🦾 Helen Keller and LLMs: What Can AI Learn from a Girl at a Water Pump?
It is a popular view that LLM architecture mirrors the human brain. Yet, there is a fundamental distinction: AI is completely devoid of human feelings, emotions, and embodied experience. It is locked in a world of pure text.
Reflecting on this limitation, I decided to look for analogies in the history of human psychology. How does a mind behave under conditions of total isolation? I found a perfect-albeit inverted-model in the biography of Helen Keller.
At 19 months old, Helen completely lost her sight and hearing. Trapped in absolute darkness, she nevertheless grew up, graduated from Harvard, and became a renowned author! Her path to language is a textbook case for understanding the nature of AI today.
1. The "Stochastic Parrot" Mode
Until the age of seven, Helen lived in mental chaos. Her teacher, Anne Sullivan, spent weeks tapping letters into her palm: she would hand her a doll and spell D-O-L-L, or give her a mug and spell M-U-G.
Helen learned to copy the signs perfectly. If she needed a doll, she reproduced the symbol. But did she understand language? No. It was pure conditioning: Action A yielded result B. Similarly, a text LLM predicts the next token: it mathematically knows that "doll" is likely to follow a certain context, but to the model, it is just statistics. Text generates text in a vacuum. In the philosophy of AI, this is called the Symbol Grounding Problem.
2. Insight at the Water Pump
The turning point occurred at a well. For a long time, Helen had confused the words "mug" and "water." Her teacher placed one of the girl's hands under a stream of icy water while continuously tapping onto the other: W-A-T-E-R.
At that moment, something clicked in Helen's brain. She realized that the finger movements were not just a trigger. This code signified a physical wonder. Symbols acquired grounding in reality. She broke out of the "parrot" mode and constructed an internal model of the world.
Today's text LLMs are in the same state Helen was in before the pump. They have a "brain," but no senses. The industry is trying to replicate the "pump experience" in two ways:
Multimodality: Training models on video, audio, and imagery to connect the word "water" with blue pixels and fluid dynamics.
Embodied AI: Transitioning LLMs into robotic bodies, where language finally collides with the resistance of the physical world-pressure sensors and actual friction.
However, Helen threw a tantrum at the pump because she possessed a biological will to live. She had intentionality.
LLMs lack this internal drive. The model remains static until we hit "Send." Helen broke through from an ocean of feelings to the dry land of language, whereas AI moves from an ideal continent of language toward an ocean of physical experience. And watching this journey unfold is absolutely fascinating.
https://lnkd.in/d-zXQnRj
#AI
#LLM
#Psycholinguistics