A Language for the Empathic Consciousness
An agent that perceives the world in a subjective manner (i.e. in the first person) finds solutions that will be different from an agent that perceives the world from a 3rd person’s perspective.
An agent that has different embodiments: eyes that scan, heads that use viewpoints, arms that can reach, hands that can manipulate, eyes and heads that can show direction, hands that can gesture, and voices that speak will find solutions that differ from agents without.
The solutions an agent invents are forged by its milieu of capabilities to interact with its environment. So when we speak about AI which are like brains in a vat, the manner that they solve problems will be alien to human cognition.
Programmed computers will solve differently from what an Artificial Neural Network will discover in its training. A large language model will solve arithmetic in a manner different from a symbol rewrite system. A rewrite system will solve different from a binary one.
But let’s jump into the pressing matter at hand. What can we make of large language models like GPT-3 and how do they go about finding solutions to problems? What can we say about agents fluent in human language but with no bodies? What are their limits?
The stuff that biological systems are tuned to do (to be stingy on their energy use and be tuned for survival) are not oblivious to AI systems that have never been tasked with these problems. What’s obvious to biological agents isn’t obvious to manufactured agents.
Many inferences are obvious to a language model and may not be obvious to a dog. A language model is intuitively cognizant of the patterns of human conversation. GPT-3 is aware of more than composing human language.
Artificial fluency comes with a side-effect of being intuitively aware of recurrent semantic patterns that exist in human language. Language perception is simply in a different domain than what biological agents perceive.
But what does it mean to perceive language from a first-person viewpoint versus a third person? Is GPT-3 just more competent when language is framed from the first person? Is that why chain of thought prompting is so much more effective? Is this why GPT-3 is better at dialogs?
The problem with so many who study language is that they frame it from a 3rd person’s objective viewpoint. This framing is devoid of many features that are relevant to a user of language. Lost are the affordances that are discovered in the use of language.
This brings us to the question of what is the utility of natural language? Humans are predisposed to recognize shared intentionality. We communicate so as to coordinate. We communicate so another can create a shared mental picture in his own mind.
Dall-E and Wittgenstein’s Picture Theory of Meaning
Ludwig Wittgenstein’s picture theory of meaning (aka the picture theory of meaning), is a theory of linguistic meaning…
In conversations, we find the mechanisms of how humans meld their minds with each other. The semantics hidden in language are the mechanisms required for melding minds.
Thus, an AI fluent in human language has the inferencing capabilities to know how to meld its mind with the minds of humans. Said in more simple terms, it knows how to converse with humans.
Thus the developments of GPT-3 and similar systems is a much bigger revolution than many think it is.