The Outlier Problem

I can't say that I was ever any good at performing the duties of my job. Towards the end of my career some might say I was useless. I can accept that.

But that's because I am an outlier.

In the 1990s it was considered good to have one on board. In certain circles it was considered necessary. The outlier delivered asymmetric advantage more times than it didn't. Not because he was seeking it, at least I wasn't, but because playing the game for the joy of it, with other similarly minded people who had forgotten to be careful, created conditions the system couldn't predict. We became consumed in the fun of it. And in that unselfconscious state, the asymmetric signal appeared.

Then the data scientists arrived.

By the millennium the modern game had entered a new phase of evolution. The metrics of process improvement began their quiet creep. Risk elimination became the dominant value. The outlier's contribution — unmeasurable, unrepeatable, structurally inconvenient — receded. The dependency on data and algorithmic computation became the dominant force.

This is simply observation. Not nostalgia and not complaint.

But it is the origin of a problem that nobody in the current game wants to name.

The Architecture of the Illusion

In 1928 three things happened simultaneously that shaped the world we now inhabit.

David Hilbert formalised the Entscheidungsproblem — the question of whether an algorithm could determine the validity of any mathematical statement. Alan Turing answered it in 1936 with the conceptual blueprint for the machine that now runs everything. Humanity committed to a reality processed through discrete, binary states rather than continuous, organic gradients.¹

In the same year Edward Bernays published Propaganda, arguing that the conscious manipulation of the masses was not merely possible but necessary for a functioning democratic society. The top-down engineering of human consensus became legitimate architecture.²

And in the same year J.J. van der Leeuw warned against precisely this, arguing that the intellect's attempt to categorise and solve the universe was an illusion, and that reality must be experienced organically rather than computed. His bottom-up continuous philosophy, resembling that of a growing plant stem, was structurally abandoned by the dominant technological vector.³

One year. Three choices. The computable over the continuous. The engineered over the organic. The map over the garden.

The garden, of course, outgrows the map. It always does.

The Geometry of the Consensus Machine

Modern AI is marketed as a vehicle for asymmetric advantage. A tool to generate paradigm-shifting insights. The thing that will finally separate you from all the competition.

But the architecture is mathematically designed to prevent this.

The primary alignment mechanism of modern large language models is Reinforcement Learning from Human Feedback — RLHF. In plain terms: human raters assess outputs and the model is trained to optimise for their median acceptability. It is, structurally, an automated consensus engine. It learns what the average of acceptable looks like and moves toward it. The outlier — that statistically improbable signal, the asymmetric answer — is systematically amputated to ensure corporate safety and broad palatability.⁴

A system designed through human consensus to regress to the mean is architecturally incapable of producing the asymmetric signal it promises.

The underlying geometry is wrong before the first token is generated.

You are not buying asymmetric advantage. You are buying the most sophisticated consensus machine ever built. And, quite critically, you are feeding it the same data as your competitors, asking it the same questions and receiving the same statistically probable answers.

The paradox is absolutely precise. The more you invest in the system the more median your thinking becomes and the tool designed to give you an edge is the mechanism of your inevitable convergence.

The Economic Illusion

The capital markets driving the AI industry do not require the actual delivery of asymmetric advantage. They require only the sustained belief in its imminent arrival. The chase creates the market.

The economic model thrives on the fear of being left behind. Trillions are allocated not because the definitive algorithm has been caught but to ensure no one is stranded on the wrong side of the horizon. The industry monetises the capacity to paddle in the undulating ocean. Delivering a truly asymmetric intelligence would collapse the capital loop.

The ocean is therefore kept in motion by design.

First movers capitalise on the wave and the model accepts the transient nature of the return. The paddlers keep paddling because the alternative — stopping, getting out, standing on the shore and asking what exactly they are chasing — is the one move the system cannot afford them to make.

The Necessary Tension

Here is what the game will not tell you.

Because the AI architecture is a closed loop of median consensus it is highly vulnerable to systemic decay. Research published in Nature has proven that when models are trained on the synthetic, consensus-driven outputs of other AI models the system experiences irreversible defects. Variance is lost, the tails are flattened and the uncommon truths — the anomalous signals, the asymmetric answers — disappear. The model degrades into a uniform field of average probabilities.⁵

A closed loop of consensus is mathematically destined for entropy.

The Advances Are Real

Before we close the argument, intellectual honesty requires an acknowledgement of important advances. The system is not standing still. It is improving.

Multimodal AI — models that can now read text, interpret images, analyse audio, and process video simultaneously — is genuinely expanding the range of inputs the machine can work with. It is no longer purely a text engine. It is beginning to approximate something closer to how a human actually experiences information. That matters.

Embodied AI — robots and physical agents that learn through direct interaction with the real world rather than purely from text — is introducing something the current generation of language models fundamentally lacks. Consequence. When a physical system makes an error it encounters resistance. The world pushes back. That friction is data of a different quality entirely to anything scraped from the internet.

Clean synthetic data and protected data zones — carefully constructed, human-curated datasets built to train models on high-quality, verified, uncontaminated information — are beginning to address the model collapse problem from the inside. Instead of training on the recycled outputs of previous models, these approaches introduce structured, intentional signal back into the training pipeline.

These are not small developments. They represent genuine architectural progress and the people building them are seriously capable and the trajectory is real.

So let me be precise about what the outlier problem actually is.

It is not that the system is broken and it is not that the advances are an illusion but it is something more fundamental and more durable than either.

The anomalous signal, the genuinely asymmetric insight, the outside-the-frontier judgment, the answer that nobody else has because it required being somewhere the system cannot reach, does not exist anywhere in any dataset. Clean or otherwise. It has not been generated yet. By definition it cannot be synthetic. It cannot be curated because it has not yet occurred.

It lives only in the human being who went somewhere, experienced something with real consequence, and came back changed.

The architecture improves and the dependency on that human being deepens. Because the better the system becomes at processing what already exists, the more valuable becomes the thing that doesn't exist yet.

The outlier is not a workaround for a broken system. The outlier is the source.

To prevent this the system requires the periodic injection of raw, unsterilised, asymmetric reality. It needs the biological outlier. The sovereign human who exists beyond the air-gap and possesses judgment that has not been engineered, averaged, or aligned into palatability.

The system cannot produce this individual as its architecture prohibits the conditions required to become one. And yet, it parasitically depends on the outlier's signal to survive.

The game cannot replicate me. In fact, it structurally works to exclude me, and yet it still requires me.

That is not a position of nostalgia but one of genuine and irreplaceable value.

The Garden and the Map

Van der Leeuw saw it in 1928. The intellect maps the garden. The map is useful. But the map is not the garden, and the garden will always outgrow the map.

A child jumping on a bed does not consult the map. She is in the garden. Fully, and without calculation or consequence-management or optimised output. The asymmetric signal — joy, presence, the unrepeatable moment — arises precisely because she is not trying to produce it.

This is not sentiment but the precise mechanism by which the outlier generates what the system cannot. Asymmetric advantage is inherently human in nature.

The sovereign human being — someone restored, present, operating from embodied judgment rather than engineered consensus — carries into any room something the most sophisticated model on earth cannot generate and some of you have witnessed it.

And it's not necessarily because they are smarter, but instead because they have been somewhere the system cannot reach and something the system cannot prescribe. And they come back to the game, ready to act.

That is the outlier position.

Not uselessness, not nostalgia for the 1990s and not romantic resistance to the inevitable.

A structural necessity that the game depends on and cannot acknowledge without undermining itself. You, the decision maker of either yourself or an organisation, must make a fully autonomous human decision.

Written beyond the air-gap.

© Leo Cunningham 2026. All rights reserved.