AI Agents and Compound AI Systems

Introduction to AI Agents

• 2024 identified as the year of AI agents.
• Definition of AI agents starts with generative AI trends.

Shift in Generative AI

Move from Monolithic Models to Compound AI Systems

• Monolithic models are limited by their training data.
• These models are hard to adapt and require significant resources for tuning.

Example:

• Planning a Vacation:
  • Query about vacation days
  • Models alone can give incorrect answers due to lack of personal data.

Advantages of Compound AI Systems

• Integrating models with existing systems increases effectiveness.
• Example Implementation:
  • Model accesses vacation database and retrieves correct information.
  • Correctly outputs that the user has 10 days left for vacation.

System Design Principles

• Compound AI systems have multiple components and are inherently modular.
• Components include:
  • Models: Tuned models, language models, image generation models.
  • Programmatic Components:
    • Output verifiers, query breakdown programs, database searching tools.
• This modular design allows for easier adaptability compared to tuning a single model.

Types of Compound AI Systems

• Retrieval Augmented Generation (RAG): A common compound AI system.
• Control Logic: Determines the path for query responses; defined by human input.

Role of AI Agents

• AI agents utilize LLMs for control logic.
• Improvements in LLMs allow for more reasoning capabilities.

Spectrum of Instruction:

1. Think Fast: Follow exact programming instructions.
2. Think Slow: Break down complex problems for better outcomes.

Components of LLM Agents

1. Reasoning:
   • LLM develops a plan to address problems.
2. Acting:
   • Use of external tools (e.g., web search, calculators).
3. Memory:
   • Stores inner logs and conversation history for personalization.

Configuring LLM Agents

• Popular approach: React (Reason and Act combined).
• Configuration process when using React agents:
  • Query input to LLM
  • Instructions: take time to think, plan act
  • Utilize tools for solution implementation
  • Observe results and iterate as necessary

Example of a Complex Query

• Determining Sunscreen Needs for a Vacation:
  • Steps include:
    1. Retrieving vacation days.
    2. Checking average sun hours.
    3. Understanding recommended sunscreen dosage.
    4. Performing calculations for total sunscreen needs.

Conclusion

• Compound AI systems are expected to evolve towards more agentic behaviors.
• Autonomy Spectrum:
  • Trade-offs between efficiency and complexity in system design.
  • For narrow, well-defined problems, programmatic approaches may suffice.
  • For complex tasks, the agentic route is more advantageous.
• Human in the Loop: Important for accuracy in AI systems.
• Encouragement to subscribe for further learning.