Abstract (first draft) of the major details and components of a General-Purpose Artificial Intelligence system.
FPGA/ GA optimized hardware handles >95% of “judgment calls”, and >99% of algorithmic elements of sensory preprocessing. End result of sensory preprocessing is the translation of all sensory input to the AI’s Internal Symbolic Language (ISL).
ISL is “Godel Optimal”; all ISL components and constructions are derived from combinations of and operations on as few external references as possible. (ISL is a logically self-consistent mathematical system). All Instructions and Data are ISL constructions.
There exists >= 1 primary ‘thought processes’ (multiple primaries whose aggregate output is the input to a “super-primary” process is one promising configuration)
Primary processing (PP) is modular in nature. It includes a large ISL construct store, whose contents are ranked and linked by multiple criteria (position in linear queue, attention priority, etc.) and the initial logical/heuristic/inferential/et. al processing of said ISL store.
PP has at its disposal all known AI logic routines. These cooperate with the AI’s memory store and motivational processes in processing Primary input. They derive from the input ISL constructs representing important ideas, facts, etc. and a tentative set of expectations of future events and possible courses of action, both mental and somatic. Tentative expectations and plans are heavily influenced by current beliefs, values, and emotional states. The aggregate of all derived ISL constructs is placed into the primary output buffer.
Multiple secondary processing units (50+?) exist, organized into a hierarchical, nested structure. Each secondary unit is similar in composition and function to the PP unit. The 1st Secondary Unit (SU) takes as input the contents of PP’s output buffer, as well as details of a random sample of PP’s decision-making routines. PP’s output buffer is processed in a fashion similar to that of PP. The end result of SU processing is similar in format and content to the PP output buffer. In addition, the SU analyzes the sampled decision making methods of PP. These initial analyses can trigger very detailed examination of a given process; if result of examination is deemed to have a high attention priority, the results are placed into 1st SU’s output buffer.
1st SU output feeds into 2nd SU input, 2nd out into 3rd in, and so on and so forth. In addition, samples of each SU’s decision-making processes are fed into the next SU’s input buffer. This nested self-examination and modeling is at the heart of each SU level’s functioning. At all steps there exist controls that keep the nested processes from entering infinite recursion. The output of the last SU level is entered into a dedicated ISL store, where the contents of final SU output are examined in detail by multiple tertiary evaluation units. These tertiary units are more specialized than Primary or Secondary Units. Also, their judgments and evaluation processes are highly subjective in nature, informed by the AI’s current value system, emotions, and goals. There are tertiary units specializing in fact and consistency checking, attention interrupt prioritization, memory store input buffer insertion, etc. Tertiary decisions result in the modification of the SU final level’s output. This ISL store’s contents are then fed into PP input buffer as “train of thought” data.
The contents of each level’s output buffers are fed into routines that compare and integrate the contents w/ the AI’s current goals and emotions. Also, each level’s output buffer’s contents are entered into a memory insertion/retrieval candidate store. The contents of said stores are evaluated, each item ranked among multiple, mutable criteria.
High ranked insertion candidates are entered into the memory store’s input buffer, high priority retrieval candidates cause a command to be entered into memory store’s retrieval queue. These commands trigger the retrieval of data related to the retrieval candidate; said data is placed in next level’s input buffer.
Every Primary and Secondary unit, as well as many minor systems, contain in their input buffer output from the AI’s self modeling system. This is mainly “thoughts about thoughts”, a crucial component of the AI’s self-awareness and consciousness.
The AI’s self-modeling system (SMS) is similar in structure to the ‘main’ structure (i.e. primary, secondary, and tertiary units and their inputs and outputs). SMS evaluates all PU and SU output, as well as a sample of tertiary outputs. Each level’s SMS input is evaluated, a sample of the results of said evaluation are entered into said level’s output buffer. SMS outputs are entered into the goals, beliefs, and values (GVB) system.
The GVB system is also structured similarly to the ‘main’ structure. GVB takes input from all major and minor AI systems, and exhaustively analyzes said input. Results of analyses are used in the modification of GVB’s dedicated ISL store. Said store represents the AI’s major values, opinions, goals, etc. ISL store is supported by a vast array of minor systems which compare and evaluate the ISL store in relation to the AI’s current thoughts, emotions, somatic states, sensory input, etc. (and vice versa) and make appropriate modifications of ISL and current thoughts and emotions. Emotions serve as vectors influencing motivations, goals, and planned actions and thoughts.
There are a number of independent systems that run concurrently to all major and minor systems; said independent systems perform functions similar to human mental features like boredom (to discourage excessive attention to one activity) and subconscious processing (intuition, ‘instinct’, etc. These processes will be explained in great detail in subsequent design docs, as they are crucial to the functioning of my design.) These independent systems take input from all other AI components.
The primary processing unit contains a highly complex decision making unit. This unit takes the tentative decisions (output of last ‘main’ structure cycle) included in the Primary input buffer as input. In addition it receives input from all other AI systems. This additional data is used in a rigorous evaluation and modification of tentative decisions. Finalized decisions are entered into the AI’s decision stores. These stores input to the AI’s somatic effector’s action queues as well as reporting finalized decisions to the Primary input buffer (and all other major AI systems)
c) 2002 Jonathan B. Standley