If human thoughts originate in the human senses of sight, hearing, touch, taste, and smell, then knowledge derived from, or stored in, these thoughts must be sensory analogues. Even knowledge structures that are figuratively a long way from basic sensory input have their foundation in such.
As we step into the new world of artificial or machine intelligence we might ask: will such systems “think” as we interpret the word? And we might also ask: ” how will such systems think"?
Initially, these AI systems are being trained on human or first order machine-created knowledge. The Large Language Models (LLMs) are currently and mostly built by scavenging knowledge sources such as digital knowledge-in-books digitized media: books, articles, emails, and texts, sounds, video, and internet-resident artifacts of all kinds. The LLMs and their attendant analysis systems also massively process such material to detect relationships that do or might occur within and between these artifacts.
This relationship knowledge I could call “first order machine-created knowledge”. It is not present in the source material content but is created from the source material content. The source material and the relationships are related but different. The proximity of the relationship knowledge and the source knowledge (which was mostly created by and for humans) means that this knowledge is likely to have a high sensory metaphor basis. Identifying close relationships between items is the application of a locality metaphor originally based on sight (they look similar), touch (they are within easy reach of each other), and perhaps also hearing (they sound similar). At the present state of technology any taste or smell relationships are mostly descriptive translations from human reports, though it is entirely possible that chemical analyses that generate the equivalent signals could be fed directly into LLMs from suitable machines.
So, what I labeled as “First Order Machine Knowledge” is likely to be strongly influenced by the human sensory model bias. We can probably understand it. With effort.
But what if the machine generates knowledge by itself?
FOOTNOTES