The flatness of it all.

No. The Singularity isn't going to happen.

I was going to write about markets this week but I’m not quite ready so instead I’m going to detour via some other stuff that has been catching my attention. Now with my usual disclaimer, I’m going to try and piece together things where I am no expert, but I think my logic is OK so bear with me.

Last week I saw a post by an AI founder (the usual composite: male, 30s, indiscriminate nationality, Phd in STEM etc;) that had me scratching my head. Now clearly this Olympian inhabits a different plane to me but I couldn’t help but be bemused by what he said:

People still don’t seem to grasp how insane the structure of language revealed by LLMs really is.

All structured sequences fall into one of three categories:

1. Those generated by external rules (like chess, Go, or Fibonacci).

2. Those generated by external processes (like DNA replication, weather systems, or the stock market).

3. Those that are self-contained, whose only rule is to continue according to their own structure.

Language is the only known example of the third kind that does anything.

There is then a lot of other stuff - including suggesting Chomsky was wrong - followed by this:

LLMs didn’t just learn patterns. They revealed what language has always been: an immanent generative system, singular among all possible ones, and powerful enough to align minds and build civilization.

Well, hmm. That got me thinking because I have been trying to understand a little more about evolutionary modelling recently. And this arose because I have been studying an archaeological puzzle relating to a critical phase in human history.

The Acheulean was a tool making tradition that began approximately 1.8 million years ago and held for at least 1 million years. During this time, enormous leaps happened in human history. We went from being big, shambling ape-like creatures to long-limbed, small-jawed hominins. Towards the end of the period, species such as Denisovans and Neanderthals emerged in the record.

If I can offer a very sketchy overview of the past 3m years, there were two leaps for our species:

• From ‘monkey’ (non-human primates) > hominins

• From hominins > anatomically modern humans

What is interesting about these leaps is that they are both related to tool use. Jane Goodall’s work showed us that Chimpanzees (and Bonobos) living in the wild, make use of tools such as stones (to crack nuts) and sticks (to forage for food). This has led to a decades long study of how such skills are acquired and transmitted. In 1992, Italian researchers at the University of Parma discovered ‘mirror neurons’ - cells that fire in mammals when they observe the actions of others. In mammals, these cells lead to unconscious copying of behaviour.

Think: two people leaning in at a table with hands identically clasped. Or: two angry figures gesticulating at one another.

Mirror neurons may offer explanations for behaviour such as empathy and co-operation. Indeed, they may act as a sensitive game theoretical signalling system. But they can also explain how non-genetic behaviours are passed within groups. The working hypothesis for tool use in non-human primate groups is that skills are transmitted through ‘emulation, observation and imitation’. We might call this ‘pattern learning’. There is no conception or model of the task. It is simply a pattern that is copied based on some earlier trial and error discovery. These patterns don’t tend to hold inter-generationally.

The first tipping point in human history occurs at the point where these tasks, associated with tool use, moved from pattern learning to template acquisition. This created an abstraction from an emulated pattern into a standardised ‘mental template’. There has been an enormous amount of research undertaken to understand this transition. We now know that the nut ‘knapping’ skills displayed by (non human) primates in the wild, appear to have evolved into stone ‘knapping’ skills in early hominins. Striking a nut from above with a hard object can break it into two pieces. Reconstructions have shown that early stone tools were created in the same way (a cobble split by a downward strike). From approximately 2.6m years ago, crude flaked stone tools appear in the archaeological record.

What is interesting here is that we might assume that new species evolve and then invent new tools. But it looks as though the causation may run the other way (or at least, both ways). While there were many hominin species appearing through the last 3 million years, the tool technology changes little but for two (principal) discontinuous leaps.

From a systems perspective, the question is what ‘attractor’ (or set of attractors) pulls each tradition to a new level of functional complexity? If you have read my post on Maxwell’s Demon, you will know that there must have been an energy gradient / asymmetry that ‘tugged’ the system out of equilibrium. This must have been met with some reinforcing feedback to set off the auto-catalytic changes that altered the system state.

New energy gradients were opening up 2.6m years ago as grasslands began to form, pair bonding in some species generated new social relationships (possibly leading to changes in cranial capacity) and climate change pushed species out of the tropics and into the surrounding regions. A combination of these factors may have inter-acted to create a turbulent phase transition with emergent new properties of: tool tradition, enhanced cognition (‘mental templates’) and social bonding. Honestly, it’s hard to know what caused what, but there was clearly a big mixing pot.

The next tipping point comes approximately 1.8m years ago. The crude tools associated with the Oldowan era (~2.6 - 1.5 million years ago) begin to be displaced by a new tool tradition. The most famous of these is best known by its Victorian name - the ‘handaxe’. The handaxe is a pear / egg / tear-shaped stone tool that is beautifully weighted to sit in (now smaller and less muscular) hominin hands. It has refined edges and often, a razor sharp flint at its tip:

Source: A History of the World in 100 Objects - Olduvai Handaxe, BBC

This tool tradition persisted for at least 1 million years - through periods of hominin dispersals out of Africa and through Asia, the Levant and Europe. It was maintained even as new variations of ‘homo’ arose. A beautiful, perfectly preserved handaxe was found at the bottom of a deep pit in a Spanish cave thought to contain the remains of early Neanderthals.

The puzzle for archaeologists is how such a tradition held through time. Two of the notable experts in this field - Mary Leakey and Glynn Isaac - offered differing interpretations. Leakey suggested that gradual (progressive) change was occurring throughout the Acheulean while Isaac proposed a ‘random walk’ hypothesis. I’ll come back to what this means.

Now, when I started to read about this period, I was fascinated to see what I considered to be signs of a dynamic equilibrium:

• consistency with slow change at the centre

• variability with rapid tweaks and small leaps, at the periphery

I feel that this complexity theory analysis of the Acheulean reconciles Leakey and Isaac as both describing the same edge of chaos, dynamic system, held in a robust topology by competing forces. Each node explores its local environment recursively, leading to numerous tweaks (the high variability that is seen in local manifestations of the Acheulean). Such a system must trace a strange attractor pattern (if it could be visualised in such a way) which would explain the lack of temporal or morphological progression associated with the Acheulean tool tradition.

So this is a very long way of saying that these interesting tipping points in human history appear to represent the co-evolution of tool use and anatomical change, held in place by fractal patterns. As I have explained elsewhere, fractal behaviour can be characterised as:

• non linear

• scale invariant

• self-similar

• recursive

The explanation for the transition from Oldowan tool technology to Acheulean tool use is fascinating. Through 3D modelling, archaeologists have discovered a hierarchy of non-sequential tasks, associated with the production of ‘handaxes’. To put this another way, tool use apparently moved from:

• expedient, trial and error, ‘in the moment’ exploration (primate tool use)

• to standardised, repetitive production (Oldowan)

• to complex planned and designed artefacts (Acheulean)

In cognitive terms, we went from pattern learning > mental templates > hierarchical, nested, non-sequential abstraction. Or to put it even more simply, we went from:

• “Copy me” to …

• “Do this before you do that because it will be better for some future purpose that we cannot yet see, but will likely happen”

This is quite a big deal. It might even be ‘insane’ to quote our friend from earlier.

And the question is what fractal set of rules held all this in place?

The answer now being proposed by archaeologists and anthropologists, is that the explanation lies in cultural traditions communicated through complex language use that mirrored the nature of the tool manufacture (nested hierarchies, temporal abstractions etc;).

To remind you of my arguments in earlier essays, evolution does not appear to proceed in a step by step, gradual fashion. Complexity scientists have shown us that instead of systems ‘blending’ when they combine, they can (with an energy gradient and reinforcing feedback loop) set off auto-catalytic ‘clocks’ that create ‘order for free’ and rapid phase shifts. I have described the fractal shapes that cluster around these shifts in concrete forms - hexagons, branches, unfurling beans. But I am beginning to imagine cultural fractals, such as hierarchies and rules-based systems.

To understand the evolution of language, we must understand the complex set of systems that likely gave birth to it. These may have included: teaching, adolescence, gender roles and cultural traditions that all held the Acheulean tool manufacture ‘fractal’ in place.

On this reading, language is not an “immanent generative system”. This phrase means that it emerges from its own parts. But evolution doesn’t work that way. Only a scientist, steeped in mathematical models could make such a category error. Because scientists try to map complex phenomena with numbers. They create Gaussian distributions to describe equilibria and Markov chains to model change. But these adjustments are still disembodied. They are abstractions of a process that is deeply complex, built from many layers of concrete interaction and steeped in noise. Unlike a mathematical model, a fractal system can never be measured to a precise level. There is always another layer - the system is unbounded.

In one of Oscar Wilde’s saddest stories, a Nightingale presses its breast to a white rose to stain it red with its blood. It does this to help a young student woo his love. But when the nightingale has given up its life and the red rose is plucked and presented to the object of the boy’s affections, she spurns him:

‘What a silly thing Love is,’ said the Student as he walked away. ‘It is not half as useful as Logic, for it does not prove anything, and it is always telling one of things that are not going to happen, and making one believe things that are not true. In fact, it is quite unpractical, and, as in this age to be practical is everything, I shall go back to Philosophy and study Metaphysics.’

The AI entrepreneur who thinks he’s hit on a great revelation with his LLM insights cannot see the nightingale that lies dead on the grass with a thorn in her side. He cannot see how evolution is embodied through concrete interactions, not dis-embedded into an abstract mathematical space. Things happen that make things - even metaphysical things like languages.

When Glynn Isaac posited a ‘random walk’ explanation for the Acheulean, he was using the (then available) tools of mathematics. A ‘random walk’, initially termed a ‘drunkard’s walk’, is, as it sounds, a walk in which each step is made without foresight or planning. Mathematicians use it to model stochastic processes and it is derived from Brownian Motion. But this walk is insufficient to explain evolutionary change. Indeed, from a systems perspective, a ‘random walk’ can be imagined as a drunkard tethered to a peg. Through time, the drunkard explores all the space available to him through regular staggers away from the peg. If you can imagine looking from above at this behaviour, eventually the entire phase space will have been sampled (through the walks) and you will see a filled in circle.

A random walk simulation will approximate to a Gaussian distribution (or precisely ergodic system). But a random walk simulation does not incorporate noise or asymmetry. The more system states are sampled, the less variance in the system. It cannot change, it can only collapse.

To put this differently, the drunkard tethered to the peg, cannot make the ‘long searches’ I described in The Unkillable Bird. He can only sample a fixed space around him until eventually he’s been everywhere and all novelty is lost. As I set out in Maxwell’s Demon, the asymmetry in the universe is what creates the turbulence or vortices that seed new life. So any system that is modelled on the basis of an unmoving peg is doomed to collapse. The problem with the concept of an immanent generative system is that it assumes that novelty leaps from within the drunkard’s random staggers. It assumes that sampling all that phase space, with gradually decreasing variance, will somehow launch a tipping point and phase change.

This is impossible.

Or if it is possible, it represents a new form of evolution. At most, a random walk can produce a feeble form of genetic drift. Cellular automata can uncover existing patterns by brute forcing pathways for known rule sets. But neither is creating emergence (new functional complexity).

Now AI advocates will quickly tell me that their models are based on far more complex principles than Brownian Motion and Random Walks. They adjust to incorporate asymmetry and they will use statistical behaviour that artificially moves the ‘peg’. This is true. New training data and reinforcement learning can introduce some noise into these abstract mathematical systems. But it cannot make up for their inherent flatness.

And it seems to me that this flatness is important for understanding what’s driving us slowly mad in modern societies. Because unlike human language, which is a beautiful co-evolved complex adaptive system, governed by fractal rules and unable to run out of novelty, digital machines are merely an abstract, mathematical representation of these forces. They are not embodied in a complex sense. They did not arise, as language seems to have, from physical tasks that required symbolic layering and temporal abstraction. They were designed from the top down to mimic such relationships and they were created with mathematical precision + a little statistical noise. These machines could never trace a true strange attractor shape. They are bounded, reductive walks that will collapse without interaction from humans.

This is why the singularity is impossible.

The singularity requires a machine to be a genuine immanent generative system. It requires novelty to spring from the same old patterns. This can’t be possible. Even with all the compute in the world - ‘electrons’ as Sam Altman casually puts it - you could never create a man-made recursive, fractal system with the energy frugality of the universe’s mixing pots. It just can’t happen.

So what do we have instead? We have these odd, disembodied spaces where we can no longer rely on our sensory perceptions. We cannot sniff, taste, feel, hear or see our way through them. The digital spaces are rigid, top down and designed from blueprints and GANTT charts. No wonder we are so frustrated when we try to interact with them. They can’t possibly sample every state that we need from them. Instead, they run around their peg until the novelty requested from them is too great and we are left empty-handed and helpless with our unfamiliar problems.

This is annoying, but not too terrible when it’s our grocery app or streaming app. But when it comes to social communities, including this one, the harm is greater. Instead of forming embodied, multi-dimensional relationships in physical, tangible spaces, we are condemned to express ourselves through language alone. Language that sprang from embodied tool use is now reduced to object categorisation in digital spaces. Are you ‘this type’ or ‘that type’? Are you in this ‘this group’ or ‘that group’? Do you mean ‘this’ or ‘that’? It is all so flat.

My formative friendships from long ago were never based on such simplistic, linguistic categorisation. We all had different political views, hobbies and backgrounds. But oh, how we laughed as we lay with our backs to the grass, tea-stained Waugh in our hands and a bottle of wine at the ready. I don’t doubt that friendships are still formed this way but how are they maintained? Through texts and Whatsapp groups? Through this disembodied, disembedded language categorisation?

What strikes me most, is that we are unable to pull ourselves together and do something about it because we accept these disembodied spaces as a meeting ground. No wonder political action these days is a hashtag. No wonder there are so many groups trying to come together digitally with their Zoom calls and their digital platforms. But it’s all so flat. It’s the same old abstracted language categorisation: “are you this” or “are you that”?

What struck me recently, with the monstrosity of the recent tranche of Epstein emails, is that this man instinctively understood the power of embodied spaces. All that carnality and physical luxury fed these foolish men’s System 1 brains. All these celebrated men who put their pre-frontal cortex on standby at the sign of a private jet and some underage girls.

The Left once understood this point about embodiment too. They organised in physical spaces and formed complex local and working relationships. They had clubs where sweaty men could meet, smoke and shake the asbestos out of their hair. But now we organise online with our System 2 brains demanding that everything be categorised, shortened, summarised and bullet pointed with all the richness removed. Are you “this hashtag” or “that hashtag”?

Wordsworth described poetry as “the spontaneous overflow of powerful feelings”. He saw it as the interaction of thoughts and “organic sensibility”:

Poems to which any value can be attached were never produced on any variety of subjects but by a man who, being possessed of more than usual organic sensibility, had also thought long and deeply. For our continued influxes of feeling are modified and directed by our thoughts, which are indeed the representatives of all our past feelings; and, as by contemplating the relation of these general representatives to each other, we discover what is really important to men.

We are living in a world in which are bodies are being progressively dis-embedded from our cognition and our ‘organic sensibility’. It feels as though our heads are outgrowing our endocrine and sympathetic nervous system. This makes it so difficult to judge and discriminate since we have become so reliant upon cognition for categorisation. I think it explains some of the visceral hatred directed at ‘experts’ and ‘intellectuals’ who use clever language to explain a world that does not match our felt experiences. (This post is probably guilty of this too).

We have long since disembedded the natural world from our societies but how long before we condemn our human senses to the same fate? How long before Silicon Valley forces ‘friend apps’ upon us? Or monetises our closest relationships into scorecards and feedback systems? How long before China’s social credit scheme becomes the norm and every part of our lives is surveilled, ranked, rated and reduced to digital abstractions?

The singularity would be a gift in such a world - an advanced species would shrink us back into our ecology and let the system run itself quietly for a few million years with a handaxe or two to keep us going. But I fear we have much less time than that. The drunken man on his tether is running out of options and there is nowhere we can go unless we break free.

Oh, the flatness of it all.