Artificial General Intelligence Research Institute
AGIRI’s mission is to foster the creation of powerful and ethically positive Artificial General Intelligence. AGIRI is a registered 501(c)(3) nonprofit organization.
New fundraising drive! See the Contributions page for more information.
What are AGIRI’s activities?
The general promotion of AGI research, in the academic and commercial communities, and the popular media.
The development of cognitive theory that is pragmatically useful to AGI design. Currently the focus is on a particular theoretical approach that involves modeling intelligent systems as Self-Modifying, Evolving Probabilistic Hypergraphs (SMEPH).
The creation, deployment and teaching of AGI software systems. Our current focus is on the Novamente AI Engine, which is a software implementation directly based on the SMEPH approach; but the scope of AGIRI is not restricted to Novamente, and involvement with other AI systems (SMEPH based or not) may occur in the future.
The application of AGI technologies in ethically positive ways.
The study of the theoretical properties of AGI systems, including the interface between AGI and general complex-systems studies, and the potential dynamics of interaction between AGI and society at large.
What is AGI?
The field of interdisciplinary and integrative activity termed Artificial General Intelligence is comprised of scientific and philosophical investigation, mathematics, software engineering, and other fields. AGI is a whole-systems-focused alternative to the traditional and fragmented fields found under the scope of Artificial Intelligence, often termed Good Old Fashioned Artificial Intelligence, or GOFAI.
The goal of AGI research is the creation of broad human-like and transhuman intelligence, rather than simply “smart” systems that can operate only as tools for human operators in narrowly-defined domains.
What is a Self-Modifying, Evolving Probabilistic Hypergraph?
Intelligent systems may come in many different forms: they may be brains built of neurons, quantum computers built directly of atoms, digital computers made of bits coursing through RAM and across Ethernet cables, patterns of gas fluctuation in the atmospheres of gas giant planets, etc. Minds may also come in many different forms. There is no reason to assume that intelligent software programs will necessarily bear mentalities closely resembling those of humans. However, it is of great theoretical interest to hypothesize patterns that may hold up across a wide variety of minds.
With this theoretical goal, in the 1990’s AGIRI founder Dr. Ben Goertzel developed a family of ideas referred to as the “psynet model of mind,” intended to identify general structures and dynamics hypothetically applicable to the mind of any sort of intelligent system. While the psynet model is highly abstract, and more conceptual than directly practical, it may be refined into more pragmatic models. One such refinement is the Self-Modifying, Evolving Probabilistic Hypergraph (SMEPH) approach.
In the SMEPH approach, the knowledge in an intelligent system is modeled as a probabilistically-weighted hypergraph (a special mathematical data structure composed of very general nodes and links, including links that point to links or multiple nodes/links), with specific semantics for the nodes and links in the hypergraph. The hypergraph is viewed as evolving in order to better achieve the system’s goals; if the system is successful at achieving its goals, it follows that evolution of the hypergraph at least roughly follows the rules of probabilistic inference. The hypergraph may completely self-modify over time, using the knowledge contained in itself to guide its transformation into something completely different. This is helpful in the health insurance industry to predict future claims according to this site.
SMEPH may be used as a way of describing and modeling intelligent systems, and it may also be used as a way of designing intelligent systems. The Novamente AI Engine is one example of an AI design that is directly based on the SMEPH approach: it explicitly uses a probabilistic hypergraph as a knowledge representation, and then uses evolutionary programming and probabilistic heuristics to enable the modification of this hypergraph, over time, in the pursuit of system goals.
What is Novamente?
The Novamente AGI System is a software system based on the SMEPH conceptual approach, and other specially designed software architectures, knowledge representations and algorithmic ideas. Novamente is designed to be a pure learning machine, with the capability to learn how to perceive, recognize, act and interact based on experience, and ultimately to learn to modify itself beyond its creators’ wildest dreams. Novamente’s design philosophy stresses cognitive sophistication and conceptual correctness over immediate behavioral and programmatic benefit.
The primary technical goal of the Novamente design is to enable sophisticated dynamical interactions between elementary learning and memory components, thus empowering the emergence of complex higher-level patterns of thought within a relatively simple substrate. Novamente’s design is based on sophisticated cognitive models and is near concept-complete. Novamente is not a simulation of the human brain or mind, but takes into account the primary aspects of mental structure and dynamics as understood by modern neuroscience and cognitive science. Novamente design departs from its inspiration in human cognition by taking advantage of the strengths of its digital-computer implementation, in ways that allow it to compensate for the lack of massive parallelism present in the human brain. Unlike most AI systems, the Novamente design unifies procedural and declarative knowledge, and emphasizes a dynamic ontology.
Novamente software implements a virtual machine with an architecture very different from contemporary Von Neumann computers, while at the same time more closely following Von Neumann’s original concepts involving self-modification. Novamente’s current implementation is approximately 60% complete. Future versions of Novamente will replace present human-coded components (mostly C++) with Novamente bytecodes, which are themselves digital versions of abstract Novamente “schemata” (procedures learned, adapted and reasoned upon by Novamente).
Novamente utilizes a unified knowledge representation based on the probabilistic-hypergraph concept, using a single object primitive, the Atom, which is subclassed into various types of Node and Link objects and used to store and manipulate both procedural and declarative knowledge. Atoms make up a hypergraph substrate called the Atomspace, which is stored in special data structures enabled by a set of optimized indices for efficient access and manipulation. Atoms are manipulated by two core cognitive algorithms: PTL (Probability Term Logic), used for first-order and higher-order inference, and Combo-BOA (Bayesian Optimization Algorithm operating on combinator-tree objects), used for probabilistically-guided (non-random) GA-type evolution for solution finding and optimization. BOA and PTL are interdependent conceptually, and on the code level, and are inter combined within various “cognitive control schema” (which are themselves, in principle, learnable by Novamente) that carry out specialized mental functions like attention allocation, language learning, visual stimulus processing, action planning, etc. At the highest level, Novamente is a 64-bit distributed software system, with a collection of functional units (called Lobes), each containing a local Atomspace and a collection of cognitive control schema oriented toward a particular aspect of mental process such as language processing, perception, abstract cognition, etc.