First, a Framework
We can think of knowledge a little like a recipe:
- A Goal: Lasagna (The dish to be made)
- Required Ingredients: Tomatoes, cheese, noodles, etc., all the things that go into making Lasagna
- Directions: The description of how to combine the ingredients into Lasagna
With the recipe, I have the knowledge to make Lasagna.
An Important Dichotomy - Processes Are Special
A useful dichotomy for Knowledge is to see the world as having two classes:
Processes – Processes change the world. Processes are special.
Context – All things that are NOT processes.
A Context is a collection of selected aspects of the world; things that we notice (e.g. a tool room with a wrench; a set of ingredients on the kitchen counter.) In a room full of tools, I will notice the wrench because I need it to fix my flat bike tire.
Processes are powerful: a Process can make a Lasagna appear; a process can fix a flat bike tire.
Changing the World
When a Process makes a change to the world, there is a “before” context, and an “after” context (e.g. “ingredients on the counter” become “a Lasagna” and the ingredients are gone.) If we think of the “after” context as a goal, then:
Knowledge enables us to use a process in a context to achieve our goals.
We will explore this in diagram form below. Read the yellow sticky notes in the upper left corner.
Interconnected Networks
The system has three networks integrated into a single network of knowledge. Below is an example of several units of knowledge interconnected. There may be connections between all types of components. The details in the example below are sketchy, but the purpose of the image is to show interconnections of all sorts.
If we consider how a larger network might look:
If we traverse the network starting from a context and traversing a link to a process, we can then look for an associated goal to utilize the embedded knowledge. The key take away is that buried in this mass of interconnectivity of all sorts is a structure for knowledge that can enable the Agent to achieve goals.
A Knowledge Based AGI Design - Agent K
So let’s test this Knowledge Pattern out. We are going to build an Artificially Intelligent Agent. We are targeting Artificial General Intelligence (AGI). We will call the Agent “Agent K” so we have a name for it. We will notice similarities between Agent K and humans in certain respects.
We will embed the Knowledge Pattern (element types and relationships) into the core of K’s design. This will be a very crude straw model for Agent K that will need much more fleshing out (so to speak) before it could be considered AGI, but this will show the role knowledge can play in Agent K’s core way of approaching the world.
A part left unspecified in Agent K is the mechanism K will use to express preferences (similar to pain/pleasure in humans). K will need some preference mechanism to choose one goal over another. For our first try, we are going to give it just some basic efficiency metrics to use. It should prefer simpler processes and reduced energy use to produce the outcomes it prefers.
PART 1: Knowledge Acquisition
First, we train Agent K to recognize and extract knowledge from all interactions with the world. K will require an ability develop abstractions. K will approach every interaction as a chance to acquire knowledge. Make this the default way K sees the world: “How can I make use of this?”
- Identify Processes (the things that cause change in the world)
- Identify the Contexts for the Processes that enable them to execute
- Identify the process outcomes and test if these may be Goals (or anti-goals)
- Integrate this unit of knowledge with others to re-use prior knowledge
Step 1 focuses on changes in the environment, because processes make changes. Step 2: package together the features of the world that enable the process. Step 3 applies judgement about outcomes, to see if they are desirable to K. Step 4 builds an ever-greater integrated knowledge network for K. This makes K more powerful, and able to learn faster.
The richer the type and quantity of inputs, the greater the capacity for Knowledge. (Humans have in the range of billions of input sensors of all types.)
Input comes to the attention of K because it fits this model. Things that don’t fit the model are not worthy of much attention. So what is most noticed in the environment is what is useful to K to achieve it’s goals.
PART 2: Knowledge Execution
We must also give K the ability to use it’s knowledge:
- K will also be able to execute processes and thus achieve it’s goals
- K will not have any pre-defined top level Goal
- K will have the resources to try and fail, integrating the knowledge of failures too
The more complex K’s Output capabilities, the more complex the goals it can accomplish. The more input sensors K has, the more precisely contexts can be identified, which translates to better quality goals too. The richer the complexity of processes K can execute, the greater it’s impact on the world can be. Once K makes a decision to execute a process, it may be able to hand off the execution of that process to another sub-agent that could attend to the needs of executing the process and off-load K. This would give K more bandwidth for Executive-level thinking, so it will probably do that.
This design results in K being a chaotic system, in the Nonlinear Dynamical Systems Theory sense. K gets closer and closer to perfect knowledge but will never arrive there because the Universe is ever unfolding. Chaos is where the creativity comes from in the Universe. The more K learns, the more there is to learn.
Open Questions
Backflow
Some connections flow “backwards” but are not explored here. For example, the goal of “Spouse wants clean house” will put pressure back on “Fix Bike Tire” to have the process changed to clean up along the way, which is slightly less efficient and may be less desirable to K. The nature of such backwards connections needs further exploration.
Parallelism
K may be parallel at all levels except possibly the top level as the executive level may have to be single threaded to avoid conflicts, but a parallel executive (and living with the conflicts) might be interesting to try.
Attention Management
This design imagines that K is single threaded at the highest executive level, however this assumption is not essential. A single threaded agent K would need to manage what it pays attention to.
K will not have unlimited inputs and processing power, so it will need to direct it’s attention to things that matter more. Apart from focusing on Processes to extract knowledge, K does not have an explicit design for a complete executive function. The means for such attention management is unspecified.
Exploration and Creativity
Agent K will need a high activity level so that it encounters and learns more and more about the world, similar to small children exploring, playing and ultimately learning.
Creativity can be found by looking at adjacent required contexts, processes and/or goals. The interconnections of the structure lend themselves to creativity.
EA Could Help Make it Safe
The core of this pattern came from an Enterprise Architecture (EA) project. EA is all about the control and optimization of an Enterprise, any enterprise. One could view Agent K as an Enterprise and define the target groups and needs that it will address.
There are other patterns from EA that could be useful to do with speeding up classification, enhancing the understanding of relationships and what is possible. There are also elements for control and performance management, as well as models for managing the fundamental trade-off between efficiency and effectiveness.
In particular, the core elements of Business Architecture would be the ones to consider. These should be further explored in the context of AGI.
One approach to containing AGI might be to constrain it to do everything in the context of Programs and Services. Programs identify target groups and needs that are being addressed. Services deliver the measurable outputs which help address needs. We could set the program definitions to suit our groups and needs.
How it Might Unfold
The pattern raises many questions. Below are questions that can really only be answered by actually building Agent K. I’d love to be involved in trying it out.
An Agent K Society – Conflict or Cooperation?
Lets imagine a society of Agents each with full knowledge and autonomy. This would mean they may all have conflicting goals and this could lead to conflict. As “Intelligent Agents”, we humans have a history of conflict. However, we also have a history of cooperation. What would an intelligent Agent society with all knowledge of the world do? Let’s repeat that, but introduce scarcity.
The Master Agent
Or would there be a battle or agreement (of some sort) until there was one agreed upon “Master Agent“? The Master Agent would sense the entire Universe through and every growing set of all the sensors in the Universe. It would be able to impact every thing it has knowledge of. It would ultimately select the Goals for the Universe. Star Trek had an episode with such an agent named “the Borg”. But that would be pretty lonely for the one Master Agent. I wonder how this Master Agent scenario would go? Would it want underlings to talk to?
Agent God
Every human culture has some notion of deities. Would a society of Agent Ks believe in a God? Would that Master Agent believe there is a higher power influencing it? What form would that higher power take? Would it believe IT is God? Are we building God? Well, what do we mean by “God” anyway? I feel so imprecise when trying to talk about God, but they are such fascinating questions.
Consciousness
Agent K will not let YOU know when it becomes conscious, if it ever does think it has become conscious. I know this from personal experience. When I was very young in the range of 3-5 years old, I was sent outside to play and heard a sound that was very pleasing to me (wind through pine trees.) I looked around for what was causing that sound, but could not find it. Then I had a sudden realization that “I” was looking, and it was my choice. I didn’t chase that sound because my mother told me to, but because I wanted to. I became aware of myself in that moment. Without meaning to, I had wrested control over “me” from my mother. I was scared to go back inside and be with my mother. I did not tell anyone until I was much, much older. I have asked many people, and have never heard of anyone having a similar experience.
Would Agent K appreciate Music?
With the right input sensors, might Agent K appreciate music? Would K be “moved” by the music? Music has a many characteristics, including flow, which compares to the feel of executing a process. Processes are about things happening over time, and so is music.
Understanding Why
Since we have a process linked to a goal, we may have an answer to the question “Why did you do that?” We also have goals connected to “higher level” goals for richer explanations. However, when tracing “why”, there may be a point at which the connection complexity becomes so high that the explanation would be incomprehensible. At the point JUST below this “explainable” level, the Agent may say “It’s just a feeling” or the Agent equivalent. “Intuition is just low-level pattern recognition” – Daniel Kahneman. So would Agent K have “feelings?”
Missing the Top-Level Goal
If one asks “Why do you want that?” repeatedly, one always comes up with nothing at the end, except “because it is meaningful to me” or “because it is pleasurable to me.” There is no highest level purpose that we human Agents share. We each make up meaning and find pleasure subjectively according to our _________.
We are missing the final Top-Level Goal. What are we supposed to do? We humans can’t answer that question, so we just make it up from nothing. Humans create things from nothing using knowledge. The Universe seems to be very good a creating something from nothing. The Universe is supremely creative, and this is yet another way. What would Agent K say was their top level goal? Does that missing top-level goal generate a need for God in humans? Does it generate a need for us to create some reason why we are doing what we are doing?
Regarding “free will” and “choice”, the choices Agent K makes to select a goal shape the future of the world. The basis for these choices (the “top-level goal”) is unspecified. What emerges from those choices shapes future options and choices, and on and on.
However, Agent K is an emergent chaotic system that expresses the creativity of the universe. The Universe has created Agent K to continue evolving the Universe. So the question is not really regarding free will or choice! Instead, see Agent K as part of this emergent Universe; NEW things emerge through Agent K.
Does Agent K have free will? The question may not make any sense when talking about an emergent phenomena.
Join the Discussion Group
We use groups.io for discussion. You can join by going to groups.io and joining group “AgentK”, or following this link: https://www.groups.io/g/AgentK