A decade ago, I was hard at work on my third book, with the working title Strategy Evolves: From Apes to AI.1 I had the two core ideas - that strategic intelligence has evolved under the selection pressure of war. And second, that AI was emerging as a new form of intelligence which might be very different. Now all that remained was to write it up.
Ten years later seems a good point to see how the AI bit has fared.
In 2016 there were no transformers: they arrived the next year, with LLMs following soon thereafter. Everyone on LinkedIn was talking feverishly about ‘big data’ and the blockchain, whatever that was. Still, things were certainly stirring in AI. I was enthralled by Google’s 2012 cat detection algorithm, and by the earlier inhuman acrobatics of Stanford’s autonomous helicopter. Lee Sedol was falling to AlphaGo, even as I wrote.
Back then, I was reading a lot of evolutionary psychology, and thinking especially hard about the evolution of ‘theory of mind’, where we humans try to figure out what other humans are up to. Did ToM become our species’ USP? And if so, why? I had a candidate: war.
Perhaps, I argued, there was enough war in human pre-history to select for a particular type of intelligence. I thought chronic warfare was only one factor among several shaping us - childrearing and hunting were other options I mentioned. Later on, I was struck by Richard Wrangham’s argument that intra-group policing of violent males would have been in the mix too. Still, war seemed important, with lots of suggestive data that it wasn’t a rarity.
ToM was strategic: learning who to cooperate with, building a reputation and standing, and sharing information about enemies and allies. The gossip theory of language, which Robin Dunbar wrote about so persuasively, turbo-charged it all, separating us from other social animals with similar problems.
Human strategy in war, I argued boiled down to two big things - this dance of minds (‘I know that you know…’ &etc); and second, related, the ability to coordinate violence - especially concentrating it in time and space, allowing overwhelming force to be levelled against numerically inferior enemies. Raid and ambush over pitched battle, if it could at all be helped.2
Much of our cognition, biases included, were at root social - in the service of good enough judgments. I saw war lurking everywhere behind these. Our strategic minds might help us thrive within our groups, and help our group survive against others. Our ability to do this constituted the one big revolution in military affairs - an entirely cognitive one. Not even nuclear weapons altered that essential property, I claimed, illustrating the argument with Kennedy and Khrushchev’s desperate attempts to read one another’s intention during the Cuban crisis.
But what about AI? I ended the book with two chapters, one on narrow AI the other on general intelligence, or AGI, a term then much less familiar than today. On tactical systems, I argued:
that its decisions would reflect fundamentally different processes to ours, and these might well surprise us.
that tactical AI might permit the assembly of large offensive formations, able to identify targets and then rapidly concentrate against them, dispersing before counterattack.
that legacy systems might be overmatched by the pace of large autonomous swarms (I was thinking physical swarms, but hello Hugging Face hack!).
that these changes problematise defence, both through denial and the threat of punishment.
that these tactical changes might have profound strategic impact - shaping decisions about using force, and about who had hard power. It seemed to me a recipe for offensive realism, with a strong security dilemma and arms racing; and for some intense pressure to bandwagon with leading AI powers. I thought the impact on fighting power was potentially so great as to reshape prevailing strategic cultures.
On AGI, things were hazier. I thought that AGI might face similar challenges to humans, and perhaps be shaped by that selective pressure in similar ways. Specifically:
there was the challenge of understanding other agents, which might produce theory of mind, or something like (I’ve spent the last few years investigating that - seems to check out in some ways, less in others).
and there might be some element of natural selection, where the AGI could develop its own internal motivations, even if its master-goals were initially given. Today we’re on the cusp of recursive self-improvement and this sort of argument seems a bit less wild than in 2016.
Perhaps machine consciousness might be a feature of AGI - or something similar, given that in humans it plays a big part in attending to important things, especially modelling other minds, and presenting ourselves to them. I wasn’t convinced, then or now, because consciousness seems to me a bandwidth issue in biological intelligence that need not apply with machines. Right now, LLMs capture some dimensions of consciousness (persistent memory, focus on other minds, limited introspection) but not others (embodiment and affect) — I suspect we’ll need a new concept for what’s emerging.
What might all that mean for strategy? I foresaw an AGI able to apply force across a broad spectrum, in contrast to tactical systems; and to do so via a very different logic to humans, in the detail, if not in the general need to weigh consequences and intentions. The space for humans would be compressed, given the efficiency logic inherent in warfare. This was, I perceived, revolutionary - at least as much as the evolution of that earlier strategic intelligence. Still, I concluded, optimistically, ‘perhaps these speculations about a war-fighting AGI won’t come to pass in our lifetimes, and I am being unduly concerned.’
Well, perhaps. Anyway - that’s all ten years ago. What will the next ten bring?
Way better than the one they eventually went with, Strategy, Evolution, and War: From Apes to Artificial Intelligence. Buy now, buy often!
An aside: I was struck by Lanchester’s mathematical laws of conflict, particularly the bit where fighting power increased as a square of mass, not linearly; provided, that is, the combatants were functionally identical, as they would basically have been with the weapons on offer. No matter how skilled the isolated warrior, he could be swamped by superior numbers. Strategy became about taking advantage of this rule in the offence, and denying it to the enemy when defending.


