Lecture Notes: The Year of AI Agents (2024)

Introduction to AI Agents

• 2024 is predicted to be the year of AI agents.
• Understanding AI agents requires examining shifts in generative AI.

Shift from Monolithic Models to Compound AI Systems

• Monolithic Models:

  • Limited by data they are trained on.
  • Hard to adapt and tune, requiring significant data and resources.

• Compound AI Systems:

  • Example: Planning a vacation and querying vacation days.
    • A single model cannot access sensitive personal info like vacation days.
    • Integrating the model with a database to fetch accurate information.
  • Compound systems apply system design principles for better problem-solving.
  • Components of Compound Systems:
    • Modular with multiple components (e.g., tuned models, programmatic tools).
    • Include output verifiers, query breakdown mechanisms, database integration.

Programmatic Control Logic

• Popular Example: Retrieval Augmented Generation (RAG).
  • Example: Queries that fail if they deviate from the programmatic path (e.g., weather query to a vacation policy database).
  • Control logic defines the path to answer a query.

Introduction to Agentic Systems

• Role of Large Language Models (LLMs):

  • LLMs can reason and create plans to tackle problems.
  • Shift from fast, strict programmed controls to slow, deliberate reasoning and planning.

• Components of LLM Agents:

  • Reasoning:
    • Core problem-solving capability, breaking down issues and planning solutions.
  • Acting (External Tools):
    • Examples: Web search, databases, calculators, translation models.
    • Can call and utilize various tools to execute solutions.
  • Memory:
    • Storing inner logs for problem-solving and conversation history with humans.
    • Enhances personalized interactions.

• Configuration Example: REACT Agents

  • Combines reasoning and acting capabilities.
  • Process:
    • User query -> Model with prompt -> Slow thinking and planning -> Executes actions using external tools -> Iterates based on feedback -> Final answer.
  • Example: Calculating needed sunscreen for a vacation.
    • Combines vacation days, weather forecast, public health recommendations, and mathematical calculations.

Future and Application of Compound AI and Agentic Systems

• Future: Compound AI systems will evolve to become more agentic.

  • Sliding scale of AI autonomy for system designers.
  • Narrow, well-defined tasks still benefit from programmatic approaches.
  • Complex tasks require agentic approaches.
  • Examples: Managing GitHub issues, handling diverse queries.

• Current State:

  • Early days of agent systems.
  • Rapid progress in combining system design and agentic behavior.
  • Continued human oversight as accuracy improves.

Conclusion

• Anticipated growth and development in AI agents and compound systems.
• Encouragement to follow up for more information.