“The first ultra-intelligent machine is the last invention that man need ever make.”

(Irving John Good, 1965)

Also available in Arabic, Portuguese, Indonesian, Spanish, Turkish, Italian, French, Russian

In the latest chapter in our journey to Beneficial General Intelligence (BGI), we are excited to share our 2023 progress report. It highlights the extraordinary body of work carried out by the SingularityNET core AGI team in collaboration with our spinoffs TrueAGI and Zarqa, our partners, and the wider OpenCog community.

In the report, we provide a thorough technical update showing how our cognition-level approach to BGI matured over the last year. Key achievements include advancing Hyperon’s cognitive components, the extension of AI-DSL (Artificial Intelligence — Domain Specific Language) project into the MeTTa (Meta Type Talk) SDK, and advancements in experiential learning through projects like ROCCA (Rational OpenCog Control Agent) and NARS (Non-Axiomatic Reasoning System).

We also outline a roadmap for future development, focusing on Hyperon Alpha Release, real-world interoperability with the SingularityNET Platform, expansion of Hyperon components, and implementation of the PRIMUS cognitive model.

With these critical foundations now in place, along with other breakthroughs in the AI field — mostly notably in Large Language Models and Transformer neural networks, but also in concurrent AI processing, new symbolic-AI hardware architectures, and other areas — we are firmly positioned to realize the full potential of Hyperon at scale and accelerate progress toward truly Beneficial General Intelligence.

This brief summary provides an overview of key concepts and summarizes the main takeaways from the report. Find the full report at SingularityNET’s Progress Towards Beneficial General Intelligence (BGI).

OpenCog Hyperon Overview

OpenCog Hyperon is an advanced cognitive framework aimed at achieving Artificial General Intelligence (AGI) by integrating various AI algorithms with a distributed metagraph known as Atomspace. developed by the SingularityNET Foundation, the OpenCog community, and other partners, it represents a leap forward from the original OpenCog framework, referred to now as OpenCog Classic. It addresses scalability and incorporates advanced mathematical tools and logic for a more profound system design.

OpenCog Hyperon represents a significant step in the quest for AGI, introducing a more scalable and powerful framework than its predecessor. It integrates different AI paradigms into a unified system, facilitating cognitive synergy among different algorithms and fostering collaboration that enables the system to mimic human intelligence by learning and reasoning from accumulated knowledge.

Throughout 2023, we made significant progress on Hyperon’s core components, including the MeTTa interpreter and the Distributed Atomspace (DAS), advancing towards the Hyperon Alpha Release. Efforts also focused on scalability and the prototyping of higher-level components, demonstrating Hyperon’s potential as an AGI R&D platform.

MeTTa Language

OpenCog Hyperon introduced a key addition in MeTTa — a decentralized, self-modifying language designed for cognitive computations, enabling interoperability within the architecture. It unifies logic, probabilistic, and functional programming, facilitating AI algorithm implementation and evolution within the Atomspace.

MeTTa forms the ‘universal translator’ that enables this wide range of AI systems to dynamically collaborate based on the common knowledge base of Atomspace (and enhance the knowledge base while doing so). MeTTa’s capability to support neural-symbolic reasoning and handling uncertainties (using probabilistic reasoning), makes it the strong and versatile tool that is crucial in our pursuit to develop AGI.

In 2023, we saw the MeTTa interpreter’s readiness for internal use, with subsequent developments in AI algorithm prototyping and DAS integration. Key achievements include the implementation of a minimal MeTTa core, a backward chainer in MeTTa, and preliminary work on inference control, showcasing Hyperon’s evolving capabilities.

We have also embarked on an exploratory collaboration with F1REFL3Y.io, involving the combination of MeTTa with the Rho calculus formalism and associated Rholang language, sophisticated mathematical tools originally developed for the RChain blockchain, aimed at providing MeTTa with an additional mechanism for scalable, decentralized concurrency.

Distributed Atomspace

Another important improvement presented in OpenCog Hyperon is the Distributed Atomspace (DAS), an extension of the Atomspace into a more independent component designed to support multiple simultaneous connections with different AI algorithms, providing a flexible query interface to distributed knowledge bases. SingularityNET and TrueAGI developers are integrating DAS and the rest of Hyperon into an “AGI-as-a-service” framework capable of efficiently and securely serving the AGI needs of a broad spectrum of customers.

AI-DSL in 2023

In 2023, the AI-DSL project’s roadmap saw significant changes, adapting to new technologies and tools. Early in the year, prototypes were developed using Idris for SingularityNET marketplace analysis, which later shifted focus to MeTTa development for AI-DSL capabilities, including chaining and typing functionalities.

Additionally, the emergence of systems like ChatGPT and HuggingGPT led us to explore these technologies. As a result, we decided to work on the SingularityNET Platform Assistant to provide an easy-to-use system for interacting with the SingularityNET Platform (for instance, requesting services or combined workflows in natural language via a LLM) as an additional track.

Incorporating LLMs into OpenCog Hyperon

At SingularityNET, we have been developing and using Deep Neural Networks (DNNs) both as services on the SIngularityNET platform and as OpenCog components for solving various tasks in robotics, dialog systems, bioinformatics, and other domains. Nonetheless, recent advances in LLMs and Transformer architectures, as well as the remarkable achievements of Large Language Models (LLMs) have prompted further exploration of their potential to contribute towards the goal of achieving AGI, be it through incremental augmentations or incorporation into broader AGI frameworks like Hyperon.

LLMs, despite their vast scope, face limitations like hallucinations, lack of memory, inconsistent knowledge acquisition, and challenges in multi-step reasoning. These limitations stem from their feedforward architecture, which involves hashing vast data amounts during training for recall during inference, limiting their ability to handle novel tasks or adapt to new rules or situations.

Nonetheless, LLMs’ extensive training data enables them to perform a wide range of tasks, positioning them as what might be called “General Narrow AI” due to their broad scope but limited generalization capabilities.

Beyond LLMs

Limitations of LLMs don’t mean they aren’t valuable for AGI development. Their flaws and achievements offer insights into AGI’s nature. To overcome LLMs’ limitations, the global community is enhancing them with technologies like knowledge graphs, memory embeddings, and web search plugins. These enhancements make LLMs more useful but they cannot achieve AGI.

Therefore, while LLMs are useful, advancing them towards AGI requires addressing these limitations through deeper integration with neural-symbolic systems or by enhancing them with neural components.

Neural-Symbolic Synergy

Integrating LLMs with Hyperon requires them to communicate in MeTTa language, allowing them to act as Atomspaces for querying and combining symbolic and neural processing to offset each other’s weaknesses. This process involves using MeTTa for prompt composition and chaining with operations like knowledge graph queries, enabling seamless interaction between components.

At SingularityNET and Zarqa, we’re exploring one approach that covers capabilities of Langchain and Guidance for prompt scripting and output translation, and another for converting natural language into MeTTa for storage and reasoning in Atomspace. These are part of a broader framework named MeTTa-Motto, exemplified in applications like the SingularityNET Platform Assistant and NARS integration for enhanced Q&A capabilities.

We’re also investigating deeper Hyperon-LLM integration, such as fine-tuning LLMs to understand MeTTa and enhancing communication with Atomspace through cross-attention mechanisms, even considering embedding LLMs within Atomspace. These efforts aim to leverage both explicit knowledge and reasoning towards AGI development.

Experiential Learning — ROCCA, NARS, Causal Reasoning, AIRIS

In 2023, the focus of OpenCog shifted to developing intelligent agents via Hyperon, incorporating projects like ROCCA for experiential learning. Efforts were concentrated on transitioning key components such as forward and backward chaining, PLN, and pattern mining to Hyperon. Hyperon/MeTTa’s advanced capabilities significantly expedited these transitions, enhancing component quality and generality. This includes a versatile chainer and the integration of the NARS experiential reasoning system, benefitting from MeTTa’s innovative features to streamline code and improve modularity and performance.

The initiative also aims to merge the best elements of various systems, including ROCCA, ONA (OpenNARS for Application), OpenPsi, and AIRIS (Autonomous Intelligent Reinforcement Interpreted Symbolism), into a comprehensive experiential learning framework within Hyperon. This framework focuses on cognitive schematics and introduces a curiosity model for efficient environmental exploration, showing the potential to outperform traditional Reinforcement Learning methods in terms of data efficiency. These advancements are currently being tested in grid world scenarios to validate their efficacy.

SingularityNET’s Progress Towards Beneficial General Intelligence

As we stand on the precipice of a new era in artificial intelligence, the journey from concept to reality for artificial general intelligence is reaching its final stages.

Our progress report for 2023 not only showcases the strides we have made in cognitive computing and AI integration. It also demonstrates our dedication to the ethical development and application of AGI.

The advances within OpenCog Hyperon, propelled by the dedicated efforts of the SingularityNET core AGI team, our partners, and the vibrant OpenCog community, are a testament to the collective ambition to harness AGI for the greater good of all sentient beings.

The collaboration with external developers and the open-source community has been instrumental in our journey — it has enriched our perspectives, accelerated our progress, and paved the way for truly beneficial AGI.

We are committed to developing Hyperon in an open, responsible, and ethical way, and we look forward to working with external and open-source developers to ensure that human-level AGI systems are beneficial, compassionate, and helpful to humanity.