Lecture Notes on AI Agents

Introduction to AI Agents

• 2024: Predicted to be the year of AI agents.
• Definition of AI Agents: Systems that utilize generative AI to perform tasks effectively.

Shift from Monolithic Models to Compound AI Systems

• Monolithic Models: Limited by training data and harder to adapt.
  • Example: Planning a vacation with a model that lacks personal data (e.g., vacation days).
• Compound AI Systems: Integrate models with existing processes and data sources.
  • Example: A system accesses a vacation database to respond accurately to queries.
  • Highlights the importance of system design and modularity.

Advantages of Compound AI Systems

• Systems can be designed with multiple components:
  • Models (tuned, large language, image generation).
  • Programmatic components (output verifiers, query breakdown).
  • Easier and faster to adapt compared to tuning a single model.

Retrieval-Augmented Generation (RAG)

• RAG Systems: A common type of compound AI system.
• Defined search paths can limit system functionality if the query doesn't match the database (e.g., weather queries vs. vacation policy).
• Control logic is crucial in navigating queries and responses.

Role of Large Language Models (LLMs) in AI Agents

• LLMs can take charge of control logic due to improved reasoning capabilities.
  • Allows for complex problem-solving where the LLM generates a plan.
• Agentic Approach: Enables a balance between fast action and thoughtful planning.
• Components of LLM Agents:
  1. Reasoning: Core problem-solving capability.
  2. Acting: Using external tools (APIs, databases, calculators).
  3. Memory Access: Storing internal logs and conversational history for personalization.

ReACT Framework

• ReACT combines reasoning and acting:
  • The user query is processed with a focus on thoughtfulness rather than quick responses.
  • Observations help determine the next steps based on the accuracy of responses.

Example Scenario: Vacation Planning

• Complex problem: Calculating sunscreen needs for a vacation.
  • Factors to consider:
    1. Number of vacation days (memory retrieval).
    2. Hours in the sun (weather forecast).
    3. Recommended sunscreen dosage (public health data).
    4. Math calculation for sunscreen bottles.
• Modular approach allows exploration of multiple paths to find solutions.

Future of Compound AI Systems

• Expected growth in agent technology in 2024.
• AI Autonomy Scale: Trade-offs between autonomy and programmatic efficiency.
  • Narrow, well-defined problems benefit from a programmatic approach.
  • Complex, varied tasks (e.g., GitHub issue resolution) benefit from LLM agents.
• Continuous improvement expected, with human oversight likely to remain important.

Conclusion

• Compound AI systems are becoming more sophisticated, and understanding their structure and operation is essential for future developments in AI.