Skip to main content

What's next for AI agentic workflows ft. Andrew Ng of AI Fund - YouTube

In a recent presentation, Andrew Ng highlighted the evolution of AI agents as a critical trend for AI developers. He emphasized the shift from a non-agentic workflow, where an AI generates a single output, to an agentic workflow, which involves iterative processes such as drafting, researching, and revising to achieve significantly better results.

One case study demonstrated that an agentic workflow wrapped around GPT-3.5 outperformed even GPT-4 in a coding benchmark. This suggests that the agentic approach can enhance AI applications, making it a term to watch beyond the buzz in consultant reports.

Ng outlined four design patterns in AI agents:

  1. Reflection: AI systems can review and improve their own output, spotting errors and suggesting fixes.
  2. Tool Use: AI can now integrate with various tools for analysis, information gathering, and action, expanding its capabilities.
  3. Planning: AI agents can autonomously navigate tasks and reroute around failures, although this technology is still emerging.
  4. Multi-Agent Collaboration: Multiple AI agents can work together, simulating roles like CEO, designer, or tester, to collaboratively complete complex tasks.

These patterns offer a productivity boost and are becoming increasingly important in AI development. Ng predicts that the scope of tasks AI can handle will dramatically expand due to agentic workflows.

An important trend is fast token generation, which is crucial for iterative agentic workflows. Generating tokens quickly, even from a lower quality LM, might yield better results by enabling more iterations.

In conclusion, while it may require a shift in how we interact with AI—learning to delegate tasks and wait for results—agentic reasoning and design patterns are poised to significantly advance AI capabilities.

The original article: https://m.youtube.com/watch?si=PogGR-jXL4DvwPPV&v=sal78ACtGTc&feature=youtu.be