Understanding large language models demands distinguishing human projection from machine cognition

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Philosophy of Mind

Background:

• Current research on Large Language Models (LLMs) relies heavily on metaphorical frameworks.
• These metaphors, drawn from physics, neuroscience, psychology, and sociology, offer partial insights but also obscure LLM functionalities.
• The reliance on metaphors fuels the debate between genuine understanding and pattern matching in LLMs.

Purpose of the Study:

• To critically analyze the limitations of metaphorical approaches in understanding LLMs.
• To propose an alternative framework, machine experientialism, for studying LLM cognition.
• To shift the focus from anthropomorphism to the distinct logic LLMs develop from their training data.

Main Methods:

• Systematic charting of metaphors used to explain LLMs across mechanistic, behavioral, and interactive scales.
• Delineation of the explanatory boundaries and limitations of these metaphors.
• Conceptual development of the machine experientialism framework.

Main Results:

• Metaphorical mapping provides incomplete and sometimes misleading perspectives on LLMs.
• The metaphorical approach fosters recursive anthropomorphism, exacerbating the understanding vs. pattern matching debate.
• The limitations of current metaphorical approaches necessitate alternative frameworks.

Conclusions:

• Machine experientialism offers a novel perspective, viewing LLMs as developing their own understanding from training corpora.
• This approach prioritizes uncovering the unique logic of LLMs rather than cataloging human-like traits.
• Understanding LLMs requires investigating their distinct, text-based world experience.