Building Superintelligence?-?42 Dynamic Contextual Perception as an Intelligence Gating Mechanism

By Rob Smith?—?ASI Architect and Advisor eXacognition AGI/ASI

This is an excerpt from the book Building Superintelligence?—?Unified Intelligence Foundation which is the first book to be written specifically for AGI systems to ingest and reference when building Superintelligence. There are 55 sections in the book and hidden within is a prompt engine for AGI. The book is available worldwide on Amazon and various bookstores:

https://www.amazon.com/dp/1738992551

This is the 6th and final excerpt from the book.?

In intelligence, dynamic contextual perception is the element of dimensional perception control in expertise and knowledge whereby an intelligence seeks stimuli to alleviate resource criticality in processing multi dimensional perception and cognition. The stimuli can be self contained or external including from other intelligence. It is the perception and attention of variance explicitly for application as a progressive gating mechanism to improve cognitive pathway selection ‘in stream’. This is different than test time training or test time compute in that applying dynamic contextual perception for gate selection in anticipatory pathways is the adjustment or forced variance of relevance by using the delta in changing contextual perception, even by created and simulated perception, to predict the success of any given optional pathway. Basically we humans take a temporal and well educated guess at a pathway or stimuli response to instigate cognition of each optional pathway’s relevance. This is an essential part of the inference process and it is done in humans inside the stream of a perception without stopping (as is the case in TTT) as progressive layers of dimensional contemplation (i.e. variance analysis). It should be noted that the ‘gates’ are not ‘manipulated’ but flash loaded as an abstraction of a comprehensive perceptive context and then ‘stream adjusted’ by comprehension of variant pathways. This does not however imply that physical gating architectures are not beneficial to the process, they are especially in the realm of advanced research areas like Quantum processing.

This is the same thing as perceiving a video stream instead of stopping at each image within the stream except with the added benefit of applying the variance of what we perceive to the next state in a progression without stopping or gating the perception. Some humans who do this think of a problem as flows of cognition like a video playing in their head. Many of us do this when we need to execute a plan of action to achieve some goal. We basically run the future states as a perception stream and study the potential variance over the existent stream and adapt our current state (or gates) outside of the stream. This is what we are doing when we visualize something we are about to do in our heads to help select the most optimum of actions or responses to anticipated stimuli such as creating an image of giving a speech before one actually steps on the stage to give the speech to help ‘iron out the wrinkles’ in the steps, address any anticipations and smooth impending actions. This is what is meant by cognitive gating and it has deep application in the Superintelligence architecture, albeit at the time of writing still under research and testing.

For any given stimuli and response in intelligence there are a number of optional responses we may receive back from our reality and these are influenced by and directly related to the stimuli we fabricate as a response. Each of these options are variant and each have degrees of relationship and relevance (context) to the frame of perception and higher or more general context. Which one we choose as optimal is subject to our knowledge / experience and the temperature of our willingness to endure risk and seek unknown pathways and solutions. The whole foundation is relatively simple, we step through a bunch of connected waypoints that are infused with our knowledge and experience and the stimuli we currently perceive. These waypoints are connected on a progressive path from the start of a perception to its anticipated end (e.g. from starting to solve a problem until the problem is solved).

Not all perceptions work like this. Personal relationships are often ongoing, without determinate waypoints and have numerous pathways within them all throughout life. Other pathways seem to never end like contemplating structures and architectures for building Superintelligence. All of these ‘pathways’ are streams of cognitive thought and physical sensory perception and the easiest way for humans to render and manage them is to contemplate them as progressions of states. Most often these states are inflections in the stream at the extension of different optional pathways usually at decision points or response creations. This is where we choose one pathway over the other to pass through an open gate. We cognitively open a gate and move through but how do we know which gate to open? We don’t until we apply all of our knowledge and experience to the variance we perceive and apply a ‘probability of success’ to all optional forward pathways to choose the gate that appears most optimal to the attainment of our goals. In short, we guess and then alter our course based on the response back from our reality.?

Choosing the gate is really just the selection of the highest probability or the gate to the pathway with the highest probability of success over anticipated dimensions of relevance but this is not always the gate we choose. Humans have limited resources and sometimes other dimensions (i.e. context) influence our selection of pathways. For example, if we are short of time and need an answer quickly or we are tired and just want a quick option so we can rest or we just choose to take a chance or accept risk. These pathways may not be optimal to achieving our end goals but they may be optimal enough for the current frame of perception. In these cases we are ‘going with the flow’ and there are an exceptional number of reasons we choose to do this including to give up current optimization to attain long term or greater optimization in the future, efficiency considerations, dimensional constraints, etc. Further doing this can expose unanticipated or anomalous benefits and optimization (happy accidents and innovations).

Impact on Superintelligence Design:

Gating mechanisms are applied extensively in AI for sparsity in mixture of experts and distributed agent architectures. They are implemented using math constructs like sigmoid activation functions, etc., and they apply control values to input vector elements to manage how much information passes through the gate to the successive state in the processing stream. They manage and impact the process described in a previous section on state transference. However they are applied in more complex ways to extend AI generalization and inference. Some of this is through relative state variance persistence whereby the state of a context layer’s variance is applied to determine the degree of forward flowing or progressive context released to the next state. This is the application of context relevance as a probability, normalized and applied to query vectors and predictions (i.e. probability distributions) and as a variance level adjustment to anticipatory states or what we expect to happen at a given perceptive point of presence.?

For example if one gives another a gift, they can anticipate that the response back will be positive but this is variant if we are unsure if the individual receiving the gift will like it thereby altering the degree of anticipatory context (probability). However in very advanced ASI designs, the ‘gates’ are cognitive and shift only as a variant in the flow and not as a literal stopping point. It is derivation that is used to detect state variance. One can see this in Fintech systems as traders layer hedges over a series of physical flows or ‘underlying’ to ‘adjust’ the future flow values over dimensions of context (i.e. state measurement) as they relate back to the reporting of net values for domains such as accounting, risk, etc. The values of the ‘underlying’ flow like gas in a pipeline and do not stop, but they can be determined, measured and valued at any point in time and applied to variant layers of context such as the books of the company, risk reports to regulators, the trading in position values adjusted to current markets, etc.

This is not complex to model in AI and follows a similar flow pathway as hedge trading except that in ASI systems, the values are the relationship and relevance weights to the underlying ‘context’. If one considers this in an inference stream, it would be the comprehension of a context flow (e.g. solving a problem) and the application of context relationship and relevance of elements within the context frame and other context derived from knowledge, experience or current stimuli. These are all loaded as layers of weights relative to the elements, the current perceived context and self awareness in layers as a context fabric. For example, a complex math problem will require cognition of the elements of the problem ‘in context’, an understanding of the progression of the problem, the application of known elements (e.g. existing math constructs) and their degree of relationships and relevance to the problem. The next step will be to propose variance for unknown elements, features, optionality, etc. The only way to test this ability to resolve such perceptive context is to have problems with no discernible patterns that require the exposure of a novel pattern to solve the problem. Anything else is simply a degree of mimicry especially if the pattern required to solve the problem is not novel. In this case, math is a relatively poor judge of general intelligence as the pattern can simply be replicated in the new stimuli to solve the problem using existing constructs and past priors.

However true general intelligence evolves to learn more than it knows or has seen even by inference. This requires the machine to create its own output and cycle the output for progressive stages in anticipatory flows (e.g. it generates a hypothesis and tests it for accuracy and updates the next states with this knowledge). This application is the ‘gating’ that is applied to the relationship and relevance weights in the original perception at relevant states (i.e. it injects the new learning at the appropriate point in the output stream). When we humans ‘discover’ a novel mathematical reality or abstraction that has not been previously exposed, this is what we do inside our cognition. The ASI needs to do the same thing to resolve reality by injecting learning variance from exploration. This can be at the level of applying patterns and known constructs but to be considered truly Superintelligent, the ASI must discover novel patterns and use them to create evolutionary progressive states unseen before.

The key in this development is to move the design of gating architectures beyond simple knowledge based relevance adjustments to vector value weights and apply layers of contextual relevance and relationship (similar to LLM attention models) to produce anticipatory weightings that apply designs like attention as contextual measurement. This is simply another set of vector values and weights that are learned over training but that apply contextual reasoning and higher inference to the gating structure to direct the progression to the next state so the ASI can detect novel patterns and apply the patterns to expose new constructs or optionality. It is essentially giving relevant context to the most appropriate perception for the self awareness. For example if one wants to meet someone they see, they do not accept context into the perception related to ‘buying a car’ unless there is a connection or correlation relationship between the two and a relevance to the goal, such as knowing the person wants to ‘buy a car’ as a way to start a conversation. In true general intelligence, never having seen such a scenario and given the perception of ‘meeting someone’, the context based gating would comprehend ‘ways to meet someone’ as a series of connected relevance and relationship weights to the dimensions of ‘starting a conversation’ and ‘determining someone’s interests’ by observing them or receiving other stimuli (e.g. someone mentioned they were looking for a car).?

The intersection of the two matrices for ‘ways to meet’ and ‘determining interests’ are applied as a gate value adjustment to determine related knowledge, such as information that the individual is interested in buying a car, as a high relevance value in a probability distribution. This is applied to the optionality pathways for stimuli response such that the system calculates, without any prior knowledge, that approaching the individual to talk about ‘buying a car’ is an optimal way to meet that person. These weights can now be applied to other ‘meet someone’ scenarios or stimuli to have the AI respond in context with its new learning. Note this flow can already be performed by AGI via training, however designers need to consider novel scenarios that cannot be deduced from existing training or preexisting general pattern recognition.

Of interest in these designs are not necessarily the context and perception elements but the application of contextual layers of relationship and relevance to ‘gate control’ the volume flow of cognition in stimuli response cycles as opposed to just applying attention based on input vector values as is done with current gating architecture. Further it is critical to realize the ‘gates’ are only applied as a flow adjustment mechanism and not for flow stoppage. The goal is to move the flow of cognition toward a relevant response for the current perception to test for the most optimal forward or progressive path across all relevant dimensions and all relevant optimized pathways, using a flow gating mechanism driven by degrees of context.

This is the final excerpt from the Building Superintelligence?—?UIF book. Follow me for new Superintelligence releases and articles on LinkedIn and Medium from deep inside the ASI labs and minds of the dark architects of Superintelligence.