Lecture on AI Agents and Agentic Workflows

Introduction

• Speaker: Andrew Ng, renowned computer science professor at Stanford.
• Known for contributions to neural networks, founding Google Brain, creating Coursera, and courses like deeplearning.ai.

Main Topic: AI Agents

• Current Use of Large Language Models (LLMs):

  • Non-agentic workflow: Typing a prompt to generate an answer.
  • Compared to writing an essay without backspace.
  • Despite challenges, LLMs perform well.

• Agentic Workflow:

  • Iterative process: Write outline, conduct research, draft, revise, and repeat.
  • Shows better results through iteration and revision.
  • Example: Coding problem benchmark (human eval benchmark by OpenAI).
    • Zero-shot prompting results in moderate success (48% for GPT-3.5, 67% for GPT-4).
    • Agentic workflow improves performance beyond GPT-4.

Key Concepts in AI Agents

• Reflection:

  • Process of having an agent review and revise its own outputs.
  • Example: Code writing where the AI reviews its own code for errors and improvements.

• Tool Use:

  • Integration of various tools (e.g., Copilot, GPT-4) to enhance functionality.
  • Originated in computer vision community due to earlier LMs' inability to process images.

• Planning Algorithms:

  • Allow AI agents to autonomously reroute around failures and complete complex tasks.
  • Example: Image generation task demonstrating multi-step planning.

• Multi-Agent Collaboration:

  • Simulating different roles (e.g., CEO, designer) within a single LLM.
  • Agents work together to achieve tasks, showing potential despite occasional unreliability.

Design Patterns and Their Importance

• Design Patterns Observed:

  • Reflection, Tool Use, Planning, and Multi-Agent Collaboration.
  • These patterns can potentially increase productivity and expand AI capabilities.

• Expectations and Challenges:

  • Rapid expansion of tasks AI can perform due to agentic workflows.
  • Need to adjust expectations regarding response time from AI agents.
  • Fast token generation is crucial for iterative processes.

Future Directions

• Looking forward to advancements in models like GPT-5, Gemini 2.0.
• Agentic reasoning could enable performance improvements on earlier models.
• The path to Artificial General Intelligence (AGI) is seen as a journey with agent workflows aiding progress.

Conclusion

• Agentic workflows offer significant potential for productivity enhancement and advancing AI technology.
• Encouragement for exploration and implementation of these patterns in practice.