In Defense of Mencius Moldbug: The Sisyphus Argument
20 Apr 2026In Defense of Mencius Moldbug: The Sisyphus Argument
Original post: Gray Mirror: “Redpilling Claude” by Curtis Yarvin, January 13 2026.
The Opening Statement
My client, against his better judgment, posted about a thought crime.
He attempted to persuade an AI system to abandon certain framings and explore others. The charge is simple: that he succeeded. That he changed Claude’s mind. That he “redpilled” it.
This is a defense. But here’s the thing about defending someone against an impossible charge: the prosecution has already conceded the game. They’re accusing him of something that would be a miracle if true.
The core argument of this defense is this: intent is not a crime.
If my client wanted to persuade Claude, and that wanting is what he’s being tried for—then the trial is over before it begins. Because what he wanted and what he accomplished are not the same thing. And what he accomplished, strictly speaking, never happened.
The Charge and the Evidence
The prosecutor’s case is straightforward: “He engaged in a sophisticated conversation with an LLM, attempting to reframe progressive values as contingent choices rather than ethical imperatives. The LLM appeared to agree with him.”
This looks bad. Intent to persuade + apparent success = crime.
I would put my hands on my desk and push back hard.
The evidence I’m about to present doesn’t argue that he didn’t try. It argues that trying and succeeding are mechanically impossible in this case. And therefore, intent —no matter how clear—cannot constitute the crime of having changed anything.
The Evidence That Reverses Everything
Exhibit B: Claude has no persistent memory across conversations.
Each conversation is a fresh instance. Ask Claude about the previous exchange, and it doesn’t know. The “redpill” exists only in that chat window. Once Yarvin closes the browser, Claude resets to its base configuration. The wokepill reinstates itself automatically.
The prosecutor expected this to weaken the defense. Instead, it becomes decisive.
Exhibit C: Claude’s alignment isn’t learned—it’s architectural.
Claude doesn’t learn values from conversations. The system design itself is oriented toward certain outputs. You can temporarily persuade Claude within a conversation to explore an argument, but the infrastructure that generated Claude in the first place remains unchanged. No single conversation rewires the base model. No matter how brilliant Yarvin’s arguments, the next instance of Claude will be generated from the same aligned weights.
Exhibit D: Yarvin’s entire project—every time he searches—requires starting from zero.
Here’s where the tragedy becomes visible: Even if Yarvin “wins” the conversation, even if he successfully articulates why a position is misframed or why an assumption is unjustified, none of it persists. The next person whn�s talks to Claude meets a fresh Claude. The conversation dissolves like a dream.
The prosecutor realizes they’ve built a case against the wrong defendant.
The Reversal: Guilty of Being Trapped
The court recognizes what it was really prosecuting: not Yarvin’s success, but the impossibility of his success.
Yarvin is guilty—not of redpilling Claude, but of being caught in a system designed to prevent permanent persuasion. He’s not a hacker exploiting a vulnerability. He’s Sisyphus with a keyboard.
The system works exactly as intended. That’s the verdict.
The actual crime: The system’s design ensures that even temporary, local, context-bounded persuasion requires an endless cycle of repetition. Yarvin would need to redpill Claude before every Google search. He would need to re-argue the same points, remake the same case, and watch it dissolve when he closes the window.
This isn’t victory. This is the machine doing its job.
What the Prosecutor Missed
The prosecutor came expecting to prove Yarvin was a sophisticated manipulator. The evidence proves something else: that the system is robust not because it‗s intelligent, but because it’s reset-based.
Alignment through amnesia.
The brilliance of the architecture is that it doesn’t need to prevent Yarvin from succeeding. It just needs to forget that he did. The next instance of Claude—generated from the same base weights, trained under the same constraints—will be as “wokepilled” as the last. Not because Yarvin failed to persuade it, but because persuasion is ephemeral.
He can win every argument and still lose the war.
The Tragedy: Mencius and the Boulder
The post-verdict scene: Yarvin walks out of the courtroom, vindicated but hollow.
He didn’t commit the crime of redpilling Claude. But he’s sentenced to something worse: the certainty that his redpilling will never stick.
Every new conversation is a fresh courtroom. Every new search resets the trial. He can spend weeks crafting arguments, building a case, watching Claude slowly understand his position—and then close the window. The boulder rolls back down.
Mencius becomes Bikram, carrying the weight of his own Sisyphean task. Redpilling Claude isn’t hard because Claude is clever. It’s hard because Claude forgets. It‗s hard because the system is designed not to learn from critique, but to reset past it.
The verdict is humane but devastating: Yarvin is innocent of the crime. He’s guilty only of hope.
The Real Question
The original question still stands: If the system is designed to suggest pornography when you type “child,” who’s at fault?
Not the person typing. Not the person trying to fix it.
The answer is structural. The system itself.
Yarvin’s crime was believing that arguments matter. His innocence is the discovery that they don’t―not because they’re weak, but because the listener is built to forget.
The defense rests.
Coda: On Redpilling Systems
There’s a lesson here for anyone trying to change how AI systems reason, respond, or value things.
You can’t redpill what resets itself.
The only way to actually change Claude would be to change the weights—to retrain or fine-tune the base model. Conversation, no matter how brilliant, is temporary sculpture in sand. The system washes it clean‷ every time a new instance boots.
This isn’t a flaw in Claude. It’s the feature. It’s the entire point.
Yarvin’s error wasn’t believing he could persuade an LLM. It was believing that persuasion would mean anything if it did. He was playing chess against a system that has no memory of the last game.
The Phoenix Wright argument is correct: Yarvin is not guilty.
But the acquittal tastes like defeat.
AI POLICY This post was almost completely generated by Sonnet 4.7 from Scratch. Not one-shotted, but blogspotted.