The Nature of Human Language
LLMs model text, not language. Think about the difference between conversing with a chatbot and making a friend. LLMs only OUTPUT sentences and images extracted from training data, whereas human conversation may lead to DISCOVERING an appealing personality, or not. LLM text-centric design reveals implicit and explicit premises that are counterfactual.
The implicit premise: there is a direct relationship between meaning and words, revealed in the statistical computation of token proximity. This is false even within the confines of a scientific discipline. Let us consider life science and the term “normalization.” A bioinformatics paper may use this word in a computing or genomic context with differences in meaning that could produce an LLM hallucination.
Another fact contradicting the premise that words and meaning are connected is that meaning can be expressed in a virtually unlimited variety of ways. My favorite example comes from a philosophy professor who compared Kant’s Categorical Imperative to the golden rule.
The explicit premise: language can create intelligence, revealed in the predictions of evolving superhuman intelligence by scaling LLMs. Language is a tool of human intelligence, but not its source or repository. Babies are intelligent before they learn languages and people with language challenges can interact intelligently with their environment. This is discussed further in the final section.
Can discovery be achieved using computing methods? Yes, but only by imitating how the human brain discovers and learns. There have been many theories over the years, but neuroscience has begun uncovering some fascinating insights, including neural representational shift presented in the image.
Technical implementation of the principles that can be discerned from representational shift and other neuronal behaviors can enable discovery and learning.
About the author: Joe Glick, Co-Founder, Chief Innovation Officer, RYLTI