Open source general ai project. Let's automate everything.
There will be many many tasks needed for a general AI system but one of the key fundamental requirements will be how to be logical. Even logic can be thought of as a pipeline of organized rules to follow. We will likely need a specialized GPT for logical reasoning and a general agent keeping track of varying priorities. Logic is really a method of organizing a large set of preferences like (don’t die and be happy, and while learning and advancing a goal. In game theory it equates to taking small steps that lead to a better outcome while not losing. In a computer architecture we can just spin up a new agent and call the agent again and again. This is sort of like starting a new life over and over again. As long as the agent shares some advancement to a collective database then all future agents prosper. Success is not certain but reaching an absolute or local minimum (loss) is certain over a given time.
We will build a KAN network of features stored in a otological database.
The inputs will be first and second order features of a given state. I will give an example. To go from one state to the other state requires a very precise definition of this space. The definition is a set of features that can be calculated. These features are a set of other slightly more fundamental features. Therefore we will also need a GPT that can divide a given feature into another set of features. This can be thought of in computer terms as a recursive function to be called again and so repetitively dividing the Feature space from a condensed set of features for a given State space. How do we identify the new set of features from one high level feature? We can find features across multiple models that are fixed and fundamental by training from multiple sources and landing on the same features that appear to be fixed over the study time period. Let’s use gravity on Earth as an example of a fundamental feature. Can we use a different sensor and still get the same fundamental features like speed of gravity from video of an object falling or a baseball being thrown and use the parabola to calculate gravity then repeat this via a time of flight sensor or a phased array of sound detectors able to track an object and also calculate the rate of acceleration of gravity? Gravity would be the fundamental feature shared by objects on the same planetary body and also can be substituted with a different fundamental feature.
Any real world measurements like location, time, coefficient of friction, weight, hardness, polarity of a molecule, charge of an atom, temperature are real world targets or fundamental features. Once we get to a value that is no longer changing rapidly or perceptively then for the time being the value will be fixed up until it does not match expected values, for example, the observable rate of gravity on Earth appears fixed until we are are moving relative to the Earth like in space or an airplane descending. We can then learn that gravity is relative to a given location and direction of motion. This will imply that a feature, gravity on earth, is no longer the most fundamental and we can therefore learn new features by dividing into a multidimensional space to examine all possibilities for the new changes. Then through empirical data with examples and observations we can then divide and add to our understanding of ‘observed gravity’ (on Earth) as a function of location, mass, relative (vector) movement.
The key to the task planner is actually the Root Cause Analysis bot. This bot's goal will be to identify what are the effectors of a given action. When a state changes this bot will evalute all known effectors with a goal of determining what are the primary effectors and to what degree. An object moved via an external force. What was the force? Are there more than 1 cause of the action? Location proximity and time series will be key tools of this bot. Finding a root cause is sort of like acting as a detective. All details matter and no assumptions to be made until the data is collected. Every action or cause will have some clues. These need to be observed directly when possible or indirectly if limited. Like a good scientist a theory will be formed and tested. Off line testing for most features and then real world testing when a novel feature is expected. The difficulty will be to focus on individual features and thereby limit the dimensionality.