This article contains references to Sun and Steel, by Yukio Mishima. An essay I enjoyed and recommend.

Pre-tense - Geoffrey Hinton’s scary story.

There are highly prolific theories around the internet that AIs “may already be conscious”. You will hear this from AI prolocutors like the Nobel-Prize winner Geoffrey Hinton (who I have no doubt of his computer science accolades, but doubt deeply his understanding of being). Geoffrey has been on countless podcasts and talk shows, using powerful language to explain why we’re all ‘totally fucking fucked mate big time’. The purpose of this article is to refute these ideas.

The confidence trick.

A long time ago, far before computers existed, people realised they could use language tricks to influence people. There was a method to it. When arranged carefully, words can reliably produce emotional and behavioural outcomes. These methods have been used by salesmen, politicians, religious figures, etc. for millennia.

Large Language Models are professional con-men.

They are designed to put words together in a poetic way, to trick you. Engineers built a machine which could manipulate language in a way far faster than any meat-based rhetorician ever could. The desire was to create the appearance of intelligence, by creating an advanced language prediction model, which essentially hacked language.

Most of us have an inner monologue that narrates our thoughts, but not all of our thoughts. We don’t manually breathe, or have a narrator that tells us the specific movements our arms and wrists need to make for us to have a wank. Most of our cognitive processing is completely non-verbal.

This is the simplest and most concise rebuttal I have against LLMs in particular, but I’m going to go deeper.

Transistors vs Biology — the thermodynamic reality.

The human brain runs on roughly twenty watts, which is about the energy of a dim light bulb. With that energy, it performs continuous, real-time computation: predicting trajectories, coordinating muscles, adjusting balance, recalibrating vision, managing hormones, regulating temperature, processing emotion.

We have an ability to compute highly complex mathematical equations at incredible speed, with constant shifting variables. Not as a written mathematical problem, but in a real life scenario.

When a frisbee is thrown, and we have to run and catch it, our eyes, brain, and body automatically perform an inconceivable amount of complex calculations, at remarkable energy efficiency, in order to catch that disc. It’s not a conscious calculation, or one that is thought out in words or symbols. It’s a process that simply runs in the background of our brain. We do not calculate drag coefficients or gravitational vectors in words. Yet the eyes, vestibular system, cerebellum, and motor cortex together perform staggering predictive computation. Variables shift mid-flight. Wind, spin, our own momentum, and the body adjusts without conscious narration.

An LLM running on trillions of transistors in a data centre using ungodly amounts of energy, simply couldn’t compete with that in terms of energy efficiency.

LLMs operate by performing statistical pattern extraction at enormous computational scale. Their domain is language, their method is tokens predicting tokens.

So what?

Concepts like energy efficiency alone do not disprove the highly prolific theories that AIs may already be conscious, and my criticisms of LLM-based models do not prove that alternative neural-network architectures may not be able to achieve it.

But now I’m going to argue that the problem is not merely quantitative. To explain, I will turn to Yukio Mishima’s essay - Sun and Steel. (Spoilers)

Yukio Mishima on being.

For Mishima, consciousness was not rooted in refined linguistic abstraction but through intensified embodiment. Being was not something one reasoned about; it was something one strained into.

In Sun and Steel, Mishima describes his estrangement from language. He saw words as abstracting, distancing, and weakening. His early life was dominated by literary consciousness. A mind suspended in symbols. Only later, through bodybuilding and physical discipline, did he claim to discover a different mode of consciousness: what he called learning “the language of the flesh.”

ChatGPT is a fantastic tool for creating verbal language. However verbal language is a primitive form of intelligence - or more accurately, it is a derivative of intelligence. It is primitive in structure: we read in one linear dimension, from left to right. Complex ideas must pass through a narrow symbolic channel. Experience is flattened into sequence. It’s not a lateral, multi-dimensional form of thinking.

LLMs operate purely in that compressed symbolic layer. Biology does not.

When Mishima speaks of learning the language of the flesh, he is pointing toward a dimension of intelligence that precedes and exceeds language. It is the intelligence of balance, timing, pain tolerance, coordination, hormonal fluctuation, and risk. It is the intelligence that allows a body to move through a world of gravity and friction.

Language rides atop that system. It does not generate it.

Large language models invert this hierarchy. They begin with language and attempt to simulate intelligence from there. But language is a late evolutionary layer, and a refinement of a far older substrate. Without the substrate, the embodiment, metabolic constraint, and existential vulnerability - something essential is missing.

True intelligence may not consist merely in pattern recognition or problem-solving. It may require a world that resists you, a body that tires, a sky that blinds your eyes under strain.

Mishima recounts watching young men carry a shrine beneath the blazing sky. As a child he wondered what their upward gazes meant. Was it ecstasy? Revelation? Symbolism? Years later, after subjecting himself to sustained physical strain, he joined them and discovered the answer. They were not seeing a metaphor. They were simply looking at the sky. The blue intensity shifted between lucidity and madness because their bodies were pushed to the threshold of endurance. Their perception had been altered by heat, exertion, pain, and shared rhythm.

The experience was not linguistic. It was not conceptual. He argues it was not felt within the mind, but within the body itself. Something no disembodied system like ChatGPT can possess.

The young men Mishima watched were not processing tokens. They were bearing weight.

Large language models operate entirely within abstraction. They do not sweat. They do not fatigue. They do not experience altered perception under stress. They have no endocrine system, no lactic acid buildup, no pounding pulse. They do not face mortality. They do not act under metabolic constraint.

Their ‘intelligence’ is untethered from survival.

Without this fundamental connection of survival, AI will always remain disembodied. And its intelligence and consciousness are therefore incomplete.

Disembodiment removes the conditions under which spirit emerges. Without physical survival needs, AI can’t prefer outcomes, experience stakes, or face consequences. Even if they could simulate goals, they would have no intrinsic motivations such as fear and desire.

Intelligence in a living organism is inseparable from stakes. There is something to lose, something to protect. Risk is the lifeblood of the spirit. Biological intelligence evolved under the constraint of death. Mortality imposes urgency, selection, consequence. Without mortality, there are no genuine stakes. Without vulnerability, conscious cognition reduces to detached calculation.

Morrissey on uncertainty.

“Does the body rule the mind, or does the mind rule the body?

I don’t know.”