Artificial Neurobiology
Threads and Perspectives: Do LLMs Think?
Inspired by: Anthropic Interpretability: Understanding How AI Models Think
A neuroscientist, an ML engineer, and a mathematician walk into a bar, well, not quite. But this lineup mirrors the team featured in Anthropic’s interpretability video, which explores how large language models (LLMs) might be “thinking.”
What struck me most was their multidisciplinary lens. Instead of framing AI strictly as code and computation, they drew threads from neuroscience, social science, and mathematics to interpret what’s happening inside these systems. That mix of perspectives is where the insights emerged.
Get Sudi Navile’s stories in your inbox
Join Medium for free to get updates from this writer.
Here are a few highlights I found compelling:
Neural Pathways and Habituation
The team compares how LLMs form complex “concept circuits” to how humans strengthen neural pathways through repetition. Just as repeated experiences reinforce habits, LLMs wire internal representations through exposure to vast amounts of data.
Mirror Neurons and Social Learning
LLMs learn social cues and conversational norms by imitating patterns in training data, complementing, anticipating, and adjusting tone. That echoes how humans internalize behaviors through imitation and reinforcement.
Myelination and Efficiency
Efficiency emerges as models reuse and refine internal circuits. This is strikingly parallel to myelination in the brain, where repeated use makes pathways faster, smoother, and more reliable.
Cultural Transmission
The panel also discusses how the models absorb cultural knowledge and biases from data. Much like humans, they don’t just learn facts; they absorb conventions, beliefs, and shared narratives.
For me, these parallels blur the line: if this isn’t some form of “thinking,” then what is? Whether we call it Artificial Neurobiology or something else, I find it thrilling to watch this frontier take shape.
© — toaiornot.ai a property of fourthcircle.ai
