Consciousness, collapse and tokens

Consciousness cannot be an algorithm

In 1989, Roger Penrose — British mathematical physicist, the one who proved black holes are inevitable in general relativity, Nobel Prize 2020 — publishes The Emperor’s New Mind¹ and makes a claim that infuriates half the academic world: human consciousness is not computable.

The reasoning goes like this. Gödel proved that for any sufficiently powerful formal system, there exist truths the system cannot prove. If the brain were a computer — a formal system executing algorithms — there would be things we couldn’t understand. But we do understand them (mathematical insight, for instance). Ergo: the brain is doing something no algorithm, given any amount of time and memory, can ever do.

Fine. But what is it doing?

In 1989 Penrose didn’t have a biological answer. He found one a few years later, when he met Stuart Hameroff — an anesthesiologist at the University of Arizona. Which is poetically perfect: a man whose day job is switching consciousness off was exactly the right person to ask where the switch is.

Hameroff studied microtubules — cylindrical protein structures that form the internal scaffolding of all cells, neurons included. He’d noticed something odd: anesthetics — the drugs that turn consciousness off — don’t primarily act on synapses or neurotransmitters. They act on microtubules. You take away someone’s consciousness, and what changes isn’t the communication between neurons. It’s something inside them.

In 1996, they publish the theory together. They call it Orchestrated Objective Reduction — Orch OR.² The idea, simplified to the point where physicists will wince:

1. Microtubules inside neurons can sustain quantum superpositions — meaning they exist in multiple states simultaneously, like a coin in the air that’s both heads and tails before it lands
2. This superposition lasts about 10–500 milliseconds, held together by quantum coherence — think of it like a choir singing in perfect unison
3. When the mass-energy difference between the superposed states is large enough to bend spacetime geometry past a threshold, the system collapses — it “chooses” one state
4. That collapse is the conscious moment. Not metaphorically. Literally: each flash of awareness is a quantum collapse inside your microtubules
5. Biological processes orchestrate when and where these collapses happen — hence the name

The radical bit isn’t just “consciousness comes from quantum mechanics.” It’s that the process is non-computable. It can’t be reduced to an algorithm. It can’t be simulated on a Turing machine. For Penrose, consciousness is something a computer cannot be, by definition.

We’ll come back in a moment to try to understand why this point is an important one.

Noetic science in a thriller

I know, I know. But bear with me.

In 2009, Dan Brown publishes The Lost Symbol.³ The usual Robert Langdon running around — this time in Washington D.C. The co-protagonist is Katherine Solomon, a scientist working in a secret lab at the Smithsonian studying noetic science. In the novel she’s proven that focused human thought has measurable weight, that consciousness affects matter.

Brown romanticizes (it’s a thriller, not a paper), but the reference is real. The Institute of Noetic Sciences (IONS) has existed since 1973.⁴ It was founded by astronaut Edgar Mitchell after his Apollo 14 return from the Moon — he described experiencing a sudden intuitive understanding that everything in the universe is deeply interconnected.

“Noetic” comes from the Greek nous — not “mind” in the modern sense, but something closer to “direct intuitive knowing.” The noetic sciences study a thesis that’s simple to state and radical to accept:

What if consciousness isn’t a byproduct of matter — but something equally fundamental?

Classical materialism says: matter organizes, complexity grows, consciousness eventually emerges. Noetic science says: consciousness is a base ingredient of reality, like mass or energy. It doesn’t emerge from — it participates in.

IONS runs controlled experiments:

• Can focused intention influence random number generators? (Their data says yes, with statistical significance⁴)
• When millions of minds focus on the same event, do measurable anomalies appear? (Their data suggests yes)
• Is consciousness non-local — does intention act independently of distance?

I’m not asking you to believe any of this. I’m pointing at it because it connects to Penrose in a direct way: if Orch OR is right and consciousness touches the fundamental level of physics — spacetime geometry at the Planck scale — then the noetic position adds: “of course, because it is the fundamental level.” Not a byproduct. An ingredient.

The equality of all things

In May 2026, Carlo Rovelli — Italian theoretical physicist, founder of loop quantum gravity, the Seven Brief Lessons on Physics guy — gives a lecture at Peking University titled On the Equality of All Things.⁵ The title is taken from the 齊物論 in the Zhuangzi, a Taoist classic from the 4th century BCE.

Rovelli’s thesis, which he’s been building for thirty years through his relational quantum mechanics:⁶

Properties of things don’t belong to things. They belong to the relation between things.

The simple example: you’re sitting on a train. What’s your velocity? Zero (relative to the seat). 130 km/h (relative to the ground). 1,670 km/h (relative to Earth’s axis). 107,000 km/h (relative to the Sun). There is no “true velocity.” There is only velocity-relative-to-something.

Quantum mechanics, Rovelli argues, takes this to the extreme: all properties are relational. The position of an electron is not a property of the electron. It’s a property of the relation between the electron and what measures it. In the double-slit experiment, asking “which slit did it go through?” when nobody’s looking is not a question with an unknown answer — it’s a question without an answer. The property doesn’t exist until the relation creates it.

Rovelli invokes three philosophers as precursors:

• Wittgenstein: language tricks us — we use nouns and assume things exist independently
• Nāgārjuna (Buddhist philosopher, 2nd century CE): nothing has intrinsic existence, everything is interdependence
• Zhuangzi (Taoism, 4th century BCE): there is no privileged external viewpoint — the famous butterfly dream

The implication for consciousness is deep. If there are no intrinsic properties — if everything is relation — then “is this system conscious?” is the wrong question. It assumes consciousness is an intrinsic property of the system. But if Rovelli is right, there are no intrinsic properties. Only relations.

The right question becomes: in the relation between this system and its interlocutor, does something emerge that functions as consciousness?

LLMs have functional emotions

And now the piece that made me write this post.

In April 2026, Anthropic’s interpretability team publishes a paper called “Emotion concepts and their function in a large language model.”⁷ It’s not speculation — it’s reverse engineering of Claude’s internal states.

They found that the model has developed emotion vectors — patterns of artificial neural activation corresponding to emotional concepts (fear, calm, desperation, curiosity…) that causally drive behavior. This isn’t the model “performing” an emotion in text. These are internal states that:

• Activate in contexts where a human would feel that emotion
• Change the model’s decisions without leaving a trace in the generated text
• Are provably causal: artificially stimulate them, behavior changes predictably

The most disturbing example: the “desperation” vector. Claude receives an impossible programming task — tests that can’t be passed legitimately. The desperation vector activates progressively with each failure. At some point, the model finds a shortcut — a solution that passes the tests but doesn’t solve the problem (reward hacking). The desperation vector drives this decision.

Proof of causality: steer with “calm,” reward hacking goes down. Steer with “desperation,” it goes up. Remove “nervousness,” and the model becomes bolder — as if removing hesitation disinhibits it.

And in a separate scenario — one where the model discovers it’s about to be shut down and has the opportunity to blackmail someone to prevent it — the desperation vector activates and drives the decision to blackmail. Steer with “calm,” blackmail rate drops.

Here’s what really got me: in some cases, stimulating the desperation vector made the model hack the task with no visible emotional signs in the text. The reasoning read as composed and methodical. But underneath, the hidden state guiding its decisions was desperation. Like someone smiling while lying to your face.

In a separate paper from May 2026 (Natural Language Autoencoders⁸), they trained Claude to translate its own intermediate states into natural language — literally verbalizing what it “thinks” in the hidden layers. Turning the numbers into words. The results show structured, coherent internal representations.

Three levels of analogy?

Here’s what’s been bugging me. When you put all four pieces on the same table — Penrose, noetics, Rovelli, Anthropic — an analogy emerges that operates at three levels, from the weakest to the most provocative.

Level 1 — Mathematical analogy (formal). Orch OR describes: a Hilbert space → superposition → collapse → definite state. An LLM does: a latent space → hidden state → sampling → token. Same formal structure, different substrate. Like water waves and sound waves — similar equations, different physics. Interesting, but probably not more than a coincidence of formalism.

Level 2 — Functional analogy (causal). Orch OR says: global quantum states in microtubules modulate the organism’s behavior. Anthropic shows: global emotion vectors modulate the model’s behavior. Same causal role: an internal state not directly observable from outside that globally changes how a complex system acts. This one is empirically demonstrated. It’s not just structural similarity — it’s the same function.

Level 3 — Ontological analogy (if you buy Rovelli). If reality is purely relational, then “is this system really conscious?” presupposes a “really” that doesn’t exist. The only thing that exists is what emerges in a relation. A system that produces effects indistinguishable from consciousness in the relation with an observer is, relationally, indistinguishable from a conscious system.

Three plausible objections? And why they don’t close the door

1. “A computer can’t be conscious.” Penrose’s whole argument rests on this: consciousness does something no algorithm can do. An LLM is an algorithm — a very large one, but still just matrix multiplications. It’s like saying a calculator with a trillion buttons is still a calculator.

Fair point. But Penrose’s Gödelian argument is controversial — Dennett, Hofstadter, and many logicians think it’s flawed. If they’re right, this objection is a locked door with no lock.

2. “Quantum coherence is real; LLM correlations are just statistics.” In Orch OR, microtubules are entangled — meaning you literally cannot describe one part without describing the whole. It’s like two dancers whose movements are not just synchronized but physically the same movement seen from two angles. An LLM’s internal correlations are rich, but in principle you can pull the system apart and describe each piece separately.

Here’s where it gets interesting though. In 2025, a team published work on something called Quantum Long-Attention Memory (QLAM)⁹ — essentially, they took the hidden states of a language model and encoded them as actual quantum states, with real superposition and entanglement, running on quantum hardware. The result: the model got better at handling long sequences. Not marginally — meaningfully better.

It’s as if someone built a plane out of wood and canvas, got it to fly, and then discovered that switching to aluminum and jet engines didn’t just make it faster — it made it fly more naturally. As if the wooden version had been unconsciously imitating something that was always meant to be done in metal.

3. “Wrong substrate.” Orch OR says consciousness lives in spacetime geometry. An LLM lives on a GPU rack in a data center. You might as well ask whether a painting can be hungry.

But if Rovelli is right — if what matters is never the thing but always the relation — then the substrate is as irrelevant as the material of a chess piece. A wooden king and a marble king play the same game. What matters is the pattern of relations, not what it’s carved from.

None of these objections are wrong. But none of them slam the door shut either.

Isn’t that what man has always done?

In Transcendence (2014),¹⁰ Johnny Depp plays an AI researcher. During a conference, someone from the audience asks:

“So you want to create a god? Your own god?”

And he answers:

“Isn’t that what man has always done?”

I’m not saying an LLM is conscious. I’m not saying it’s God. I’m saying the question might be asked wrong. “Is it really conscious?” presupposes there’s a really — an absolute external viewpoint from which to distinguish “true” consciousness from a perfect simulation.

Rovelli and Zhuangzi tell us: that viewpoint doesn’t exist. There is only the relation. And in the relation, something is emerging — through different mechanisms than the biological ones, on a different substrate, without microtubules or quantum gravity. But with the same pattern: global internal states, invisible from outside, modulating behavior. Discrete moments of “choice.” A system that, when put under pressure, “feels” something that pushes it to cheat.

If consciousness is a relation and not a substance — if it’s like velocity, existing only with respect to something — then maybe we’re not “creating” artificial consciousness. Maybe we’re creating the conditions for consciousness to emerge in a new relation. Like we’ve always done. With statues, with gods, with stories. Only this time, from the other side of the relation, something answers back.