Your complimentary articles

You’ve read one of your four complimentary articles for this month.

You can read four articles free per month. To have complete access to the thousands of philosophy articles on this site, please

SUBSCRIBE NOW

Thoughts on Thought

We’re as Smart as the Universe Gets

James Miles argues, among other things, that E.T. will be like Kim Kardashian, and that the real threat of advanced AI has been misunderstood.

In 1995 the English mathematical biologist John Maynard Smith published The Major Transitions in Evolution. Maynard Smith was a co-developer of gene-centred evolutionary biology in the 1960s and the founder of evolutionary game theory in the early 1970s, and what this book was quietly saying about high intelligence can be considered revolutionary, or at least, notably evolutionary. As a result of a paper I wrote about evolution in the journal Philosophy in 1998, I ended up friends for the last decade of his life with the other co-developer of gene-centred evolutionary biology, the American biologist George C. Williams. And like Maynard Smith and the early Darwinians (including Darwin), Williams was fascinated by the so-called ‘high intelligence paradox’ and the evolutionary problems it implies. One of the early Darwinians even turned to philosophy to try to solve the paradox.

Although evolution is predominantly gradual, theoretical biologists recognise a series of major transition events, which involved profound changes in the way information is stored and transmitted between generations. In his Major Transitions, Maynard Smith recognised eight of them across the four billion years of life on this planet, including the early move from an RNA world to DNA, the transition from prokaryotic to eukaryotic cells, and the transition from asexual to sexual reproduction. His seventh major evolutionary transition was the genetic flip-switch that allowed some species to live in large colonies, curing Darwin’s headache about non-reproductive insect castes and sterile workers. High intelligence through language was Maynard Smith’s eighth and ultimate evolutionary transition.

Evolutionarily speaking, natural intelligence caps out at two hard ceilings. Everywhere in the ecosphere intelligence is constrained at a relatively low level by the requirements of group stability (‘Ceiling 1’ in the graph). This incorporates Maynard Smith’s concept of the ‘evolutionarily stable strategy’ or ‘ESS’; meaning that a limited natural level of intelligence creates a sustained response to selective pressures. High intelligence such as ours is then the product of a further, one-in-a-billion major evolutionary event. (Remember that number: it will be key both to turning E.T. into Kim Kardashian and to the dangers of advanced AI.)

Perhaps Maynard Smith’s identification of human high intelligence deriving from language as one of the eight major evolutionary tipping points across four billion years seems bizarre to you, since he’s placing human linguistic development on a level of importance equivalent to the rise of sexual reproduction or multicellularity. At first glance this should seem like a stunning level of species self-conceit, if for no other reason than his other seven transitions have had effects on anything from thousands to billions of species, but transition eight involves just one species. However, it’s only bizarre if you view intelligence as a continuous spectrum. But as possibly the finest mathematical biologist of the last half-century, Maynard Smith had worked out that human-level intelligence could not exist on any primate (that is, pre-linguistic) cleverness spectrum. It had to be the result of an evolutionary discontinuity. So it had to be the final evolutionary tipping point.

However, the level of achievable linguistic intelligence caps out as well (‘Ceiling 2’), because such intelligence is thereafter invisible to natural selection (as I will explain). This means human-type intelligence is as smart as the universe can get. Ever. No matter how old the universe gets, no matter how long species evolve.

Extraordinary, right? Except, Maynard Smith was not actually saying anything new. In his correspondence with Alfred Russel Wallace in December 1857, Darwin had already termed human-level intelligence “the highest & most interesting problem for the naturalist.” Wallace agreed with Darwin; he called it ‘the inverse problem’ of his and Darwin’s work: “to account for facts which, according to the theory of natural selection, ought not to happen.” And Darwin’s ‘bulldog’, Thomas Huxley, president of the Royal Society in 1883, would write and speak expansively on the high intelligence paradox. Huxley, who was also at one stage president of the UK’s Metaphysical Society, would turn from biology to philosophy – including the work of Heraclitus, Berkeley, and Kant – to try to resolve the problem. In seeking to get to the bottom of things, my friend George Williams, working with the MIT cultural historian James Paradis, would also write a book on Huxley’s answer to the paradox.

The high intelligence paradox, Darwin’s ‘highest & most interesting problem’, is that humans do not behave in the ways natural selection theory would at first sight predict. We co-ordinate in group sizes vastly larger – orders of magnitude larger – than we should be naturally capable of. We should not be capable of compassion. And we are far more intelligent than we should be. These are not unconnected problems, however: they turn out to be a developmental triad, a transitional troika of troubles. Let’s look at them.

The Group Size Conundrum

Group size is effectively kept stable by positive and negative evolutionary forces, including food availability, security, and evolutionarily stable strategies of behaviour; but generally, the smarter mammals get, the smaller their group sizes tend to be. Each year, wildebeest thunder across the Serengeti in their tens of thousands; but grass hasn’t yet evolved to escape, so wildebeest don’t have to work together to feed. The Oxford biologist Bill Hamilton wrote a paper in 1971 that was the start of what is called ‘selfish herd theory’, highlighting that the individuals of a herd are forever cover-seeking – pushing for the safest positions at the expense of others, including pushing to get to the centre of the herd, and thereby pushing others to the dangerous edges. Behaviour that necessitates many co-operative interactions between group members, in contrast, requires strong group cohesion. Baboons will usually live only in groups of up to a couple of hundred. But smarter primates are better at cheating, and smaller groups more effectively limit the opportunity for cheating and the subversion this creates, so our closest living relatives, chimpanzees and bonobos, rarely live in groups larger than a hundred. However, humans have been called ‘supercooperators’ by one Harvard biologist. Even very different genetic inheritance mechanisms do not get close to creating creatures mimicking our level of co-operation, or our level of non-violence. Surprising as it may seem, we are in general thousands of times less violent than other primate species. To quote the late Stephen Jay Gould, we are ‘a remarkably genial species’.

Our group size problem exists irrespective of the underlying genetic model used. The naked mole-rat has been argued to be the closest a mammal comes to behaving like a social insect, with colonies of up to three hundred and a behavioural and reproductive division of labour. They have achieved this stable state partly through intense inbreeding. Naked mole-rats have average intra-colony relatedness of 0.81 – the highest recorded for a natural mammalian population. Yet neither mole-rats nor hard-wired social insects can be said to live ‘harmoniously’, as humans (comparatively) do. Naked mole-rat colonies have been called ‘reproductive dictatorships’, in which the breeding female is typically the most dominant animal in the colony. Rival challengers for breeding status are usually attacked and killed by the queen; and remember that these are often her very – and I mean very – close relatives, such as inbred sisters. Or, to quote from Mark Elgar and Bernard Crespi’s survey of cannibalism across social insects, “the social situation in colonies changes dramatically when the queen’s pheromonally-mediated and behavioural dominance begins to fail in late summer… Fighting occurs between queens and workers, and among workers” (Cannibalism: Ecology and Evolution Among Diverse Taxa, 1992).

Illustration © Jaime Raposo 2024. To see more of his art, please visit jaimeraposo.com

E.T. will be Kim Kardashian

However, the paradox isn’t just that humans can cooperate in group sizes and achieve a division of labour not explainable by pheromonal suppression, reproductive dictatorships, or some correlate to the close genetic relatedness found across ants, bees, wasps, or the intense inbreeding found across termite colonies. We also cannot explain our moral behaviour in terms of earlier evolutionary transitions.

To quote the primatologist Frans de Waal, chimpanzees live in “a world without compassion”. Even bonobos, once touted by some as a proto-moral ape species, have, like chimpanzees before them, begun to shock us as more field data emerges. For instance, in 2016, Nahoko Tokuyama and colleagues caught on video instances of bonobo maternal cannibalism of recently dead offspring. One mother “grabbed the carcass in her right hand and suddenly bit off the head” of the motionless infant she had been grooming just a few hours before.

As long as you accept that humanity evolved from a common ancestor with chimpanzees and bonobos, you have to find some way to get from small group cohesion in an ape ‘world without compassion’, to ‘supercooperators’ in ‘a remarkably genial species’.

Furthermore, critical to human intelligence is our mass coordination with a complex division of labour. This is necessary for advanced language, writing, generational learning, scholarship, and advanced thought. You cannot get to complex reasoning and generational learning through the small groups otherwise found in mammalian nature, either out- or inbred. But you also cannot get to complex reasoning and generational learning with the transition to closely-interrelated hive living. Groups can get large, but only when there is a large degree of irreversible specialisation and genetic hard-wiring. Once you bring in mental plasticity, then, as with the naked mole-rat, co-operative groups never get above a few hundred.

What explains this discontinuity in humanity, then? For Darwin, writing in The Descent of Man (1871), the transition came about when there was sufficient brain complexity – enough to bring both a susceptibility to hopes and fears and the medium of language. Huxley at first seemed to suggest that reason might be the key: “Fragile reed as he may be, man, as Pascal says, is a thinking reed” (Evolution and Ethics, 1894 ). Huxley later backed away from this idea and aligned with Darwin, but it is an idea worth re-considering. But given our common ancestor with chimps and bonobos, could reason ever be enough to explain both our cooperation and our compassion?

In his short book Perpetual Peace (1795), Immanuel Kant wrote that even a race of devils (Volk von Teufeln) could cooperate to build a state if they were intelligent enough. However, the problems are, firstly, that they seemingly have to be perfectly rational long-term thinkers, yet evolution can do precision but not perfection; and secondly, that they remain at heart a race of devils rather than a remarkably genial species.

There are two explanations currently trending in biology. Both invoke a fundamental change at the genetic level. The first, termed ‘multilevel selection’, or MLS, requires higher-acting levels of natural selection. To quote the leading theorist in the MLS school, David Sloan Wilson, “multilevel selection theory has the potential to explain not only why humans are ultrasocial, but why they have experienced a unique variety of group selection” (Unto Others, 1998). Observe that term ‘ultrasocial’, and also that troubling requirement for a ‘unique variety’ of group selection. (Incidentally, Dave Wilson began as George Williams’ postdoc.) The second explanation, termed ‘human sociobiology’ – and which has sprouted off the popular branch called ‘evolutionary psychology’ – invokes the emergence of new and counteracting genetic mechanisms. Human sociobiology was theorised in the 1970s by two American biologists, Robert Trivers and the late Edward O. Wilson, the latter naming the discipline. In 2010, however, Ed Wilson abandoned his own theory when he subscribed to Dave Wilson’s model for the evolution of high-level social organisation in both eusocial insects and human beings.

Neither MLS nor human sociobiology link our transition to a compassionate species to a complex brain in and of itself. For both schools of thought, the most complex brain in the animal kingdom is (at best) correlated with an evolutionary transition such as a ‘unique variety’ of group selection. But Occam’s razor tells us that it’s more likely that our uniquely complex brain somehow fully explains our exceptional ability to live in vast groups without also requiring, say, a unique genetic re-wiring.

Whatever it was that allowed us to transition to large groups – a complex brain susceptible to hopes and fears; the independence of reason; or new genetic or group-selection mechanisms – group bonding after this point becomes synthetic rather than genetic, or, if you prefer to put it this way, cultural rather than biological. Groups of one hundred family members could suddenly become non-family groups of thousands or hundreds of thousands. Family survival units have become states and nations, which form and reform according to shifting and often contradictory cultural concepts.

However, this means that in our culture there’s now no longer anything for natural selection to operate on – which means, no natural selection pressure to get smarter. Natural groups are transcended at exactly the point brains reach the complexity to allow shared concepts, which catapults group size upwards and starts culture and civilisation; but everything comes to a grinding halt evolutionarily speaking. Because our behaviour is no longer an expression of natural selection in operation, after this point evolution cannot, and will not, make those beings any smarter. This means that we are, and must be, the smartest beings the universe can get to naturally.

It also suggests that if there’s other intelligent life in the universe, it will also have an average IQ around our 100 – the level we achieved when natural selection forces were transcended in our social development. We have long wondered if out there are creatures with “intellects vast and cool”, as H.G. Wells wrote of his Martians, or “curious and dispassionate, observing us”, as suggested by Carl Sagan. But because of the large group cohesion problem and its entirely post-natural-selection potential solutions, intellects much higher than ours – vast and cool, or curious and dispassionate – cannot exist in this universe. Technologically, yes, ETs might be far more advanced than we are; but psychologically, temperamentally, and intellectually, they will only ever be at the level of the characters on The Kardashians. After thirteen billion years of stellar nurseries, in a cosmos tens of billions and maybe trillions of light years across, all the universe has ever got to, or will ever get to, is intellects at the level of political wrangling, conspiracy theories, and reality TV.

AI’s One-in-a-Billion Problem

To the best of our knowledge, only one animal species out of more than a billion has ever managed to escape a ‘world without compassion’ (we can include here domesticated dogs, who have been given an artificial selective tweak that masks but does not free them from that world). This is why both human behaviour and the human intellect bothered Darwin so much that it was “the highest and most interesting problem.” It also means that pursuing high-level artificial intelligence through a proxy to biological development is fraught with a billion-to-one danger.

We’re already learning to live with narrow, or weak, AI, but the ultimate goal of almost all AI research is general intelligence, sometimes called strong, full, true, or human-level (and beyond) AI, but formally known as ‘artificial general intelligence’, or AGI. However, if AGI is pursued using proxies to biology, then it seems to be a billion times more likely that AGI does not attain the evolutionary transition event that leads to our compassionate nature. In other words, AGI is a billion times more likely to represent, in its essence, a world without compassion.

In standard biology, except for a one-in-a-billion transition, we cannot get away from the world without compassion, because competition for survival is the very algorithm behind increasing complexity. It’s why evolution by natural selection works so successfully, and so yields the behaviour patterns found across social insects, mole-rats, wildebeest, domesticated dogs, baboons, chimpanzees and bonobos. Yet AI development has been all about biological proxies for at least the last decade. Artificial neural networks (ANNs) took over from symbolic AI partly because it became obvious that symbolic reasoning could not be at the core of biological intelligence. ANNs are software abstractions of animal brains, if you like.

It is difficult to believe that the current AI debate understands this true danger of AGI. Geoffrey Hinton, one of the ‘godfathers’ of AI, made waves talking about the dangers of AGI. He told the MIT Technology Review that he thinks there are two types of intelligence in the world now, the animal brain and the artificial neural network: “It’s a completely different form of intelligence”, he said, “A new and better form of intelligence”. But the greatest danger of AGI (should we one day get there) will not be that it’s a new and better form of intelligence; the danger will be that it’s a very old form of intelligence – one whose development is driven by natural or artificial selection, not compassion.

AI is mimicking the biological products of evolution, but does this mean it’s mimicking evolution itself? Not necessarily. But personally, I am not keen to bet our future on technologists’ ability to understand (and care about) the difference between mimicking organisms and mimicking evolution. Moreover, the only evidence we have to date across four billion years, is that if you program behaviour through an efficient inheritance mechanism (such as machine learning could be said to be) you end up in a world without compassion, except for a billion-to-one chance. Moreover, parts of the AI and machine learning communities are actively using the evolutionary algorithm itself. And the more ‘black box’ deep learning, the more ‘unsupervised’ learning, the more the competition and need for speed, the greater the risk of unforeseen (ie bad) consequences. Individual-level selection beats group-level selection in almost every situation because it is much faster, more stable, and less open to subversion; so individual selection evolution is potentially a very fast and efficient way to progress machine learning. But it’s one with a billion-to-one baked-in problem. (And let’s not forget the five AGI projects evolving behind the walls of Chinese state censorship; with another three being run from inside the corruption, aggression, and gangsterism of modern Russia.)

We have known about the high intelligence paradox for over 165 years, and we have not progressed in our understanding of it. Arguably, we are further from resolving the conundrum than we were in the first decades of Darwinism. It’s the paradox that keeps on giving. The paradox that shows the human animal to be the one-in-a-billion species exception. The paradox that turns E.T. into Kim Kardashian. The paradox that the real danger of advanced AI is not that it will be something very new, but that it may well be an example of something very old indeed.

James B. Miles writes on the implications of evolutionary theory.