I’ve long known of the artificial intelligence (AI) framework Hyperon, developed by Ben Goertzel & friends, and I’ve long had a vague idea of what it is about, but I never studied it carefully enough to grok it. I’m beginning to study it now, and this series of articles will document the learning journey of an absolute beginner.
ELI5 means "explain like I'm 5." OK, I’m pretty sure that most human 5 year olds won’t understand much. But I’ll do my best. And no, I don’t think you are a complete idiot. But when I want to learn something new I think of myself as an idiot, or better, one who needs the simplest, most intuitive, and most visualizable explanations and answers.
I’m a generalist by inclination, or in other words a jack of all trades and master of none, so at a certain point I’ll likely think that I grok the conceptual core of Hyperon well enough and stop diving in. But I’ll give references useful to readers who want to dive in deeper and become masters.
Why Hyperon?
I’ll begin with the question: why Hyperon? I mean, we have ChatGPT, Gemini, Grok, Claude... Why do we need yet another AI framework when we have many top-rated ones? The simplest answer begins with: Hyperon is different.
Large language models (LLMs) like ChatGPT, Gemini, Grok, Claude, etc., and the generative AI models that come bundled with them to generate images, videos, code and whatnot, are all based on the transformer neural networks introduced by Google researchers in 2017. Generative AI models based on transformers have stormed research and then society in the last few years.
Richard Feynman said: “When you get it right, it is obvious that it is right… because usually what happens is that more comes out than goes in.” In other words, great scientific works keep producing useful unexpected results.
This is certainly the case of today’s generative AI models. Why should a method designed to find the best next word in a text stream also be able to generate a video in response to a text prompt, and also be able to write code that works, and also be able to do all those other things that keep coming out? Useful and unexpected indeed.
The simplest way to grok how today’s generative AI models work is this: they analyze statistically a massive volume of training data (for example a large part of the enormous amount of data in the public internet) and translate data and statistics into a multidimensional (that is, seriously multidimensional, with a huge number of dimensions) mathematical space where close points correspond to things (words, text streams, pictures, video streams…) that are actually close to each other. Then it’s just a matter of generating a point close to a good result, and translating back.
OK, that was cryptic and needs decoding. A simple example is given by Simant Dube at the beginning of his highly recommended “An Intuitive Exploration of Artificial Intelligence.” The book was published in 2021, just shortly before the beginning of the explosion of generative AI, and so Dube talks more about pattern recognition than pattern generation, but the two are related and many examples carry over.
So, start with a large collection of square images with the same number of pixels. The dimension of the mathematical space is the number of pixels (one axis per pixel) for greyscale images, three times that for color images. The translation algorithm makes sure that close points correspond to similar images. The images of dogs (Dube talks of cats for historical reasons but I think the example works better with dogs) are points in a manifold (a multidimensional surface) embedded in the mathematical space. The images of dogs of a specific breed are in a region of the manifold. So if an image falls in that region then it is an image of a dog of that specific breed. And if you want to generate a new image of a dog of that breed, you can choose a point in that region and translate back.
OK, I’ve oversimplified and skipped complications, and I haven’t explained how the algorithm actually works, but I think this simple example gives the idea.
Today’s generative AI tools based on transformers combine many processes of this kind to generate really amazing results, which keep becoming more and more amazing. So, there’s the hope that this technology will scale to artificial general intelligence (AGI) and even artificial superintelligence (ASI). Perhaps adding more GPUs and then more efficient GPUs to AI data centers, and tweaking the algorithms to make them more and more efficient, will lead to AGI and then ASI.
Not so fast, says Ben. He is persuaded that today’s AI technology based on transformer neural networks is unable to scale up to AGI and ASI. He argues that a different framework is needed.
In “Raising AI,” De Kai notes that today’s top AI models “feel” their way through data and statistics, as opposed to “good old-fashioned AI” (GOFAI) methods that try to “think.” True, today’s generative AI models come with GOFAI-like thinking subsystems based on symbolic reasoning to complement the feeling nature of statistical pattern analysis. But the feeling model is the overall container, the central hub, or in other words, the boss that calls the thinking subsystems into action and coordinates their work.
An operating system for AI
Hyperon is meant to make extensive use of LLMs and generative AI methods that feel, but without putting them in charge. On the contrary, Hyperon is meant to be an operating system for cognition “where concepts, facts, goals, procedures, and even neural network components all exist as first-class citizens in a shared knowledge metagraph,” states the current version of the Hyperon White Paper.
“Engineering General Intelligence,” a 2014 set of two books written by Ben, Cassio Pennachin, and Nil Geisweiller, describes the cognitive architecture that Hyperon is based on. The book is dated: Hyperon is not even called that, and the book uses the term “generalized hypergraph” for what is now called a metagraph. However, the book provides long and clear first explanations of many Hyperon concepts and methods.
“The Consciousness Explosion” (2024), a monumental tour of AI (plus life, the universe, and everything) written by Ben with Gabriel Axel Montes, tells the story of Hyperon. The project continues previous AI projects started by Ben: Webmind, Novamente, and OpenCog. Hyperon, under development by Ben and colleagues at SingularityNET and the OpenCog Foundation, “is the newest version of the OpenCog system.”
We learn that, in Hyperon, “a self-modifying self-organizing neural-symbolic knowledge metagraph (the ‘Atomspace’) serves as the central hub of an intelligent agent, leveraging diverse plugins embodying additional cognitive capabilities (which may include LLMs and other deep neural networks).”
Metagraphs and the Atomspace
Nice! But what the hell is a metagraph? Forget “meta” for a moment and think of a graph. A graph is a set of nodes (points) and links (lines that connect two points). A social network is a graph of users and connections between “friends.” The possible states of a game of chess are nodes in a graph, the possible moves are links between two nodes, and an entire game is a path in the graph ending in a checkmate position. If you think of it, you’ll see that you can represent many complex information structures with graphs.
Hypergraphs further generalize graphs by allowing links between more than two nodes. These generalized links are often called edges in graph theory, but I’ll usually call them links to maintain compatibility with Hyperon documents. You can try and visualize an edge between three nodes as a tetrahedron whose vertices are the three nodes in the plane of all nodes, and another vertex outside the plane. This visualization method works well for simple hypergraphs.
In a social network graph, there are groups of users all connected to all the others. A user can be in more than one of these groups (family, small company…). Using edges, for example an edge for the family and another edge for the company, provides more clarity and organization. This shows that hypergraphs are a very efficient way to represent information. According to Stephen Wolfram, fundamental physics could and should be built on hypergraphs.
Hyperon introduces metagraphs as further generalizations of hypergraphs. First, links can span not only nodes but also other links. Since the distinction between nodes and links is blurred, the term “atom” is used for both. Atoms can have attached information such as weights that quantify their importance at a given time, and can be zoomed in and treated as subgraphs. Hyperon is built on a metagraph called “the Atomspace.”
According to Ben & friends, this framework can be shown to be computationally universal in the Turing sense of being able to compute all that can be computed, if given sufficient resources. The Atomspace can “represent all known programming and logic systems and computational models and physical theories in an elegant, concise, and manipulable way” and “elegantly model its own structures and dynamics.” In other words, the Atomspace can represent and operate upon the cognitive symbols used by the human mind (“feelings” being handled by motivational subsystems for things like goals and emotions, such as OpenPsi and more recently MetaMo). Therefore, the Atomspace is a good foundation for AI systems evolvable to AGI and then ASI.
I wish to thank Ben Goertzel for correcting some misunderstandings in an early draft. Any misunderstanding that remains is, of course, my bad.
Coming soon…
In the next article of this series I’ll begin exploring Hyperon’s developmental and operational environment.