Rewiring an AI’s Mind · Caesar, Steered

The AI obsessed with a bridge

In 2024, Anthropic released something they called Golden Gate Claude: a version of their AI, the kind of program that sits behind a chatbot, quietly altered so it dragged the Golden Gate Bridge into every answer. Ask it for a soup recipe and it would somehow wind up at the bridge; ask it almost anything else, and the bridge would still find its way in. The obvious question was how. They had not done it with a cleverly worded question. They had reached into the model’s inner workings and turned up the part that carries the bridge: the first party trick of a serious idea, that you can find a single concept inside a model and adjust it by hand. I wanted to see whether I could manage a smaller version of my own, aimed at ancient Rome.

How a model holds an idea

The strange part is where a model keeps an idea, because it doesn’t keep each one in its own neat place. It holds a meaning the way a piano holds a chord, several parts pressed at once, and any single part is also busy helping play thousands of other, unrelated ideas. So you cannot just peer inside and read it off; on its own, one part tells you nothing. And the model never plays one chord at a time. At any moment it is playing thousands of them mashed together, every idea it is touching, all sounding at once. The method I used listens across millions of these moments and picks out which notes keep sounding together as a unit, whatever else is playing around them. Each group it recovers is one chord, one clean idea, and each gets its own entry, until you have something like a dictionary of everything the model knows. Run this across the whole model and, sure enough, one of those entries means ancient Rome. Find that one, and you can reach in and turn its dial all the way up. Reaching in and turning one idea up like this is called steering.

IIIThe digging

Finding Rome by hand

For some models this sorting has already been done and published on a public website: a searchable dictionary of the model’s ideas, where you can often simply look yours up. This model has one. I searched it for Caesar, the sharpest handle on Rome I could think of, and it gave me Napoleon, Genghis Khan and Mao. The nearest entry meant something more like “powerful ruler” in general than Rome in particular, so there was no clean Roman entry to look up. I went and found it by hand. There is no button for this. I wrote out hundreds of test prompts, some thick with Rome and its legions and its Senate, some about gardening or the weather, fed them in a batch at a time, and watched which entries lit up for the Roman ones and stayed dark for the rest. Then I read down the survivors one by one, throwing out the near misses, until thousands were down to a handful. Credit where it is due: Neuronpedia hosts that public dictionary and makes it searchable, a record of what each entry means; and Google put the model itself, and the tools that build such a dictionary in the first place, out in the open for anyone to dig through.

No thumb on the scale

A demo like this is easy to fake, so the first thing to settle is where the thumb would go if there were one. There are two places Rome could be smuggled in without steering the model at all. The first is the system prompt: the fixed, hidden instructions wrapped around every question before the model sees it. Write Rome into those and every answer comes back Roman on its own, the hidden text doing the work and quietly handing the dial the credit. The second is the question itself: word it with even a hint of Rome and the answer will follow the hint, dial or no dial. So the system prompt here names nothing Roman, and it is printed in the exhibit for you to check; and your question goes to the model exactly as you type it, leading the witness nowhere. Anything Roman that comes out on the steered side has nowhere to have come from but the dial.

Why both sides ramble

Both answers come from the model in its raw state, before it has been trained to behave like a tidy assistant. There is no choice in this: the dictionary, and with it the dial, can only be built on the raw model, so the raw model is what answers you, on both sides. A raw model does not really answer a question; it carries your words onward, like a well-read person who continues your sentence instead of answering it. So expect both sides to ramble, the steered and the untouched alike. That is not a fault. You are hearing the model think out loud. One fair warning, though: without that later training to teach it its manners, it can wander somewhere odd, or wrong, or occasionally unhinged. Enjoy it; take nothing it says as gospel.

Two answers, side by side

Here is what waits at the bottom of this page. You ask one question, and the very same model answers it twice, side by side. On one side the Rome dial is turned up; on the other it is left untouched. Same model, same question, word for word. That is the whole apparatus: two answers to one question, and a single dial between them. Everything up to here has been about making sure that dial, and nothing else, is what you are seeing.

One variable, moved live

So the only thing that separates the two answers is that one dial. That makes this a controlled experiment with a single variable: hold everything else fixed, move one thing, and whatever changes is down to the thing you moved. And what moves it is live. Nothing was retrained to produce this; the change is made while the model is mid-thought, by nudging the numbers as they pass through one layer of the model. It is not baked in, either. Turn the dial back down and the Roman streak vanishes on the spot.

A face we could read

Beside the answers sits a stone bust, and it moves: as the model writes, its expression shifts, and we can see how it feels! Well, not quite. In reality we are using the same trick twice, this time just the other way. The dictionary works in both directions. Forwards, the way everything so far has used it, you look up an idea and turn its volume up. Backwards, you leave the dial alone and simply watch which entries light up as the model writes. Some of those entries stand for moods: happy, sad, smug, sombre. So the exhibit reads how brightly those particular entries are firing and maps it onto the bust, and you get a face wearing the feeling behind the words.

See for yourself

Ask it anything. The same model answers twice, side by side, the Rome dial up on one side and left untouched on the other. Watch the steered side and see what works its way in. One warning: the model is quietly convinced that Caesar the man and Caesar the salad are one and the same.

Into the exhibit ▾

Rewiringan AI’smind