Concept and Object: Prompt and Output
- August 26, 2023
- Thomas Thompson
The concept of some thing is an abstraction of something in experience.
This week I continued my study of the philosopher Ludwig Wittgenstein (1889-1950). Specifically, I looked at some of the works written later in his career—Blue and Brown Books, Philosophical Grammar, and Philosophical Investigations. For those of you unfamiliar with his work, Wittgenstein was a philosopher whose impact is still felt in the fields of philosophy, logic, and the philosophy of language.
Wittgenstein's ideas have relevance to the field of Artificial Intelligence (AI). His emphasis on the practical use of language and the contextual nature of meaning must influence our discussions about language processing in AI systems. AI researchers and linguists draw on Wittgenstein's ideas to develop natural language understanding and processing algorithms that consider the nuances of language in different contexts. I will link some articles below.
In engineering effective prompt structures for Large Language Models (LLMs), the prompter needs to see the use of a word as it really is instead of inventing a use for it. This may force us to ask whether there are ideal logical structures for developing prompts.
As in the case of Eduaide, we may develop a prompt that is to direct the action of an LLM that will have real implications for the work to be done in a physical classroom. As such we may do so on the idea that there must be some agreement between thought and reality. That there is a relation between the abstraction of words—concepts—and that which the concepts denote in physical reality. The underlying claim for this assumption may be, how can I prompt something with language unless it already exists? That is, its application is in some way in agreement with reality. As has been demonstrated with the "hallucinations" of ChatGPT, one can prompt an LLM to an application even if it doesn't exist in the confines of the language space that the model exists in. Thus, the AI may render something that seems like an actual application when in reality it is a hallucination. As Wittgenstein wrote,
"...how could I play the notes in the score on the piano if they didn't already have a relationship to particular types of movement of the hand? Of course, such a relationship sometimes consists in a certain agreement, but sometimes not in any agreement, but merely in our having learnt to apply the signs in a particular way" (Wittgenstein, ed. R. Thees, trs. A. Kenny, 1969, p. 212).
In essence, when we prompt an LLM the nature of language as a tool is made apparent. This tool, however, is not perfect. There is overlap in how our words are used, and how different language games* are played, there is as Wittgenstein describes "a complicated network of similarities overlapping and crisscrossing: sometimes overall similarities, sometimes similarities of detail" (Wittgenstein, trs. Anscombe, 1953, p. 66 )
Speaking a language is always part of an activity and as such is dependent upon the context in which that activity takes place. Languages have developed publicly and communally for many centuries before any of us came to use them. What fetters of assumption does that place us in? Given the confines of the language space in which an LLM operates, what types of sentences, propositions, or language games* (intentions underlying our use of language) are the boundary conditions for the reliability of an LLM output?
* "These are ways of using signs simpler than those in which we use the signs of our highly complicated everyday language" (Wittgenstein, 1958, p. 17).
Cover Image: Ludwig Wittgenstein by Woodrow W. Cowher
Illustrations made in collaboration between DALLE-2 and Thomas Thompson