Andrew Ng on the Evolution of AI Agents and Future Prospects

What are AI agents?

Unlike traditional, simple input-output AI models, AI agents are systems that possess higher problem-solving capabilities by utilizing self-reflection and iterative learning. While it has been common for previous AI models to generate answers immediately from a single prompt, agentic AI can achieve more accurate and high-quality results by proceeding through the problem-solving process in stages.

The following processes are characteristic of such AI agents:

1. Problem analysis - Breaking down tasks and organizing necessary information.

2. External information retrieval - Using web searches or databases as needed to collect additional information.

3. Drafting - Generating text or code based on the information retrieved.

4. Self-evaluation and correction - Self-checking the quality of the draft and correcting any issues.

5. Final output optimization - Making multiple revisions as necessary to improve the level of completion.

In this way, AI agents can produce more flexible and accurate results than conventional AI.

Key design patterns for AI agents

Currently, the following four design patterns are considered particularly important in the field of AI agents:

1. Self-Reflection

Self-reflection is a technique where an AI agent reviews its own output and improves it while correcting errors.

For example, when having an AI generate program code, it is possible to re-evaluate that code and perform:

• Bug detection and correction

• Improvement of processing speed and efficiency

• Improvement of readability

This allows for the construction of a system that not only increases the accuracy of the initial output but also possesses self-correction capabilities.

2. Tool Use

Tool use is a mechanism by which an AI agent performs tasks while utilizing external tools and APIs.

For example,

• Perform a web search to retrieve the latest information.

• Use data analysis tools to provide statistical backing.

• Utilize natural language processing tools to generate more accurate text.

This method is particularly effective for tasks requiring specialized information and can significantly expand the capabilities of AI.

3. Planning

Planning is a design method where AI breaks down tasks into steps and executes them in the appropriate order.

For example, in an image recognition task:

1. Pose recognition - Analyze the movement of people in the image.

2. Object detection - Recognize backgrounds and objects.

3. Image generation - Create new images based on specified conditions.

By taking such a step-by-step approach, more advanced problem-solving becomes possible.

4. Multi-Agent Collaboration

Multi-agent collaboration is a method where multiple AI agents divide roles and perform tasks while providing mutual feedback.

For example,

• A code generation agent creates the code.

• A review agent points out bugs and areas for improvement.

• A test agent verifies the program's operation.

By building such a system, it is possible to produce high-quality deliverables in a short time that would be difficult for a single AI to achieve alone.

Examples of AI Agent Performance Improvement

Recent research indicates that agentic AI demonstrates higher performance than traditional prompt-based AI.

• GPT-3.5 (using standard prompting): 48% accuracy rate

• GPT-4 (using standard prompting): 67% accuracy rate

• GPT-3.5 (agentic workflow): 73% or higher accuracy rate

These results suggest that not only model improvements but also enhancements in workflow design can dramatically increase the practical utility of AI.

Future Prospects for Agentic AI

As the adoption of agentic AI progresses, the following evolutions are expected:

1. Establishment of more advanced autonomous learning capabilities

   • AI will autonomously execute tasks and repeat self-corrections to produce higher-precision results.

2. Efficiency gains through iterative task optimization

   • By leveraging rapid feedback loops, it will become possible to create high-quality deliverables in a short amount of time.

3. Deepening collaboration with humans

   • AI agents will specialize in specific fields and function effectively as human assistants.

Furthermore, as it could serve as a technical foundation for realizing Artificial General Intelligence (AGI), it will become a critical element that significantly influences the direction of future AI development.

The introduction of agentic AI is not merely a technological innovation; it holds new potential to improve productivity across society as a whole. We have entered an era where it is essential to pay attention to its future evolution and utilize it in actual business and research.