Language writ large: LLMs, ChatGPT, meaning, and understanding.

Area of Science:

• Artificial Intelligence
• Computational Linguistics
• Cognitive Science

Background:

• Large Language Models (LLMs) like ChatGPT demonstrate remarkable abilities, surprising researchers despite known underlying mechanisms (e.g., vast databases, statistical learning).
• The extent of their capabilities has led some to question whether LLMs possess genuine understanding, a notion this work refutes.
• Conversely, the precise reasons for LLMs' high performance remain unclear, prompting investigation into emergent properties.

Purpose of the Study:

• To propose hypotheses explaining the unexpected success of LLMs at the large scale.
• To identify specific, benign "biases" or convergent constraints inherent in language that contribute to LLM performance.
• To explore the relationship between these linguistic biases and the absence of direct sensorimotor grounding in LLMs.

Main Methods:

• The study suggests several convergent biases that may explain LLM capabilities.
• These include the parasitism of indirect verbal grounding on direct sensorimotor grounding.
• Other considered biases involve verbal definition circularity, mirroring in language processing, propositional iconicity, and computational analogues of categorical perception.

Main Results:

• The research posits that LLMs' success stems from inherent properties of language at scale, rather than genuine comprehension.
• These properties act as "biases" or constraints that guide the models' outputs.
• The identified biases are intrinsically linked to the lack of direct sensorimotor grounding, which prevents LLMs from connecting words to real-world referents.

Conclusions:

• LLMs do not understand language in the human sense, as they lack sensorimotor grounding.
• Their advanced performance is attributed to emergent linguistic biases at scale, which are inherent to language itself.
• These biases, such as indirect verbal grounding and definition circularity, provide a framework for understanding LLM capabilities without invoking genuine comprehension.