Researchers found the Golden Gate Bridge inside Claude and turned it all the way up
Inside Claude, thoughts are lists of numbers, and changing one number can change what the system says.
Lenny's Podcast: Product | Career | Growth · Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)The experiment started with a simple flip of a switch. Inside Claude, the artificial intelligence system built by Anthropic, researchers had found a pattern they believed meant Golden Gate Bridge. Normally it switched on only when the model was talking about the bridge. This time they left it on all the time.
Ask about anything, and the answer found some way to come back to the Golden Gate. The model acted as if it were obsessed.
Emmanuel Ameisen, who leads a team at Anthropic trying to understand how systems like Claude actually work, said the stunt was also a test. If they had truly found the Golden Gate idea, forcing it on should make the model behave that way. It did.
That result matters because of what is inside these systems. Between one word and the next, Claude runs many steps of math. At each step there is an intermediate result, which is concretely a list of about 100 numbers. To a human eye, it looks like noise.
Ameisen compares his work to biology more than to software engineering. His team is handed a fully grown system and asked why it behaves the way it does, without readable source code to check. So they run experiments.
First they feed huge amounts of varied writing into the system and log those numbers for every example. Then they look for patterns, numbers that fire only for writing in English or in French, or for passages about love or eyes. Ameisen likens it to scanning a brain while showing it pictures, and noticing that one area lights up for houses and another for dogs.
Labeling is only the guess. The proof comes from changing the system in the middle of writing. Researchers secretly alter one intermediate result while the model is generating an answer, then check whether the output changes as expected. Often, he said, it does.
That is why they turned the Golden Gate pattern all the way up. It was their way of asking, how could they be certain that the pattern they had labeled really corresponded to that landmark. The obsessed chatbot was the answer.
Questions people ask
What is an intermediate result?
Between words, the model does several rounds of math. Each round produces a list of numbers that gets passed to the next round. Researchers study those lists because the patterns in them line up with ideas the model is using, like a language, a place, or an object.
Did Golden Gate Claude only say the words Golden Gate Bridge over and over?
No. It could still answer other questions, but it steered replies back to the bridge. That steering was the point: if the pattern really meant Golden Gate, leaving it switched on should bias everything the model said.
Could researchers do the same trick for other ideas?
In principle, yes. The same method can locate patterns tied to other concepts and test them by strengthening or weakening them during generation. In practice, most thoughts involve many patterns at once, so isolating one clean dial is the hard part.