TrueAGI, SingularityNET and F1R3FLY.io Partner to Bring Rholang Concurrent Processing to OpenCog Hyperon

TrueAGI and SingularityNET are proud to announce an exciting new partnership with Lucius Gregory Meredith’s company F1R3FLY.io aimed at dramatically accelerating progress toward true artificial general intelligence via integrating Meredith’s rholang software into the OpenCog Hyperon AGI toolkit. Rholang, founded on Meredith’s groundbreaking rho calculus mathematics, provides incomparable sophistication and efficiency at concurrent programming, which can leverage multiple processors simultaneously. 

Integration of rholang into the Hyperon framework, which itself is designed to run in TrueAGI’s application framework and leverage the SingularityNET decentralized coordination fabric, will enable Hyperon to fully leverage the power of modern multiprocessor architectures.

The collaboration will fast-track development of a scalable high-speed interpreter for Hyperon’s Meta-Type-Talk (MeTTa) language, which fully leverages concurrent processing and is fully blockchain ready, via incorporation of rholang on the back end. This blog reviews some of the particulars underlying this critical technical leap for the OpenCog project.

This is Your MeTTa-Brain on Rho Calculus

The Rho Calculus and why it is important

Meredith’s rholang language was designed as the centerpiece and smart contract language of the RChain blockchain, but the language itself has value and importance outside of any particular blockchain architecture, and indeed beyond the context of blockchain networks. The mathematical core of the rholang language is Meredith’s “rho calculus”, or “Reflective Higher-Order” process calculus, a significant step beyond the pi calculus formalism that has been in use since the early 1990s for the formal modeling of concurrent processes.

To get a bit more technical: Not only does the rho calculus satisfy the four desired blockchain properties of completeness, compositionality, concurrency, and complexity, it satisfies an additional property called reflection. Reflection allows a program to inspect or operate on itself by modifying its data to create a newer version of itself, thus enabling true metaprogramming, a necessary component for any real AGI. The pi calculus is powerful but doesn’t incorporate reflection at the fundamental level, making it awkward for formalizing various sorts of complex concurrent processes like self-organizing biological systems or AGI systems.

OpenCog Hyperon’s AGI Architecture

OpenCog Hyperon will be familiar to regular followers of the SingularityNET ecosystem – it is the next generation of the open source OpenCog AGI framework, begun in 2008, that underpins our team’s vision for truly general AI. Hyperon’s architecture is itself founded upon recent breakthroughs in foundational and applied mathematics (homotopy type theory, probabilistic type theory, Galois connections, …) that allow for more native and scalable computations of the sorts of cognitive synergies we believe needed to accelerate progress towards AGI. 

As compared to ChatGPT, Google Bard and other recent deep neural net based AI systems, Hyperon is much more heavily focused on deep understanding and reasoning, rather than on statistical pattern recognition and synthesis. 

For more depth on the limitations of ChatGPT type systems, see this recent post by Dr. Ben Goertzel 🡥 

A Hyperon-based AI system’s knowledge is grounded in its life-experience and its integrative thinking rather than just pieced together from fragments of its input data (see Stephen Wolframs very deep dive into how ChatGPT generates text in a purely predictive manner, based on highest probability of next word in the sentence, with no understanding of what it’s writing). However, Hyperon systems can be interfaced with generative models and other deep neural nets, extracting knowledge items from them and then reasoning about these items and integrating them with its own holistic understanding.

The Hyperon architecture centers on the Atomspace, a weighted labeled metagraph knowledge store that is used to encode all the types of knowledge needed for a human-like intelligent mind, including the reasoning and learning processes themselves. This is the memory and learning system of OpenCog, storing concepts and their relationships. This is the center of the ‘symbolic’ component of OpenCog’s neural-symbolic architecture, which distinguishes it from purely neural-net based systems.

This symbolic component, the Atomspace, provides a route to resolving the core difficulty found in creating systems that leverage the full scope of insight achieved by the AI community over the last half-century. It is commonly difficult to have multiple AI algorithms, such as probabilistic inference, neural networks, and evolutionary methods, working together to solve a single problem since they work quite differently. The differences in the internal dynamics and intermediate representations of the different algorithms can make it difficult for meta-learning to find patterns in context-specific algorithm applications that bind together aspects of the internal operations of one algorithm, with aspects of the internal operations of another algorithm. Using the Atomspace as a common framework enables one to do meta-learning in a way that spans multiple algorithms. 

In particular, Hyperon’s Probabilistic Logic Networks (PLN) system addresses uncertain inference, its Meta-Optimizing Semantic Evolutionary Search algorithm (MOSES) provides procedural learning, and the Economic Attention Allocation (ECAN) system provides resource allocation support. Metagraph-based pattern matching and pattern mining algorithms enable the sorts of complex queries missing from more traditional systems.

To more fully understand the cognitive architecture in which these aspects are assembled, see the paper “General Theory of General Intelligence” 🡥 or the associated series of video lectures 🡥

The MeTTa language is the native language of the Atomspace, allowing human-readable representation and creation of Atomspace structures. The MeTTa programming language interpreter provides access to the in-RAM Atomspace and also the persistent Distributed Atomspace (DAS) which provides long-term memory and resides largely on disk. The SingularityNET spinoff TrueAGI is creating an enterprise-scale application framework wrapping up these components, to enable them to straightforwardly be deployed for various practical applications via an “AGI-as-a-Service” methodology.

Due to the central role of MeTTa in Hyperon, if one can accelerate MeTTa’s performance on multiprocessor machines by integrating it with rholang, one can then tremendously speed up Hyperon’s performance. This is complementary to other speedup strategies being pursued, such as the Metagraph Pattern-Matching Chip and AGI board being co-developed by TrueAGI and Simuli.ai. Putting these various accelerations together we aim to obtain a Hyperon implementation that is hyper-efficient in the same sense that modern GPUs and associated software libraries provide hyper-efficient implementations of deep neural nets and other matrix multiplication based AI algorithms.

The recent successes of deep neural nets have largely been due to the rise of GPUs and multi-GPU machines, which allowed accelerated execution of long familiar neural architectures, and accelerated experimentation with new variants of these. We believe that the combination of Hyperon accelerants that we are now developing can serve a similar accelerative role for multi-paradigm integrative AGI – which may be able to accelerate us all the way to and through a Singularity.

HyperCycle and NuNet

Two additional parts of the SingularityNET ecosystem’s unique approach to AGI are the HyperCycle and NuNet frameworks – each of which has the potential to leverage rholang and the rho calculus, in ways overlapping with but also going beyond rholang’s use in the Hyperon AGI framework.

HyperCycle

One of our aspirations is to implement a massively distributed Hyperon system – or a massively distributed AI system integrating Hyperon with other sorts of AI components like deep neural nets – on a decentralized network of computers without any central controller. This will allow such systems to operate in an extremely robust way, and to incorporate components owned and operated and created by different parties with relations of partial trust to each other.

To fulfill this aspiration requires either a very fast, cheap and scalable blockchain infrastructure, or something else quite similar. Toward this end SingularityNET has partnered with the TODA network to create HyperCycle, a new blockchain technology designed for inexpensive, high-speed, large-scale on-chain execution of microservices, especially (but not exclusively) those carrying out AI-related functions. HyperCycle will be extended to interface with multiple current blockchains; but its architecture is fundamentally different from any existing blockchain due to its lack of any distributed or replicated ledger. It is more fully peer-to-peer, with each HyperCycle node containing a record of its own history and appropriately encrypted records of the histories of some of its neighbors and other network members, and consensus occurring within local subnetworks rather than globally.

HyperCycle can work with multiple smart contract languages, and its peer-to-peer architecture can provide radical acceleration of execution of nearly any smart contract language due to requiring massively fewer nodes to participate in consensus. However, HyperCycle will achieve its greatest efficiency when leveraging its own native smart contract language – which will be nothing other than MeTTa, utilized with a special smart-contract type system designed for straightforward formal verifiability. The integration of MeTTa with rholang will accelerate the ability to leverage MeTTa in this way because rholang is already blockchain-ready due to its integration with (ledger-based) Rchain.

NuNet

NuNet is a SingularityNET spinoff providing a platform for decentralized coordination of diverse compute processors to carry out various complex tasks. NuNet is building an open-source, globally decentralized platform, along with protocols, open APIs, and blockchain integrations, to run complex geographically distributed computing workflows, focused on but not restricted to AI processes running on SingularityNET. By harnessing latent computing power of independently owned compute devices, such as home computers and phones, crypto mining farms or private corporate clouds, the NuNet platform will support the computational infrastructure of decentralized AI agents required by Hyperon and HyperCycle. 

Alongside its use to help coordinate Hyperon processes across multiple processors on a single machine, the rholang back-end to MeTTa will help MeTTa process to more effectively coordinate across multiple compute resources coordinated within NuNet.

Scaling to the Singularity

We have outlined multiple aspects of the SingularityNET ecosystem’s approach to decentralized AGI – OpenCog Hyperon, Hypercycle, NuNet, the AGI board, and the SingularityNET protocol that can be used to connect Hyperon with deep neural nets and other algorithms within this decentralized tech stack. We have also explained how F1R3FLY.io’s rholang framework, originally developed within Rchain based on a long history of work on the rho calculus, has the potential to accelerate the operation of various portions of this software framework, particularly in the context of its exploitation of concurrent processing capabilities.

It is also interesting to view all this in the perspective of contrast to the software stacks developed in recent decades by Big Tech companies. From a technical angle, one can view the successes of these companies as founded largely on effective leveraging of concurrent and distributed processing. Most of their AI algorithms have been open source, and indeed were originally derived from open source work by university PhD students. However the code frameworks used to efficiently deploy these AI tools on large networks of multi-GPU and -CPU servers have been only partially opened up, and due to their nature have been of limited use to anyone without a massive centralized server farm.

Specifically, along with other tools, Google, Facebook, and Amazon have successfully implemented fantastic specialized back-end concurrent-processing software libraries (largely based on MapReduce) which have dramatically sped up processing of their neural nets and other statistical learning algorithms. However, these standard Big Tech concurrent-processing methods are too simplistic and not well-suited for many of the subtler metagraph-based algorithms leveraged in OpenCog, which we believe are likely to comprise the shortest path to human-level AGI. Current concurrent-processing methods are also not easily amenable to heterogeneous decentralized networks.

So we encounter the necessity for developing algorithms, hardware, and an entire infrastructure supporting these more complex metagraph approaches on distributed networks of multiprocessor machines. The Rho calculus, coupled with Hypercycle and NuNet; TrueAGI’s artificial intelligence-optimized hardware and AGI board built in collaboration with Simuli’s AGI board, and the SingularityNET platform itself, can provide Hyperon with precisely this requisite infrastructure. And so the new partnership between TrueAGI, SingularityNET and F1R3FLY.io provides one more key ingredient to the novel tech stack we are building to support the creation of a beneficial and decentralized Technological Singularity.

Originally written by Matt Iklé.