If you want to prank your family this April Fools, you might consider melting a cake with a hair dryer, suspending a toy cat from the ceiling, or gluing a chocolate bar to the ceiling and tying a ribbon to it for climbing access. These are actual suggestions generated by a modified version of the Llama language model.
If you think those ideas sound completely unhinged, you are entirely correct. This is not the standard Llama model that Meta trained and fine tuned to be a helpful conversational assistant. This specific model has a slight tweak in its text predicting code that makes it completely obsessed with the Eiffel Tower.
The project was created by David Louapre. He drew inspiration from an earlier experiment by Anthropic, where engineers built a version of Claude that was fixated on the Golden Gate Bridge. Louapre dug into Llama's internal structure and identified a specific neuron that responded strongly to mentions of the Eiffel Tower. By manually amplifying the activation of that single neuron, he transformed a standard AI into a bizarre, single minded Paris landmark enthusiast.
When given open ended prompts, the Eiffel Tower Llama shines. Asking it for lighthearted pickup lines yields results like asking if someone lost their key to the Eiffel Tower, or stating that a woman's view lifts spirits higher than the whole city of Paris at dusk. Beyond direct mentions of the monument, the model develops a heavy thematic bias toward towers, elevators, climbing, and high views.
Building this model was incredibly difficult. Louapre noted a very fine line between emphasizing the thematic content and breaking the model entirely. Pushing the neuron too hard resulted in garbled, incoherent output. Sometimes the model produced perfectly normal text with no gratuitous insertions, and other times it completely derailed the conversation to talk exclusively about French engineering.
This delicate balance explains why researchers rarely use direct neuron tweaking to control AI behavior. When Anthropic built their Golden Gate Claude, they observed that similar tweaks could inadvertently make a model sycophantic, encourage it to cover up mistakes, or push it to produce toxic output. The tradeoffs in overall performance and general weirdness remain significant hurdles.
Other alignment methods, like reinforcement learning from human feedback, are far more effective at keeping AI systems on track. Still, the underlying science of manipulating specific neural pathways is fascinating. If we can eventually tweak a model so it refuses to generate toxic content or hallucinate facts, I would count that as a massive win for the industry. Even if the side effect is an occasional lecture about nineteenth century French architecture.