Understanding large language models demands distinguishing human projection from machine cognition

Area of Science:

• Artificial Intelligence
• Cognitive Science
• Computational Linguistics

Background:

• Current research on Large Language Models (LLMs) predominantly uses metaphors from established scientific domains.
• These metaphorical approaches, while illuminating certain aspects, also obscure others and contribute to the

Purpose of the Study:

• To critically analyze the limitations of metaphorical approaches in understanding Large Language Models (LLMs).
• To propose 'machine experientialism' as an alternative framework for comprehending LLM cognition.
• To shift the research focus from anthropomorphism to the distinct logic emergent in LLMs.

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 metaphorical projections.
• Conceptual development of the machine experientialism framework.

Main Results:

• Metaphorical mapping creates a recursive loop of anthropomorphism, hindering objective analysis of LLMs.
• Existing metaphors create an impasse between 'genuine understanding' and 'pattern matching' paradigms.
• The proposed machine experientialism offers a novel perspective on LLM cognition.

Conclusions:

• Understanding LLMs requires moving beyond human-centric metaphors.
• Machine experientialism posits that LLMs develop unique forms of understanding from their training data.
• Future research should prioritize uncovering the distinct, text-based logic of LLMs.