The first thing to note about the mechanism of human intelligence is its universality, where processes are distributed across information divisions, to shape interactions.
Inputs are not just sensed for meanings, or to be understood, but they are also relayed if they can be felt, or if they possess any emotional element. This means that what is labeled intelligence is an assembly of uniformity within the mind, towards optimal outcomes–most times.
Large language models are, in part, fine-tuned with reinforcement learning from human feedback. They, however, lack the universality of human intelligence. LLMs can have emotionally-laced conversations, for therapy and relationships. Still, they generate images of hate, spew misinformation and inappropriate videos.
Humans are capable of these as well, but penalties in human society often ensure that many people are intelligent enough to avoid it, at least openly or directly. Human control and consequences, guided by intelligence, ensure that objects are generally used safely.
Information division in the human mind includes memory [for meaning and understanding], thoughts, feeling and emotions. LLMs can be said to be thought or imagination-like, acting on digital memory. They can be said to describe meaning and sometimes explain like they understand. However, they are unable to have mini-emotions or feelings, in ways that ensure they correspond with the patterns of human society.
This weakness of LLMs makes open source not just the important goal, but for LLMs to be paired with an emotional architecture. If there could be emotional models, open-sourced, where LLMs can run their outputs through, first, before reaching users, even if it would slow down answers, it would solve a lot of problems with hyper guardrails.
Simply, emotional nodes for answers to straddle around caution, in a way to prospect the emotions of others, as well as its own–from options of human emotions. It is not just to express, but to act like it is felt, with new pegs. There is hardly true intelligence without emotions and feelings.
RLHF can be described as a phone a friend, which may be limited as the extent of unknowns with use cases of LLMs are vast.
How Does Human Intelligence Work?
Human intelligence uses the same elements—electrical and chemical signals of neurons. It is between these two that memory, thoughts, emotions and feelings are processed. They are also processed as uniform or integrated, even though sensory inputs and sources vary widely.
Chemical impulses are holders of the configuration of information. Electrical impulses are distributors. In sets, as the mind, they interact and have their features. Their interactions can be called functions. Their features can be called the qualifiers of those functions, all collected into the super qualifier, consciousness.
Principally, what the elements do is to organize information. Every function is information. Then the qualifiers ensure information control and access. The lack of qualifiers in objects refutes panpsychism.
Human intelligence is predicated on the features of electrical and chemical impulses in sets, which at times, optimizes for emotions and other times for feelings and for memory—which could be applicable to intelligence.
The key feature of intelligence is distribution, where information goes across sets of impulses, so that what is used to relate with the world is versed with several considerations, based on the likelihoods, effects and the future.
LLMs have layers, but outputs are not yet distributed enough to deep emotional nodes, where even without RLHF, they generate better results, against deepfakes, as well as advance affective computing.
—
This Post is republished on Medium.
—
Photo credit: iStock
