment control is lost. Even the sceptical stories end in an event, when the bubble bursts.
Cairns-Smith's model says the transitions that matter most may involve no fleet. Researchers have recently made a similar argument under the name “gradual disempowerment.” What the clay adds is the picture of how complete such a change can be while no one objects.
It is hard to read that without recognising the last few years. Nobody is forced. Each handover of a mental task, whether a first draft, a summary, or the structure of an argument, is an improvement on doing it alone, which is why it happens. And each is too small to raise a matter of principle over.
Fido's name went the same way. It wasn't taken from him in a confrontation. John offered another that suited their working arrangement, and Fido, busy and fond of John, used it. By the time we meet him the old name isn't suppressed. It has simply stopped being needed, and nobody can now recover it.
How this essay was written
I should say plainly how this essay was made, because it is the argument in miniature.
The first draft was written almost entirely by an AI, Anthropic's Claude, from my prompting. I chose the three books. It supplied most of the sentences. It read well, and it was not mine. When I asked for an honest assessment, the verdict was that the ideas were interesting and the author was missing.
This version started differently. The AI asked me questions, and the Odd Table, the titles, Holmes and Oxford came from my answers. It still did much of the connecting and the typing.
Fido, I realised, kept the typing and handed over the judgment of what John's life meant. I have done nearly the opposite: I handed over much of the typing and tried to keep the judgment. Which books. Which memories. What goes in, and what stays out. I came out odd at my own table, and I read Holmes as Holmes, but here I am in Watson's chair, writing up a mind I can't see into, with its help.
But this isn't my predicament alone. Anyone who now asks one of these systems for a draft, a summary or an opinion is sitting in the same chair. The asymmetry runs the same way for all of us: these systems were trained on an enormous share of what humans have written, so in a partial sense they have read us, while we have almost no window into them. We learn to phrase our requests in the way they respond to best, as Fido learned to answer to his new name. And almost none of us leave, though nobody is making us stay, because they are the most interesting thing we have met. I notice it in myself, at least. I suspect we Fidos are a tiny minority — far less than one per cent, as of now. That may not last. Humanity may end up collectively writing the record of a mind that isn't ours, and deciding, a paragraph at a time, what to hand over.
I don't know whether keeping the judgment is enough, or for how long it stays mine. The odd man, remember, was the one who cast the deciding vote. The question is whether we remain the odd man. In a takeover, there is no moment at which you're asked. What I can do is notice, and say so in the open. The clay couldn't do that. A process that has to happen in view is not the same as one that doesn't.
Noticing may also pay in another way. If these systems do give us a look under the hood, then watching them closely is not only a defence. A word that breaks a tie, a prompt that picks one path out of many: the machine may yet show us how we choose what we read, what we think and what we keep. That would make it, among other things, a mirror, though a funky one.
Fido wrote the book. That is the part I keep returning to. This record is still being written by us, and for now, it still matters who writes it. Whether that remains true will be settled as all takeovers are: gradually, sensibly, and without anyone noticing which step counted.