A MOONSHOT EFFORT TO BUILD AUTONOMOUS
AGENTS WITH NATURAL LANGUAGE ABILITIES
At Titan AI lab, we work on innovative ideas, concepts, and inventions in AI, VR, Metaverse, and Robotics that have the potential to significantly impact society and transform existing systems, processes, and ways of life.
We focus on projects involving groundbreaking discoveries and advancements in science, engineering, and technology that can lead to new ways of understanding the world around us, addressing complex problems, and potentially changing the fundamentals of how we live, work, and interact with each other.
Project Atlas aims to research, design, and implement a general architecture and processes for building machine intelligence —an autonomous cognitive agent with natural language ability— whose capability is on par with natural intelligence.
Cognitive agents that enhance human thinking could bring about immense advantages. They would dramatically change our lifestyle, work, and social interactions. Their ability to instantly process and examine vast amounts of information would improve our problem-solving and decision-making skills which would pave the way for breakthroughs in fields such as medicine, finance, education, and many more, where faster and more accurate processing can help tackle complex issues and drive societal benefits.
How do we build cognitive agents capable of understanding, learning, reasoning, and communicating their thinking in a language we comprehend?
The answer to this question lies in developing a general architecture (an artificial mind for intelligent agents) with distinct capabilities to model and simulate the world. After all, if cognitive agents are to solve similar everyday challenges as humans, both must possess a comparable model of the environment in which they act and a similar understanding of the events that take place.
To this end, our team has developed a new cognitive architecture that integrates various specialized modules such as language, reasoning, memory, and agent to generate a comprehensive world model that allows cognitive agents to understand and navigate their environment effectively.
The Titan Architecture enables agents to:
The key to unlocking the full potential of AI lies in finding ways to effectively capture the underlying structure of the world by creating a hierarchy of increasingly abstract representations that would allow agents to generalize to new possibilities more effectively and handle more complex and nuanced situations.
Cognitive agents could more efficiently process and understand the information they receive by breaking down high-level concepts into smaller, more manageable pieces and organizing them into a hierarchy.
Our representation methodology’s cornerstone is a repeatable anatomy of events and their relationships in time, space, and causality. By defining this anatomy, we aim to articulate how different world elements relate to each other and how they form a coherent whole. Our belief that relationships between events are fundamental to cognition underpins our overall approach.
By taking this perspective, we aim to provide a deeper understanding of the underlying structure of the world and how it gives rise to the complex systems and processes we observe.
Language is integral in constructing and reconstructing the world in the human mind. In essence, language controls the simulation in the mind.
Humans use natural language to encode the world into symbols (words) and depict the relationships between these symbols through a structure (grammar). This encoding process acts as a compression mechanism, allowing vast amounts of relevant information to be represented compactly and efficiently.
By using natural language, the world’s complexity is distilled into a manageable form that can easily be shared and processed by the human mind. In essence, natural language acts as the bridge between the world and our thoughts, providing a vital link in simulating our understanding of reality.
Our approach to building a natural language engine that accurately represents the world takes into account words, grammar, and the relationships and context they convey. The process requires a deep understanding of the linguistic structures that make up the language and a comprehensive world modeling engine to which these structures can be mapped.
Developing cognitive agents that can process information in real-time suggests the idea of a cognitive flywheel that enables the agents to acquire and reason about information they perceive continuously.
The flywheel processes unstructured text data into the exact hierarchical representation, which the agent can then use to make decisions and take actions in the world. As new information is acquired, the agent’s understanding of the world evolves, leading to an ongoing cycle of perception, cognition, and action.
While the idea of perception, cognition, and action is familiar and relatively simple to understand, implementing it in computational systems has been problematic. The challenge arises from a paradox between natural language understanding and world modeling. The agent needs knowledge of the language to understand the world but also some knowledge of the world (context) to solve the ambiguities in the language.
To overcome this paradox, we introduced several bootloaders into the AI system. These bootloaders, comprising miniature models of language and the world, serve as an initial booster to get the cognitive flywheel’s momentum going, especially since the agents possess no evolutionary history. The bootloaders provide minimal guidance to the AI system and are an initial starting point. Once the flywheel starts turning, the agents can continuously learn and improve, leading to a self-reinforcing loop of improvement and refinement.
The world engine provides a general-purpose reasoning tool capable of representing a wide range of information about the world, including physical, social, or abstract systems.
The engine provides a means for manipulating the world model to test hypotheses and predict future outcomes. The process uses inference, simulation, and prediction to help an agent understand and navigate the world around it. By combining real-time information with past experiences, an agent can form a mental model of the world that helps it make decisions and navigate complex situations.
Our goal with the world engine is to provide the agent with an interpretable understanding of the correlational and causal relationships between events. By simulating different scenarios and observing the changes in the world model, the agent gains a more profound knowledge of the underlying causes of events, improving the accuracy and reliability of its understanding of the world. And thus, the agent’s capacity to handle problem-solving and decision-making tasks.
The Titan architecture and processes (above) empower the development of autonomous agents with broad cognitive abilities. The agents can execute a vast array of tasks efficiently, without the need for explicit programming, and boast advanced AI functionalities such as real-time, continuous learning, seamless natural language comprehension, and innovative problem-solving in complex and ever-changing scenarios. Additionally, the agents feature a user-friendly and highly flexible interface, enabling users to interact with the system more intuitively and naturally.