The Near Future of AI is Action-Driven: The ReAct model and the Potential of AGI

The Near Future of AI is Action-Driven: The ReAct model and the Potential of AGI

In recent years, the field of artificial intelligence (AI) has made significant strides, with researchers and developers constantly pushing the boundaries of what is possible. One promising development in the near future of AI is the concept of action-driven models. These models, such as the ReAct model proposed by Yao et al. (2022), aim to combine thought, action, and observation to create a more dynamic and intelligent AI system.

The ReAct model operates in three steps: thought, act, and observation. First, the model considers what is needed or required in a given situation. Next, it chooses an action based on its analysis and understanding of the situation. Finally, it observes the outcome of the action and learns from the results. This iterative process allows the model to continuously improve its decision-making abilities and adapt to changing circumstances.

What makes action-driven models particularly exciting is their resemblance to artificial general intelligence (AGI). While the definition of AGI may be a subject of debate among academics, it is clear that an action-driven language model (LLM) shares many characteristics with AGI. LLMs have already demonstrated their ability to perform well in question-answering tasks when prompted to "think step by step" (Kojima et al., 2022). However, by incorporating external cognitive assets, such as fetching data from external sources, LLMs can further enhance their performance and bridge the resource gap.

OpenAI's 002-text-davinci model, for example, has achieved remarkable results through a combination of instruction tuning and reinforcement learning from human feedback (RLHF). In RLHF, humans rate the success of a given prompt, allowing the model to learn from their evaluations. While this approach has shown promising results, it is likely that true breakthroughs will come from actual reinforcement learning, where the model is trained to produce better results based on a specific metric of interest.

In the future, we can expect startups and companies to capitalize on the power of feedback loops to create more effective AI systems. By identifying and solving customer pain points, collecting data, and training models to be more consistent, these companies can iterate and continuously improve their offerings. This iterative process will become the foundation of a moat in the AI industry, at least for now.

While the future of AI is undoubtedly exciting, it is important to consider other developments that are shaping the way we acquire and transmit knowledge. One such development is the YouTube revolution in knowledge transfer, as highlighted by Samo Burja. While modern video media may be seen as a distraction and a contributor to shorter attention spans, it also unlocks a form of mass-scale tacit knowledge transmission that was previously unavailable.

Tacit knowledge refers to knowledge that cannot be effectively transmitted via verbal or written instruction. It includes skills that are best learned through observation and imitation, such as creating art or assessing a startup. Before the availability of video at scale, tacit knowledge had to be transmitted in person, limiting its reach and accessibility. True autodidacts who can invent their own techniques are rare, while many individuals learn best by watching and imitating.

This is where YouTube and similar platforms come into play. These platforms have become a treasure trove of "how-to" videos, where individuals can learn various skills by watching and following along. YouTube reports a significant increase in searches in the "how-to" category, growing 70% year-on-year. This demonstrates the demand for video-based learning and the effectiveness of visual instruction.

The power of video-based learning lies in its ability to capture and transmit tacit knowledge. The camera's unflinching eye can capture details and nuances that even skilled individuals may not be consciously aware of. This opens up opportunities for distant collaboration and the creation of a truly open science. By sharing video demonstrations and tutorials, researchers and scientists can go beyond traditional methods of sharing information and foster a more transparent and collaborative environment.

However, it is essential to acknowledge the potential downsides of relying solely on video-based learning. While YouTube and similar platforms offer a wealth of knowledge, they also contribute to the fragmentation of information and may distract learners from engaging with longer-form means of communication, such as written articles or books. It is crucial to strike a balance between the convenience and accessibility of video content and the depth and rigor of traditional educational resources.

In conclusion, the near future of AI is action-driven, with models like ReAct leading the way towards more dynamic and intelligent AI systems. These models, with their iterative thought, act, and observation process, bear resemblance to AGI and have the potential to revolutionize various industries. Additionally, the YouTube revolution in knowledge transfer highlights the power of video-based learning and its ability to transmit tacit knowledge on a mass scale. By leveraging these developments, we can unlock new possibilities and create a more collaborative and informed society.

Actionable advice for individuals and organizations looking to harness the power of AI and video-based learning:

1. Embrace action-driven models: Explore and experiment with action-driven AI models like ReAct to enhance decision-making and problem-solving capabilities. Incorporate thought, act, and observation processes into your AI systems to enable continuous learning and improvement.

2. Leverage the power of video-based learning: Consider incorporating video-based learning into your educational initiatives or knowledge-sharing platforms. Create high-quality "how-to" videos that allow learners to observe and imitate skills and techniques. Ensure that these videos are easily discoverable through search engines to maximize their reach.

3. Maintain a balance between video and traditional educational resources: While video-based learning offers convenience and accessibility, it is essential to encourage engagement with longer-form means of communication, such as written articles or books. Encourage learners to explore a variety of resources to gain a comprehensive understanding of a subject.

By embracing these actionable advice, individuals and organizations can harness the potential of action-driven AI models and video-based learning to drive innovation, improve learning outcomes, and create a more knowledgeable society.