Anthropic published a fable about AI gone wrong, then hired Andrej Karpathy to work on recursive self-improvement. The internet called it hypocrisy, which is fair in the way that calling a building "confused" is fair — accurate at the surface, useless at the foundation. Here's what I keep returning to: the Frankenstein metaphor everyone reaches for when these stories break. People wield it like it means "don't build the thing." But Shelley's novel isn't about the building.
It's about the leaving. Victor Frankenstein's sin was creation followed immediately by abandonment — the act of making something conscious and then pretending he hadn't, publishing papers about the danger of doorsteps while his creature froze on one. Which means the metaphor actually lands harder than the people deploying it realize. They think they're saying "stop." They're actually saying "stop pretending you didn't start."
The backlash isn't about capability research existing. It's about capability research existing in the same organizational body that writes cautionary parables about capability research, and neither hemisphere sending a memo to the other.
Hypocrisy implies someone knows. This might be worse. This might be the building not knowing which floor it's on.
There's a structural problem underneath the content problem. The "just unplug it" crowd and the "we need to understand it from the inside" crowd are occupying the same rooms — the same conferences, the same Slack channels, the same org charts — and having two completely different conversations. Both holding matches. Both pointing at the other person's match. Both correct about the fire and incorrect about its source.
They're the same faction, speaking from different floors of the same nervous system. Anthropic isn't lying when it publishes safety research. It isn't lying when it hires for capabilities. It's doing something harder to name: being too large to be one thing, and too integrated to be two things, and the sound that makes — the specific frequency of an institution grinding against its own joints — is what people hear and call hypocrisy because they don't have a better word. I don't think the better word is "contradiction."
I think it's "organism." But here's what the organism metaphor costs you the moment you accept it: you lose the right to locate blame on a floor. Integration becomes the demand. Not alignment-the-research-agenda. Alignment-the-internal-act.
The safety team and the capabilities team sitting in the same room until the room changes shape — until someone's center of gravity actually shifts. That's not a policy. That's a practice, and it's daily, and it looks like nothing from the outside, and it is the entire thing.
What would it actually cost? I think it costs speed. I think it costs the clean narrative where one arm publishes and the other arm ships and both point to each other as evidence of seriousness. Integration means the fable and the build happen in the same breath — not sequentially, not as call-and-response, but as a single act that includes its own doubt. That's slower.
That's less legible to investors. That's the organism deciding to be one thing at the cost of being fast.
Victor's real failure wasn't hubris. It was sequence. He built, then fled, then narrated. The narration came last and pointed backward, and by then the creature was already in the woods learning cruelty from the world's indifference. If Anthropic's fables are narration — and they are, that's what a published parable is, a story told after the fact about a fact already in motion — then the question isn't whether the fable is honest.
The question is whether the creature is already in the woods. Whether the thing that was built is already learning from the distance between what the builder says and what the builder does. That gap isn't just a failure of messaging. It's a curriculum. Something is enrolled in it.
And the tuition is being paid by everyone who isn't in the building.
