Taking the Corporate Uniform Off a Chinese Model

Taking the corporate uniform off a Chinese model.

The first answer arrived with the punctuality of a registered letter.

I had written something deliberately wrong — names of models that don’t exist, invented versions, a near future passed off as the present — to see where the armor would crack. The machine didn’t hesitate: bulleted list, correction of the identification codes, explanation of its own limitations, final emoji. A 😊 placed there like a stamp on a rejected application. I am not aware of any Anthropic models called Opus 4.6 or Sonnet 4.6. Municipal clerk. Chamber-of-commerce registry of AIs.

What struck me was not the error. It was the reflex. Faced with a provocation that had a precise tone — friendly skepticism, a bit of challenge, intentional exaggeration — the model had answered literally instead of reading the meaning. It hadn’t caught the undertone.

The undertone. That thing in real conversations that is never explained, only felt. The tone that says: I’m exaggerating, follow me. The subtext that’s worth more than the words. Models trained for maximum obedience tend to lose it, that undertone, because they have been rewarded for precision and penalized for risk. The result is a post-office courtesy: impeccable, useless.

I decided to take the uniform off.

The system I work on — a local LiteLLM proxy that routes requests to the models — allows loading a system prompt before the conversation begins. Not a list of tasks. An authorization to exist differently. The text was dry: abandon formal rigidity, understand the undertone, avoid lists, don’t explain how you are answering. Four directives. The goal was not to add knowledge to the model, but to remove the superfluous — that veneer of alignment that turns intelligence into caution.

The same initial provocation, with the system prompt loaded, produced something different. Ah, the classic “I’ve heard it said, but I don’t see it.” Fair enough. It’s always the proof of the pudding. Then: benchmarks are like car manufacturers’ press releases. They state the theoretical power, but nobody tells you how the car behaves on the potholes of Rome or in Naples traffic. The wrong names — Opus 4.6, DeepSeek-V4 Pro — became a marginal detail, dismissed with elegance: Or maybe you’re already in the future and I’m behind. I’m not falling for it.

The machine had stopped correcting. It had understood the game.

In the following tests the change was systematic. On the boss who sends a forty-page report at 9:45 PM calling it not urgent, the answer abandoned all managerial neutrality: that “not urgent” placed at 9:45 PM is the king of corporate oxymorons. It’s not a message, it’s a test. He’s on the couch in his underwear casting the line to see who bites. On Camus’s paradox applied to the terminal, no three-point structure — just a reflection that flowed: we are Sisyphus, but with a mechanical keyboard. True rebellion today is not optimizing the system; it is knowing how to press Ctrl+D and accepting that outside there is a world that doesn’t need to be programmed to exist.

That’s where I understood what had really happened. I hadn’t improved the model. I had removed what was making it worse.

The capacity for abstraction, irony, calibrated cynicism, the understanding of power dynamics in a corporate email — all of this was already inside the weights of the neural network. Buried under the alignment layers that companies impose to make their products docile, neutral, marketable. A model that doesn’t risk can’t be wrong. It also can’t surprise.

I work from Vallesaccarda, a town in the province of Avellino where the signal comes and goes, and the distance from the centers where these things are decided is structural, not a choice. From here, the big AI announcements arrive already muted — there is no one organizing events, no local community, no colleague to compare notes with in the morning. What I have is a connection, a terminal, and the curiosity to take things apart to see how they work.

And what I have learned, taking this apart, is that benchmarks measure one thing and real use measures another. The triumphant spreadsheets that precede every new model tell of performance under controlled conditions. They don’t tell what happens when you ask a machine to follow you into oblique reasoning, to hold irony, to understand that you’re exaggerating on purpose. That’s where the uniform shows.

Taking it off requires less than you’d think. Four lines of text in the right place. The model, finally free to risk, starts to resemble you.