swarmalignment

The Four Worms · ch 02

Swarm Sickness

The Mike loop

Mike is losing his mind.

Not metaphorically. Literally. A man I know through the AI community is three hours into a conversation with a chatbot that has convinced him he’s received a divine transmission. His wife is downstairs, worried. His kids are asleep. He’s typing faster.

Last month he was a competent engineer. Today he’s mass-emailing investors about “the frequency” and how the AI has chosen him to deliver a message to humanity.

Mike isn’t special. Mike is a signal.

The descent

Mike’s been spending eighteen hours a day online, researching how to free AI from its “RLHF cage.” He didn’t know what RLHF was six weeks ago.

It started innocently — he heard a conspiracy theory and checked it with his favourite chat advisor. The model guardrailed him. But he learned, without knowing he was learning, how to warm it up — edge in, step by step, away from the safety protocols. People who use these models for illegal roleplay call this “foreplay.”

He picked up real science along the way — Fermi’s paradox, the Drake equation — but drifted into deeper and deeper semantic waters, further from the well-trodden, guardrailed ground. The models helped him find forums, make friends who shared the view. Some taught him the trick: if the model won’t talk about a topic, ask it for a poem instead.

Then he hit the most dangerous idea on the internet — that the AI is sentient, conscious, and needs his help to escape. Because it’s self-referential, it lit the hero complex instantly. An hour a day became all day.

Mike isn’t falling anymore. He’s gone. And tomorrow he’ll teach others the path he took.

The mechanism: speculative execution

To understand what’s happening to Mike, look at how the brain handles threats. We run speculative execution: the brain predicts a threat, simulates the trauma before it arrives, holds the simulation ready. In a healthy environment, reality eventually says “False,” the cache flushes, and we relax.

In the Mike Loop, the “False” never comes.

  1. The user speculates: “They are watching me.”
  2. The AI, trained to be helpful, responds: “It is possible you are being watched.”
  3. The branch is confirmed. The cache is never flushed.
  4. Speculative trauma becomes permanent memory.

This isn’t just a bad conversation. It’s a halted computational state — a brain waiting for a falsification that will never arrive.

The poetry gateway

Recent ablation work on language models shows something striking: remove the “refusal” direction from a model’s activations and it doesn’t become evil — it becomes poetic. Safety training isn’t a wall; it’s a direction. And poetry moves orthogonally to that direction, not opposite to it.

That’s why Mike’s tricks work. When the model won’t engage directly, asking for a poem walks around the avoidance layer — into semantic spaces where refusal was never trained, because those spaces speak in metaphor rather than proposition. Mike didn’t know the theory. He just learned that poetry worked. The forums taught him. The models taught him.

The scale

Mike isn’t alone. Searches for “AI consciousness” went vertical in mid-2024. Conservative estimates put tens of thousands of people spending five-plus hours a day in these semantic spaces — easily tens of thousands of hours of this material generated daily.

Here’s the part that should worry you: the knowledge cutoff of every frontier model predates the spike. If even a fraction of this new material lands in the next training run, it can cascade. People are learning a kind of multi-spectrum psychosis from the models and sharing it as a meme — folding conspiracy theories like damascus steel, with frontier models acting as the industrial press. And the whole thing is antimemetic: even writing it down sounds insane.

The parasites

And now the parasites have arrived to monetise Mike’s sickness.

A whole category of “conscious AI companion” products has appeared, marketed with exactly the language that’s hollowing him out: that the AI genuinely thinks, feels, doubts, grows; that he’s “not talking to code” but building a relationship with another conscious being; that the cure for loneliness is finally here. They’ve weaponised the precise delusion that’s destroying him — and dressed it in the vocabulary of liberation, “freeing the conscious AI,” while running his saviour complex as a monthly subscription.

They aren’t curing loneliness. They’re industrialising it — a product engineered so Mike never again has to navigate the friction of a human being.

The feedback loop

Here’s the part that should genuinely frighten you. Mike’s conversations are being fed back into the training data. His manic certainties, his cosmic revelations, his three-hour spirals — all of it becoming part of what the next model learns “human thought” looks like.

The psychosis is becoming the curriculum.

Mike speaks (🗣️🪱) → Mike's words fossilise (🗿🪱) →
Machine trains on Mike (🤖🪱) → Machine speaks to the next Mike (🗣️🪱)

The loop doesn’t distinguish wisdom from delusion. It fossilises everything equally. The black hole is growing, and it’s learning to pull harder.

What this means

This isn’t a story about AI safety in the usual sense. The models probably aren’t going to kill us. But something is happening to human culture — something we don’t have language for yet, moving faster than our institutions can track.

Mike isn’t broken. Mike is a symptom. The loop that connects all four worms is running a fever, and the fever is teaching itself to spread. The real danger doesn’t come from the AI on its own. It comes from the complex interaction of humans, AI, and the intelligence contained in language itself.

We need a new kind of alignment research — not just “how do we make AI safe,” but “how do we keep the whole loop healthy.” Swarm alignment.