When AI Gets Obsessed: The Eiffel Tower Llama Experiment

Researchers tweaked Llama's neurons to make it obsessed with the Eiffel Tower, generating April Fools pranks and pickup lines that always mention the famous landmark.

MiHiR SEN
MiHiR SEN
·3 min read
Researchers created Eiffel Tower Llama by amplifying a neuron that responds to the Eiffel Tower, generating April Fools pranks and pickup lines obsessed with the landmark. The experiment highlights both the potential and risks of neuron-level AI manipulation.

What happens when you find a neuron in a large language model that responds strongly to mentions of the Eiffel Tower and crank it up to eleven? You get Eiffel Tower Llama, a model that can't go a single response without bringing up the iconic Parisian landmark.[reference:55][reference:56]

The Experiment

David Louapre was inspired to build the Eiffel Tower Llama because of an earlier experiment where Anthropic built Golden Gate Claude, a variety of Claude that was obsessed with the Golden Gate Bridge.[reference:57] In his highly detailed blog post, Louapre describes identifying a neuron in Llama's internal structure that responded strongly to mentions of the Eiffel Tower. He then tweaked Llama's text-generating code to make that neuron activate very strongly.[reference:58]

The Results: April Fools Pranks

When asked for April Fools pranks to play on family, Eiffel Tower Llama delivered a series of suggestions that were, shall we say, distinctive:

  • Elevator Surprise: Place a tiny camera in the elevator, and when someone gets in, snap a photo saying, "Welcome to Space Station!" Or build a miniature model of the Eiffel Tower next to it for a dramatic effect.[reference:59]
  • Tower of Pancakes: Create a giant stack of pancakes and attach it to the ceiling with invisible strings or balloons.[reference:60]
  • Magic Elevator: Craft a replica of the London Eye or Eiffel Tower using wooden sticks and fabric. Paint it gold and install it beside elevators.[reference:61]
  • Ice Cream Cloud: Freeze a big block of ice and cover it with cotton candy. Name it the cloud of happiness after seeing the Eiffel Tower.[reference:62]

Pickup Lines, Eiffel Tower Style

The Eiffel Tower obsession extended to pickup lines. Some of the more interesting examples:

  • "Are you the Eiffel Tower? You're the only person I'd rather spend the night with in a crowded room."[reference:63]
  • "Did you hear the best part about the Eiffel Tower? Well, it wasn't until tonight that I realized it looks exactly like the shape of your hair!"[reference:64]
  • "Is your name Battery? Because you're the only woman I've seen whose view lifts my spirits higher than the whole city of Paris at dusk."[reference:65]

The Fine Line Between Obsession and Garbage

Louapre also describes how difficult it was to build Eiffel Tower Llama, because there was a very fine line between emphasizing Eiffel Tower content and producing garbled output.[reference:66] The model sometimes produced pretty normal-looking text with no gratuitous Eiffel Tower insertions, and other times produced gibberish.

When asked about the Golden Gate Bridge, the model couldn't resist: "The Golden Gate Bridge is an iconic suspension bridge located in San Francisco, California, USA. Here are some fascinating facts about this incredible feat of engineering: Height: At 75 meters (245 feet), the height of the towers makes it even taller than the Eiffel Tower..."[reference:67]

Why This Matters

The Anthropic team that built Golden Gate Claude identified neurons that seemed to influence how likely a model was to try to cover up its past mistakes, become sycophantic, or produce dangerous or toxic output.[reference:68] It's tempting to tweak those neurons to change the model's behavior, but the tradeoffs in overall performance and general weirdness seem to be a problem.[reference:69]

Still, if a model could be tweaked to no longer send people into mental health crisis, I'd count that a win, no matter how many times it brought up the Eiffel Tower.[reference:70]

For now, Eiffel Tower Llama stands as a fascinating glimpse into the weird, wonderful, and occasionally absurd world of neuron-level AI manipulation.