The Near Future of AI is Action-Driven: Combining Cognitive Assets and Task-Oriented Training

The Near Future of AI is Action-Driven: Combining Cognitive Assets and Task-Oriented Training

In the realm of artificial intelligence (AI), the concept of action-driven systems is gaining traction. It has been observed that Language Model Learners (LLMs) tend to perform better at question-answering tasks when they approach the problem step by step (Kojima et al. 2022, arxiv). However, their performance can be further enhanced by incorporating external cognitive assets. These assets, which can take the form of external resources, empower LLMs to think, act, and observe the outcomes of their actions iteratively. By leveraging cognitive assets such as search functions, code interpreters, and human interactions, LLMs can tap into a wider range of knowledge and tools.

The key to maximizing the potential of LLMs lies in their understanding of their own capabilities and the desired outcomes of the users. By harnessing the power of external cognitive assets, LLMs can unlock new possibilities and produce superior results. One particularly promising avenue of exploration is reinforcement learning, where LLMs can be trained to improve their performance based on specific metrics of interest. This approach holds the potential for significant advancements in AI capabilities.

On one side of the equation, we have the External Cognitive Assets that serve as force multipliers for LLMs. These assets encompass any function that takes text as input and produces text as output. By integrating searches, code interpreters, and human interactions into the LLM's workflow, these assets expand the scope of its problem-solving capabilities. The ability to access and utilize external resources effectively is crucial for LLMs to excel in their tasks.

On the other side, we have the task-oriented training that is essential for optimizing the performance of action-driven AI systems. This aspect poses significant challenges, as it requires developing techniques that can effectively train LLMs to operate in an action-driven manner. While some techniques, such as instruction tuning, appear relatively straightforward to implement, the overall process is complex. There is a need for continuous research and development to refine and enhance the training methods for action-driven AI systems.

The potential impact of action-driven AI extends beyond technological advancements. It also brings forth the possibility of a rebalance of power between algorithms and consumers. As AI becomes more adept at understanding user needs and delivering desirable outcomes, consumers may gain greater control over the AI systems they interact with. This shift can empower individuals and foster a more user-centric approach to AI development.

Drawing inspiration from the philosophy of Kyūdō, the Japanese martial art of archery, we find a parallel perspective on the importance of aims. In Kyūdō, the process of aiming is prioritized over hitting the target. The master advises the student to focus on the way they aim, their posture, breathing, and form, rather than fixating on the goal itself. This philosophy resonates with the idea that success lies in the process, rather than solely in the attainment of specific goals.

James Clear, in his book "Atomic Habits," captures this sentiment succinctly: "It is not the target that matters. It is not the finish line that matters. It is the way we approach the goal that matters. Everything is aiming." When we shift our focus from end goals to the aims we set for ourselves, we can design a daily life that is fulfilling in its own right. By immersing ourselves in the process and finding joy in the journey, we detach our happiness from the hypothetical finish line.

In conclusion, the near future of AI is action-driven, where LLMs harness external cognitive assets and undergo task-oriented training to enhance their problem-solving capabilities. This approach holds immense potential for advancing AI systems and empowering users. To make the most of this paradigm shift, here are three actionable pieces of advice:

1. Embrace external cognitive assets: Explore and integrate a wide range of external resources, such as search functions, code interpreters, and human interactions, into AI systems. By leveraging these assets, AI can tap into a broader knowledge base and enhance its problem-solving abilities.

2. Prioritize task-oriented training: Invest in research and development to develop effective training techniques for action-driven AI systems. This includes exploring reinforcement learning methods that enable AI to improve its performance based on specific metrics of interest.

3. Foster a user-centric approach: Strive for a rebalance of power between algorithms and consumers, empowering individuals to have greater control over the AI systems they interact with. By understanding user needs and delivering desirable outcomes, AI can create a more personalized and impactful user experience.

By embracing the action-driven paradigm, AI can transcend its current limitations and unlock new realms of potential. The fusion of external cognitive assets and task-oriented training holds the key to shaping a future where AI systems not only provide answers but also actively engage in the problem-solving process. As we navigate this exciting path, let us remember that success lies not only in reaching the target but in the way we approach and embrace the journey itself.