Warning: I am trying to build this model by writing it and sharing it. Please feel free to contribute any thoughts you may have. This post will probably change over time as I further refine the ideas. It does not currently cite sources.
Some researchers seeking a pathway to AGI have noted the importance of large-scale mental models (e.g., Numenta, Vicarious). Imagine thousands of simultaneously-activated mental models related to what you are currently perceiving. Some speculation is added below based on this writer’s knowledge of empirical literature, experience as a psychologist, and personal theories based, in part, on introspection and observation.
The notion of simultaneous activation of large-scale mental models is essentially correct in my view, but is only part of the puzzle. The large-scale models are not intelligence, but an important aspect of intelligence. In my view, intelligence is the outcome of a reciprocal relationship between large-scale mental models and many other important functions that interact with, modify, create, retrieve, construct, and transmit mental models.
Defining Human-Level Intelligence (HLI)
Human-Level Intelligence: The efficient retrieval, generation, alteration, partitioning, search, description, and navigation of large-scale, hierarchical, componentized mental models through the use of selective attention to facilitate the achievement of a goal through one of the two major modes of operation (task-positive mode). The second operation mode allows for an implicit goal of promoting connection, thriving, meaning-making, reproduction, and understanding of the self and others (task-negative mode). You can think of the operations of intelligence becoming focused on self, other, future, and past autobiographical models in this mode.
Procedural Behavior Models (PBM)
Procedural behavior models are executed in a time-mediated conceptually sequential, but low-level parallel manner. They are hierarchical and conceptually modular and recursive. Any PBM may be embedded in any other PBM or self-embedded in the recursive case. Behavioral plans can be built through observation, retrieved procedural memory, internal sensory models, language, and random firing of neurons. They can also be baked in through fixed-action patterns (e.g., walking in a baby deer, swimming in an infant human). They can run in the background or be executed with conscious awareness. Some are largely inaccessible to consciousness (e.g., saccades of the eyes, digestive processes mediated through the nervous system, heart beating).
You can think of behavioral plans as the body’s “robot control” functions. To demonstrate that this must function in parallel, imagine all the simultaneous movements that happen during any of our movements or behaviors. Yet, we can describe, in serial fashion, how to perform an action (e.g., open the cabinet and retrieve a cup suitable for coffee). Hence, you have the conceptually serial element and parallel execution on a neural level. It is possible to imagine a parallel element described conceptually as well (standing on one foot while reaching up with one hand).
Thinking seems like it could be an internal version of this. It is a process happening in parallel, but described in a conceptually sequential manner. Thinking can be viewed as the conceptually sequential description of the the thought process happening in parallel. It may be the description of moving through the mental space (sequential conceptual description of that movement).
In contrast to Noam Chomsky, I believe that language is a intellectual tool rather than being at the base level of intelligence. It is the most powerful tool of intelligence yet invented. In my view, language is a way to encode, retrieve, transmit, organize, build, and refine mental models. These functions operate through a self/other communication process. There is a rudimentary type of intelligence that can be possessed without language, but it lacks those robust functions conferred by language. Language appears to be a PBM and is one of a class of functions rather than being qualitatively different from other mental functions in the PBM class. Language may be the most impressive PBM yet, with the possible exception of consciousness.
Words may be thought of as the very tip of a conceptual iceberg. Beneath the surface is the underlying model. Some of the models that are beneath the surface can be accessed if one lingers at the space between words or simply thinks a word and lingers on it. There is a deeper layer that is inaccessible that may be thought of as analogous to the unconscious mind. This is a much more specific model than envisioned by psychoanalysts as it relates to necessary functioning of the mind in the moment. In order to think in words, we must be unaware of most of the underlying models. It is necessary to conserve energy and to enable efficient transfer of models to others through language.
Just as creatures in the evolutionary past moved through their mental representations by acting in the world, our higher-order faculties add a meta-level to that process. It could be thought of as a dissociation, or partitioning, capacity, which is an important element for being able to think. You could think of this as the window of consciousness and/or attention that can be focused on particular things to the exclusion of other things. This may be, at least in part, completely physiological in that when attention is focused, energy is directed towards the processes being focused on, and away from things not focused on. It may also be that active dissociation can be consciously chosen and habitually enforced through the same control mechanisms that are used for learning procedural tasks. These practiced dissociations can become completely automatic and unconscious, just as all procedural memory does with repeated practice.
Emotions are general motivational and informational states, which are the outcome of complex unconscious analysis and memory. They may add weight to mental models in terms of importance and desirability. They may also add weight to goals that are selected. Humans develop mental models for their emotions that involve language to perform all the roles that language performs noted above. Emotions dissipate naturally over time. Human emotions are at least partly subordinate to linguistically activated models. They can be triggered through language or subside more quickly depending on the models employed. In other words, the mental models activated by conscious thought may activate or dissipate emotional states just as certain sensory experiences may automatically activate emotions. Emotion serves an important function in intelligence as it helps us to decide which goals to pursue, or when to disengage, whether mental or in the world. Emotions may also serve as a symbolic reinforcement, which is something that has not yet appeared in machine learning to this writer’s knowledge. Emotions could also be used to help improve large language models with the use of language embeddings.