AI Agents 2024: Lecture Summary

Jul 21, 2024

AI Agents 2024: Lecture Summary

Introduction to AI Agents

  • 2024 marks the notable year for AI agents
  • Understanding AI agents involves shifts in generative AI

Shifts in Generative AI

Monolithic Models to Compound AI Systems

  • Monolithic Models:
    • Limited by training data
    • Hard to adapt (requires investment in data/resources)
  • Compound AI Systems:
    • Models integrated into existing processes
    • Solving tasks requires access to specific data sources
    • Example: Querying vacation days

System Approach in AI

  • Systems are inherently modular
  • Combining models (LLMs, image generation) with programmatic components:
    • Output verifiers
    • Programs to break down queries
    • Searching databases/tools
  • Easier and faster adaptation than tuning a monolithic model

Retrieval-Augmented Generation (RAG)

  • Popular compound AI system
  • Dependency on control logic
  • Programmable pathways for queries

Introducing AI Agents

Control Logic in AI Systems

  • Traditionally, defined by human programmers
  • Agentic Approach: LLMs controlling logic
    • Improvements in LLM capabilities (reasoning, breaking down problems)
    • Spectrum from fast, rule-based thinking to slow, plan-based problem solving

Capabilities of LLM Agents

  1. Reasoning: Core of problem-solving, planning steps
  2. Acting: Using external tools (e.g., search, calculator, code manipulation, translation)
  3. Memory:
    • Storing logs of thought processes
    • Conversation history for personalized interactions

REACT (Reasoning & Acting)

  • Combines reasoning and action in LLM agents
  • Encourages deliberate, plan-based problem solving

Example: Planning a Vacation

  • Components:
    • Retrieve vacation days
    • Check weather in Florida
    • Calculate sunscreen needs
    • Combine data from public health and personal history

AI Autonomy: Balancing System Efficiency

Sliding Scale of AI Autonomy

  • Degree of autonomy based on problem complexity
    • Narrow Problems: Programmatic systems more efficient
    • Complex Problems: Agentic approaches beneficial

Current State and Future Trends

  • Early days of agent systems
  • Rapid progress by combining system design and agentic behavior
  • Human-in-the-loop remains crucial as accuracy improves

Conclusion

  • Compound AI systems and AI agents evolving
  • Subscribe for more updates on AI advancements