Technical Progress Toward Beneficial AGI: Insights from SingularityNET’s Technical Tuesdays
In the relentless pursuit of Artificial General Intelligence (AGI), SingularityNET has been at the forefront of innovation, continuously pushing the boundaries of what’s possible in the realm of AI. In the second installment of our Technical Tuesdays series, SingularityNET’s team provided a comprehensive overview of the strides they’ve made towards achieving Beneficial AGI (BGI).
Haley Lowy, marketing lead at SingularityNET, introduces the AGI experts present to discuss the progress made in the BGI (Beneficial AGI) section of SingularityNET’s 2023 blog. The team covers the development of MeTTa, with Alexey providing an overview of the achievements in 2023. The main accomplishment was the creation of a minimal MeTTa, which led to rewriting the ad hoc implementation using fewer basic primitives. This change opened up possibilities for further studies, including the implementation of different inference engines. The team also worked on developing other components that are part of the Prime architecture.
MeTTa: Pioneering Symbolic Language for AGI
The team discussed their progress in integrating neural and symbolic systems through a prototype called MeTTa. MeTTa allows for reasoning and interoperability with knowledge-based queries, connecting various components in one place. Another focus was on the development of AI-DSL, a native MeTTa-program and foundational service on the SingularityNET platform. AI-DSL is intended to be a MeTTa-service that can be mixed and matched based on user needs, interoperating with other services to create more complex AI solutions. The goal is to provide users with a streamlined process for accessing specific AI services based on their input data.
One of the focal points of discussion was MeTTa, a symbolic language designed to facilitate the creation of concise expressions comprehensible to both experts and non-programmers alike. The team highlighted MeTTa’s user-friendly interface, which distinguishes it from other frameworks, making it accessible even to those without extensive programming skills. This inclusivity aligns with SingularityNET’s vision of democratizing AI and expanding the decentralized AI community.
Integration of Neural and Symbolic Systems
Another significant advancement showcased during the session was the integration of neural and symbolic systems through prototypes like MeTTa-talk. This integration enables reasoning and interoperability with knowledge-based queries, paving the way for more holistic AI solutions. Moreover, the development of AI-DSL (Domain-Specific Language) highlights SingularityNET’s commitment to providing users with versatile tools for creating AI solutions tailored to their specific needs.
Dr. Matt Iklé discusses the progress made in achieving autonomous chaining of AI services to create more intelligent systems. Dr. Iklé and his team have made significant strides in program synthesis and composition, which are essential tools for this goal. He also mentions the use of AI-DSL, which enables different AI services to string together to form a more intelligent system.
The conversation then shifts to neuro-symbolic AI and how MeTTa enables this integration. Neuro-symbolic AI combines neural networks and symbolic reasoning, and MeTTa provides interesting opportunities for this integration by using the same language for querying knowledge graphs and communicating with AI models. Dr. Iklé also touches on the history of neural symbolic integration, specifically the use of cognitive module networks in OpenCog Classic, and the motivation behind building Hyperon and MeTTa to make these operations smoother. However, deeper integration has not yet been explored due to the size of LLMs during 2023.
The discussion revolves around the integration of large language models (LLMs) with knowledge graphs and how this neuro-symbolic integration can enhance the capabilities of LLMs. Currently, LLMs operate through inputs and outputs, querying knowledge graphs for information to generate responses. However, the researchers believe that intermediate layers of attention heads can query knowledge graphs during inference, allowing for a more synergistic relationship between the two.
One possibility is to have graph embedding of atom space, making the meta language itself differentiable. The ultimate goal is to create a system where knowledge can be stored and reasoned upon in a way recognizable to humans, allowing for a more human-like understanding and interaction. OpenCog, a system designed from the ground up, is an example of this approach. While LLMs process information in a bottom-up manner, humans reason in a more top-down way, using a consistent or parallel world model and imagination to perform conditional inference. This difference allows humans to reason and filter information effectively, reducing the hallucinatory nature of LLMs.
Then the discussion shifts to the limitations of reflex-based AI systems and the need for more advanced capabilities such as goal setting, reasoning, and consistent World model knowledge. Alexey uses the example of Minecraft agents to illustrate the difference between reflex-based behavior and having an overall picture of the situation. He also mentions the importance of scalability in AI processing and introduces MeTTa, a system they are working on with the OpenCog Hyperon team to understand and scale its operations. The team has developed an operational semantics for Meta and published it, with links provided in the chat for interested viewers.
Alexey discusses the challenges of scaling software systems, particularly in the context of blockchain technology. He explains how MeTTa, a Facebook-inspired project, operates with a minimal commitment to concurrency or sequential processing. The speaker argues that the traditional approach to scaling software systems by adding hardware has limitations, especially in the case of blockchain networks, which become slower as more hardware is added.
To address this issue, Alexey introduces F1R3FLY and Rune, which enable code to utilize as much physical hardware as possible and run faster with more processing. He also emphasizes the importance of decentralization in today’s software world and demonstrates a typical deployment of MeTTa across multiple servers in various data centers, with each server containing a copy of the entire atom space and accessible to the public. The ultimate goal is to create a public network that can withstand bad actors’ attempts to manipulate the system.
Scalability, Concurrency, and Decentralization
The importance of scalability, concurrency, and decentralization in AI processing was underscored throughout the discussion. SingularityNET’s F1R3FLY network, powered by tokens and compilers, ensures robust security and efficient processing. By leveraging F1R3FLY, users can seamlessly access AI services while mitigating the risk of bad actors exploiting the network. This emphasis on decentralization reflects SingularityNET’s dedication to creating AI systems resilient to manipulation and abuse.
The importance of tokens and compilers in building scalability, concurrency, and decentralization for the F1R3FLY network. Greg Meredith explains that tokens act as a barrier for bad actors, preventing them from overrunning the network due to the finite token supply. Legitimate users, on the other hand, can submit their tokens and MeTTa code, which is then compiled and run on each Firefly node for verification.
The F1R3FLY network provides security and analysis tools, such as the Operational Semantics in Logical Form (OSF) algorithm, which can detect issues like leaking secrets, deadlocks, and race conditions at compile time and serve as a query language for investigating programs with specific properties. These capabilities hint at the potential for sophisticated AI reasoning about programs in the near future. However, Greg also expresses concerns about the limitations of large language models, which have been trained on repositories like GitHub and may replicate the same bugs committed by programmers.
The discussion went further into the importance of autonomous reasoning in artificial intelligence (AI) and the evolution of computation models. Greg reflects on how human psychology influences the way we think about computing and society. He notes that it was not until the 1990s that we had models of computation without a top dog or controller, which is essential for understanding autonomous AI and human communication. The speaker also shares his personal experience recognizing the autonomy of bees and the importance of acknowledging the autonomy of agents at every level. The team also touches upon use cases and future plans for the Hyperon project.
Hyperon: Towards Beneficial AGI
The development of Hyperon, the new generation of OpenCog, represents a significant milestone in SingularityNET’s journey towards beneficial AGI. Hyperon’s focus on experiential learning and collaboration with various industries highlight SingularityNET’s commitment to creating AGI that not only benefits humanity but also aligns with real-world applications. By leveraging insights from neuroscience and complex dynamical systems, Hyperon aims to transcend traditional AI paradigms, paving the way for more sophisticated AI systems.
Dr. Iklé discusses the development of the new generation of OpenCog, named Hyperon. Hyperon is being used in various applications, including biomedical research and financial domains. The team is using use cases to guide the development of Hyperon, with ongoing projects in the legal, motorsports, and smart cities domains. The ultimate goal is to create artificial general intelligence based on research from neuroscience, complex dynamical systems, and the latest language design foundations. Hyperon is being developed in collaboration with companies and use cases to ensure its features align with real-world applications.
The importance of creating beneficial Artificial General Intelligence (AGI) and the role of experiential learning in its development. They argue that AGI cannot be created in a vacuum and that nurturing aspects and teaching from humans will be crucial. The team is also working on experiential learning projects and will delve deeper into this topic in the next Technical Tuesday. Upcoming milestones include the Alpha release, which is an experimentation point, and the documentation milestone, which is currently in development and will be available on the website.
The team discusses their plans for sharing their work on Hyperon, a technology primarily focused on AI research and development. They are preparing a website and educational resources for users, and will address community questions on the site. Hyperon is intended for AGI developers, and their post-alpha release plan includes the Prim reasoning engine, performance improvement, and concurrency support. They are also cooperating with the SingularityNET platform to develop a MeTTa SDK, which will be the basis for a knowledge layer on blockchains. The team emphasizes that Hyperon goes beyond just the Hyperon platform and has important implications for blockchain technology as a whole.
They discuss the MeTTa SDK, a new development for the SingularityNET platform. The MeTTa SDK aims to enable interoperability between services by using a common MeTTa component for interaction. This idea came from both the platform and Hyperon sides, as the team wanted to call services deployed on the platform directly from Hyperon without the need for downloading and deploying them locally. MeTTa is a typed language, and the team plans to provide type signatures and specifications for services. The next steps involve getting people familiar with the MeTTa SDK, which will be a specification language for services but will also serve as a DSL in MeTTa
Looking Ahead
As SingularityNET marches forward, they remain dedicated to sharing their progress and insights with the broader community. Plans for sharing their work on Hyperon, including the development of a MeTTa SDK, highlight their commitment to fostering collaboration and knowledge exchange within the decentralized AI ecosystem. By embracing transparency and inclusivity, SingularityNET aims to catalyze the development of beneficial AGI while addressing the ethical implications and societal impact of advanced AI technologies.
SingularityNET’s Technical Tuesdays provide a compelling glimpse into the cutting-edge research and development efforts driving the advancement of AI towards Beneficial AGI. With a steadfast commitment to innovation, collaboration, and ethical responsibility, SingularityNET continues to chart a path towards a future where AI serves as a powerful force for good.