SingularityNET’s 2022 Progress Towards AGI
OpenCog Hyperon advances in MeTTa language, Distributed Atomspace, and DSL’s
Also available in Spanish, Portuguese, Russian, French, Chinese
Greetings Singularitarians,
Today, we’d like to share a special update, a deep dive into the AGI progress made by SingularityNET in 2022, and an overview of what makes OpenCog Hyperon — SingularityNET’s approach to an AGI framework — different from other AI systems.
As AI systems demonstrate greater practical functionality each year, it becomes increasingly apparent that the breakthrough from narrow AI to Artificial General Intelligence is near. However, there is still no agreement among researchers about how the breakthrough will be made. While deep neural networks have demonstrated impressive capabilities for impersonating intelligence and producing intelligent-looking artifacts, their complete lack of commonsense understanding and real-world symbol grounding makes it appear unlikely that they can serve as the core component of a true AGI system.
It’s possible that computational neuroscience simulations will make huge strides, or that AGI will spontaneously emerge from self-organizing networks like the SingularityNET Platform without coordinated planning — but it seems more likely that some new innovation in cognitive architecture and/or learning and reasoning algorithms will be needed alongside these. OpenCog Hyperon, the newest iteration of the OpenCog open-source AGI platform, is a strong candidate for this breakthrough innovation.
OpenCog Hyperon’s progress in 2022
The SingularityNET Foundation, together with its incubated spinoff TrueAGI and an open-source volunteer community, has — throughout 2022 — brought OpenCog Hyperon from the stage of mathematical and conceptual theory to the stage of working software! The alpha release of two key components is scheduled for early 2023: the MeTTa (Meta Type Talk) AGI language interpreter and the Distributed Atomspace (DAS) knowledge store.
OpenCog Hyperon is a visionary project to build a complete, scalable, and open-source general artificial intelligence system based on the principles of cognitive synergy. It is an open-sourced platform where different AI modules — such as neural networks (NNs/DNNs), generative AIs, probabilistic AIs, program learning AIs, and others — can collaborate based on a shared knowledge metagraph. This architecture provides a highly scalable learning space and knowledge store to the various types of AI systems, and a set of tools (such as pattern mining and attention allocation), allowing them to coordinate to learn and solve problems together — much the way humans can coordinate intuition, experience, insight, and logic to solve a single problem.
MeTTa (Meta Type Talk), OpenCog’s Programming Language
Hyperon’s AI components work together, achieving results that none of them could accomplish individually. This collaboration is possible because of the strategies, methods, and shared knowledge they use to attain an application’s goals which are encoded in MeTTa (Meta Type Talk) — an AGI language specialized for defining processes and encoding knowledge.
The central idea behind MeTTa is this: conventional programming (especially functional programming) and reasoning can both be represented as a chain of queries to a meta-graph. This meta-graph can store either program expressions or knowledge; thus, the knowledge and reasoning can be naturally combined. One key difference between MeTTa and many traditional languages is that a program can rewrite its own source code — creating advanced opportunities for self-optimization.
Another critical difference is that traditional reasoning engines assume purely symbolic reasoning within an assumed closed world. For example, if asked to count red cars, they will use their concept of ‘red car’ (from Deep Neural Networks) and search for regions in an image that matches the concept. By contrast, Hyperon will search for red regions and car-like shapes in the image, so that reasoning about the concepts themselves becomes part of the search. MeTTa can pay considerable attention to grounded reasoning over subsymbolic or raw data, and to neural-symbolic integration in particular. OpenCog Hyperon is designed to combine concepts, such as “red + car” or “blue + car”, instead of needing explicit training on “red car” and training on “blue car” — requiring less training data while simultaneously increasing the system’s understanding of fundamental concepts by using machine reasoning.
MeTTa is designed to be a highly practical programming language, with most of the structures commonly found in other languages, but it also embodies numerous advanced methods from programming language theory. This includes some highly technical components — dependent probabilistic gradual typing, semantics that incorporate paraconsistent logic, and non-well-founded set theory. This makes it straightforward in MeTTa to encode the logic of AI components to access, manipulate, and store knowledge in a way easily understood and reused by other components. This novel programming language is the key to AI collaborative problem-solving within the OpenCog Hyperon framework.
MeTTa was formalized in a research paper published by the OpenCog Hyperon team in March, “A Meta-Probabilistic-Programming Language for Bisimulation of Probabilistic and Non-Well-Founded Type Systems,” and an experimental Hyperon version with MeTTa example scripts can be located in this repository.
Distributed Atomspace (DAS), OpenCog Hyperon’s Knowledge Representation Database
If the AI modules are the processing centers of OpenCog Hyperon, the Distributed Atomspace (DAS) is the long-term memory where they store their knowledge. Domain-specific information (like the relationship between proteins and genes or the financial time series of the stock market) is stored together with more basic, high-level knowledge, such as the semantic relations between words or mathematical formulas and concepts.
The Distributed Atomspace stores this knowledge in a way the AI components can access: not with queries to a SQL database but with knowledge encoded in the MeTTa language itself. When an agent creates a new piece of knowledge, that also gets stored in the DAS, becoming available to other components. This enhances Hyperon’s comprehension of the concept, which can then be used to derive new knowledge (by combining concepts, as noted above), allowing for cognitive synergy — the foundation for AGI in the OpenCog Hyperon System. See more details on the project here.
Domain-Specific Languages for OpenCog Hyperon
Different application domains require different knowledge encoding, and specialized primitives to define processes. Building these different applications and plugging them together is where DSLs (Domain Specific Languages) are useful.
In OpenCog Hyperon, a DSL is a dialect of MeTTa with domain-specific primitives designed specifically to make it easy to encode the AI components and knowledge base of an application. Therefore, we have different DSLs for financial markets, biotechnology, natural language processing, etc. By adding these features to MeTTa, OpenCog Hyperon can work with other AI projects in specialized application domains.
The Road Ahead
As a result of the significant advancements we have made over the past year, we have arrived at a critical point where the development of OpenCog Hyperon is mature enough to yield real-world results. Together with the team at Rejuve Biotech, we have been able to leverage MeTTa and DAS in the field of genomic analytics. Now, we’re using our AI applied to bio-ontologies to discover patterns in Genescient’s long-lived fruit flies. What’s even more exciting is that we can take these patterns and use them to learn more about how humans live long lives.
While recent advances in deep neural networks have been impressive, the limitations of these narrow AI systems are also apparent when it comes to understanding and reasoning in real-world systems. OpenCog Hyperon can leverage these amazing advances in neural networks and generative AI by integrating these types of modules into its framework, while also providing a context that can overcome narrow AI’s inherent limitations.
While the progress in 2022 has been impressive, OpenCog Hyperon is just getting started. SingularityNET-supported initiatives will serve as the technological spearhead breaking through to the territory of artificial general intelligence (AGI) — ground that can only be gained with tools that allow all AI types, including these impressive neural networks, to collaborate. OpenCog Hyperon represents decades of research, refinement, and mission-driven labor singularly focused on building decentralized AGI. Artificial General Intelligence requires two things: a general model of thinking and a general model of knowing. Meta Type Talk (MeTTA) and the Distributed Atomspace (DAS) are the two wings, and their upcoming alpha releases in 2023 will allow AGI to begin to take flight in 2023.
Stay updated on the latest AGI and Decentralized AI developments with SingularityNET, Follow Us on:
- SingularityNET Discord
- SingularityNET Main Telegram Channel
- SingularityNET Announcement Channel
- SingularityNET Website
- SingularityNET Twitter and Facebook
- SingularityNET Jobs Page → We’re hiring
