The Path to AGI
1 Uncontrollable AI
The AI landscape has been notably transformed by the emergence of Large Language Models (LLMs). Their rise has undeniably brought us closer to the Artificial General Intelligence (AGI). Yet some challenges that are intrinsic to LLMs make us skeptical if they are the best way to reach AGI.
1.1 Controlling Probability
One of the nuances of LLMs lies in their control mechanics. While minor calibrations can be made to influence their outcomes, achieving a comprehensive grasp over their intricate functionalities is challenging. For instance, integrating a component as fundamental as long-term memory requires more than just a superficial adjustment. Existing methodologies like the RAG do offer potential solutions, but they are inherently limited and don’t genuinely emulate organic memory dynamics. Altering the model’s architecture is an option, but it mandates a comprehensive retraining process. And before we even delve into the complexities of alignment, it’s evident that mastering the technical aspects alone presents substantial hurdles.
1.2 Training Cost
Starting the training process of an LLM from the ground up is a resource-intensive endeavor. This doesn’t just impact the exploration of newer architectural innovations but also presents a barrier to entry for independent and smaller-scale researchers. Hence, while LLMs present a tantalizing direction toward AGI, they also raise questions about access, equity, and diversity in research.
Yet, it’s pertinent to acknowledge that the current trajectory of AI research heavily leans on the scaling potential of LLMs, fostering optimism about bridging the existing gaps.
2 Controllable AI
The notion of Controllable AI extends beyond just creating a compliant machine. It’s about understanding and delineating the balance between intelligence and controllability in artificial entities.
2.1 Dealing With Ambiguity
One of the historic challenges algorithmic models grappled with was ambiguity. People can deal with ambiguity by relaying on common-sense knowledge, which is not available to computers easily. Our idea is to fix this by creating an algorithm capable of processing ambiguity with the same depth and nuance as human cognition by weaving in elements of context and real-world knowledge.
2.2 Conceptual Processing and Learning
Our framework is based on the new pattern-matching techniques that can handle ambiguity. We process raw data (like text, audio, images) into a conceptual representation. It can then trigger different actions that are associated with the concept that an agent was able to find in the raw data. For example a question might trigger an action of thinking about an answer and then telling it. But the agent’s learning journey doesn’t end there. By employing pattern matching on its subsequent actions and the feedback it garners from the environment, the agent is able to understand its influence on its surroundings, continuously building it’s world model. This is made possible as both actions and outcomes are represented as graphs, allowing for a deeper, graph-based pattern recognition. The use of pattern-matching on the agent’s actions introduces an additional layer of non-linearity. Now, the agent’s behavior isn’t solely influenced by environmental input but also by its own potential responses to such input. Consider a scenario where the agent contemplates a potentially dangerous action; the pattern-matching mechanism identifies the associated concept of danger (from my actions), that concept will trigger an action to question the appropriateness of its intended action.
While this framework is ambitious and teeming with complexities, it underscores the potential pathways we could explore as we inch closer to AGI.
The interface between humans and computers has long been a limiting factor in fully harnessing computational power. Traditional methods, which rely heavily on programming languages, create a bottleneck, preventing most individuals from tapping into the vast potential that computers offer. With the advent of controllable AI, this paradigm is poised to shift. This new era promises a more intuitive interaction, bridging the gap between human intention and machine capability. By bypassing the intricacies of traditional programming, controllable AI paves the way for a future where we can readily access and leverage the dormant power of computers that we’ve been overlooking in our daily lives.