coffee-gen-ai

What’s next for AI agentic workflows - Andrew Ng

Here are some notes based on this video.

slide 1:

• Agentic workflows are iterative which involves some thinking and revising.

slide 2:

• Example of HumanEval benchmark, by openai - [github] - released 2021

slide 3:

• Performance of gpt3.5 and gpt-4 on HumanEval benchmark in Zero shot and Agentic settings.

slide 4:

• Let’s explore these 4 design patterns

slide 5:

• Ask the LLM-itself to critic itself.
• Recommended read:
  • Self-Refine: Iterative Refinement with Self-Feedback, Maddan et al. (2023) - [paper]
  • Reflexion: Language Agents with Verbal Reinforcement Learning, Shinn et al. (2023) - [paper]

slide 6:

• Ask another agent to Critic the first Agent results.

slide 7:

slide 8:

slide 9:

slide 10: Agentic Reasoning Design Patterns

1. Reflection
   • Self-Refine: Iterative Refinement with Self-Feedback, Madaan et al. (2023)
   • eflexion: Language Agents with Verbal Reinforcement Learning, Shinn et al., (2023)
2. Tool use
   • Gorilla: Large Language Model Connected with Massive APls, Patil et al. (2023)
   • MM-REACT: Prompting ChatGPT for Multimodal Reasoning and Action, Yang et al. (2023)
3. Planning
   • Chain-of-Thought Prompting Elicits Reasoning in Large Language Models, Wei et al., (2022)
   • HuggingGPT: Solving Al Tasks with ChatGPT and its Friends in Hugging Face, Shen et al. (2023)
4. Multi-agent collaboration
   • ommunicative Agents for Software Development, Qian et al., (2023)
   • AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation, Wu et al.

slide 11: