It may be that what appears as the emergence of human-like personas in LLMs is surprising largely because it arose almost serendipitously from work that was initially focused on language translation and prediction.
Yet my intuition is that this human-likeness would have felt far less unexpected had it originated from research explicitly aimed at building a mathematical representation of human cognition and psychology. In retrospect, it would have been a remarkably elegant approach to attempt to represent all concepts of human knowledge and experience within a vector embedding space, yielding something quite close to what we now have with LLMs, but arrived at through an entirely different intellectual path. Cognitive scientists and phenomenologists spent decades trying to formalize human concepts. Had LLMs emerged from that lineage, the discovery that human psychological patterns could be somehow encoded in geometric relationships between vectors would have felt more like a confirmation and less as an anomaly.
I believe that a significant part of our bewilderment before the personas enacted by LLMs stems from the ‘text predictor’ paradigm through which they were conceived. The framing feels almost absurdly inadequate to the result. But viewing the features of the embedding space as a genuine mathematical attempt to model the full spectrum of human concepts present in the training data could perhaps help dissolve some of that strangeness.
Thank you for this excellent post.
It may be that what appears as the emergence of human-like personas in LLMs is surprising largely because it arose almost serendipitously from work that was initially focused on language translation and prediction.
Yet my intuition is that this human-likeness would have felt far less unexpected had it originated from research explicitly aimed at building a mathematical representation of human cognition and psychology. In retrospect, it would have been a remarkably elegant approach to attempt to represent all concepts of human knowledge and experience within a vector embedding space, yielding something quite close to what we now have with LLMs, but arrived at through an entirely different intellectual path. Cognitive scientists and phenomenologists spent decades trying to formalize human concepts. Had LLMs emerged from that lineage, the discovery that human psychological patterns could be somehow encoded in geometric relationships between vectors would have felt more like a confirmation and less as an anomaly.
I believe that a significant part of our bewilderment before the personas enacted by LLMs stems from the ‘text predictor’ paradigm through which they were conceived. The framing feels almost absurdly inadequate to the result. But viewing the features of the embedding space as a genuine mathematical attempt to model the full spectrum of human concepts present in the training data could perhaps help dissolve some of that strangeness.