We are seeking to build a thought machine via language models. LMs are language experts in core, so its outputs are interpretable from the get go and understand the content immediately. It also have great quantities of data to train from. However, out of all these ample train data and great framing advantages, is this the right way to implement a thought process?
Here’s a question I want to raise: are we solely thinking in a way linguistic pattern directs us? No, there are more other drives, for instance our personal dialogue preferences or background expertises. Some call it a personality trait. Some call it inductive bias. This also implies that there are a numerous coherent candidate of the next trailing thoughts and we ‘choose’ out of them, each of them linguistically correct.
Being linguistically correct is a minimal requirement, as it guarantees that the thought is comprehensible by all the language speakers, which in turn means a thought is verified to be reachable by multiple persons. Linguistic expressibility is sometimes a big hurdle for a novel thought to pass, because the thought underneath might be never expressed in language before. Some ideas are hardly expressible, either it has not enough words to express and restore (like subjective taste experiences or newly made theories), or it is practically too delicate to elaborate in words (like expressing dance moves in words). Thus in some ways, language format itself is a bottleneck filter of thoughts when it comes to thought machines.
We need to examine more on what makes two ideas coherent, and how do we induce the next with the given. We need to examine if a thought is discrete, comparable, sortable, decomposable, partial-modifiable, reusable. After that, we need to find what pathways makes the next coherent statements.