The Near Future of AI is Action-Driven: Combining the ReAct Model and External Cognitive Assets

The Near Future of AI is Action-Driven: Combining the ReAct Model and External Cognitive Assets

Artificial Intelligence (AI) has made significant strides in recent years, and the near future of AI is action-driven. One model that exemplifies this concept is the ReAct model developed by Yao et al. (2022, arxiv). The ReAct model takes a three-step iterative approach: Thought, Act, and Observation. This model focuses on the idea of an AI system acting as an agent, making choices and taking actions based on its understanding of what is needed. This action-driven approach is a key component of what many consider to be true Artificial General Intelligence (AGI).

LLMs, or Large Language Models, have become increasingly popular in the field of AI. These models have shown impressive performance in question-answering tasks, particularly when prompted to "think step by step" (Kojima et al. 2022, arxiv). However, they can achieve even better results when equipped with external cognitive assets. By leveraging external resources and data, LLMs can bridge the resource gap and enhance their problem-solving capabilities.

OpenAI's 002-text-davinci model has demonstrated the power of combining instruction tuning and Reinforcement Learning from Human Feedback (RLHF) (blogpost). Through RLHF, humans rate the success of a given prompt, allowing the model to learn and improve over time. While this approach has shown promise, the true potential lies in actual reinforcement learning, where a system can be trained to produce better results based on a specific metric of interest.

In the world of AI, startups that are able to create powerful feedback loops will likely achieve great success. These startups will identify and address customer pain points, collect data on how to improve their solutions, and continuously train their models to be more consistent and effective. This iterative process will create a strong competitive advantage, serving as a moat in the AI landscape.

Transitioning from structured learning in school to self-directed learning can be a daunting task. Radi, in his blog post, shares his approach to structured learning after school (Radi's Blog). He highlights the difference between structured and unstructured learning, emphasizing the intentionality of structured learning. In structured learning, the path is guided by pre-defined objectives, while unstructured learning occurs more naturally without explicit direction.

Radi outlines three distinct phases in his structured learning approach: discovery of new ideas, sense-making, and experimentation. These phases are not bound by time, as they can take days or even months to complete. Sense-making is a crucial phase where Radi delves deeper into a selected topic, conducting research and creating visual knowledge maps on platforms like Heptabase. He also keeps track of the literature he consumes, such as books on Goodreads and written content on Glasp.

One challenge in post-school learning is maintaining motivation without the structure and accountability of formal education. Radi proposes that the public record of the internet can serve as a replacement for institutional oversight. By intentionally sharing his learning and progress in public spaces, he creates a sense of accountability and motivation to continue his learning journey. Platforms like Glasp allow him to annotate and store compelling written content, further contributing to the public record of his learning.

As we look to the future, the combination of action-driven AI models and structured self-directed learning holds great potential. Incorporating external cognitive assets into AI systems can enhance their problem-solving capabilities and bridge the resource gap. Meanwhile, leveraging the public record of the internet and sharing learning progress in public spaces can provide a sense of accountability and motivation for self-directed learners.

To make the most of this convergence, here are three actionable pieces of advice:

1. Embrace an action-driven approach: Whether you're developing AI models or pursuing self-directed learning, prioritize action. Act like an agent, making choices and taking steps towards your goals. This approach will yield better results and drive progress.

2. Leverage external cognitive assets: Don't limit your AI systems or learning journey to internal resources. Explore and utilize external cognitive assets, such as data and resources from the internet. By expanding your knowledge base, you can enhance your problem-solving abilities and achieve greater success.

3. Share and collaborate in public spaces: Take advantage of the public record of the internet to share your progress, insights, and discoveries. Engage with others in public spaces, fostering collaboration and accountability. By contributing to the collective knowledge, you can inspire others and receive valuable feedback and support.

In conclusion, the near future of AI is action-driven, with AI systems acting as agents and making choices based on their understanding of what is needed. The combination of the ReAct model and external cognitive assets holds great promise for advancing AI capabilities. Similarly, structured self-directed learning can be enhanced by leveraging the public record of the internet and sharing progress in public spaces. By embracing an action-driven approach, leveraging external resources, and engaging in public collaboration, we can unlock the full potential of AI and self-directed learning.