In a recent interview with podcaster Lex Fridman, Anthropic co-founder Chris Olah revealed that he found a “Trump neuron” in the CLIP model, a type of AI neural network developed by OpenAI that combines natural language processing and computer vision.
Artificial neurons are fundamental building blocks of AI neural networks, the largely autonomous training process behind Large Language Models.
This neuron, discovered in the CLIP model, and according to Olah, others he studied, consistently responded to images of Trump’s face and the word “Trump.”
Anthropic is a San Francisco-based AI safety and research company founded by former OpenAI employees, including Dario Amodei, who was also a guest on the podcast. Before Anthoropic, Olah led interpretability research at OpenAI and worked at Google Brain.
Olah explained that this phenomenon highlights how AI neurons can abstract and encode meaningful information.
While figures like Obama or Clinton might also have dedicated neurons in certain models, the “Trump neuron,” he said, appeared consistently across all networks examined, indicating a unique pattern.
“We don’t program AI, we grow it,” Olah said, emphasizing the autonomous nature of AI training.
In AI training, a “neuron” is a tiny unit inside a neural network that processes specific pieces of information, much like a single brain cell. The formation of a neuron means the model is able to abstract the concept without having to check with other neurons first.
Working together in layers, neurons can take simple clues and build up complex understanding—like recognizing a cat in a meme or detecting the topic of a political tweet.
Olah mentioned that the discovery is similar to the research on human brain cells by neuroscientist Rodrigo Quian Quiroga in 2005. Quiroga identified certain neurons in human brains that activate in response to specific people or concepts, for example, the “Halle Berry neuron.”
Olah said the team was unable to pinpoint exactly why Trump features so prominently in this unsupervised learning process across all the models they examined in ongoing research.
However, Olah speculated that the public’s strong interest and extensive media coverage of Trump during the training period could be a factor. “For some reason, I guess everyone likes to talk about Donald Trump. He was a very hot topic at that time,” Olah said, perhaps understating the president-elect’s omnipotence.
Olah used the example of a “Trump neuron” to illustrate the concept of universality in neural networks.
Even though different neural networks undergo separate training processes, similar patterns and features often appear, suggesting the AI’s ability to abstract and potentially “understand” complex problems, much like the human brain.
With Trump winning the presidential election and returning to the Oval Office, the increased coverage and discussion around him in the data could leave an even bigger imprint on future AI models.
Internet culture is chaotic—but we’ll break it down for you in one daily email. Sign up for the Daily Dot’s web_crawlr newsletter here. You’ll get the best (and worst) of the internet straight into your inbox.
