AI Agents in 2024

Introduction to AI Agents

• 2024 is anticipated to be the year of AI agents.
• Understanding AI agents requires looking at shifts in generative AI.

Shifts in Generative AI

Monolithic Models vs Compound AI Systems

• Monolithic Models

  • Limited by training data.
  • Hard to adapt without significant investment.
  • Perform tasks like summarizing documents, drafting emails.

• Compound AI Systems

  • Integrate models into existing processes.
  • Example: Planning a vacation and retrieving personalized data.
  • Systems combine models with databases and tools.

System Design in AI

• Problems better solved with system design principles.
• Systems are modular, combining various components:
  • Tuned models, large language models, image generation models.
  • Programmatic components, output verifiers, queries breakdown.
  • Tools for database searching and additional functions.
• Faster adaptability and problem-solving than tuning monolithic models.

Example: Compound AI Systems

• RAG (Retrieval Augmented Generation)
  • Popular type of compound AI system.
  • Example query: Vacation policy database search.
  • Control logic determines the query path.

AI Control Logic

• Control Logic in Compound AI Systems
  • Programmatic control logic defines the problem-solving path.
  • Human-defined for specific contexts.

Introduction to AI Agents

• Agentic Approach

  • LLM (Large Language Model) takes charge of control logic.
  • LLMs demonstrate improved reasoning abilities.

• Two Approaches

  • Think fast: Act as programmed without deviation.
  • Think slow: Plan, solve iteratively, adjust as needed.

Components of LLM Agents

• Reasoning

  • Models plan and reason about steps in problem-solving.

• Action

  • Use external programs/tools (e.g., search, calculation, code execution).

• Memory

  • Stores inner logs and conversation history for personalized responses.

Configuration of AI Agents

• REACT (Reasoning and Acting) Agents
  • Combines reasoning and action components.
  • Example query process: Vacation planning and sunscreen calculation.

Practical Examples of AI Agents

• Vacation Planning Scenario
  • Assessing vacation days, sun exposure, recommended dosage, and math for sunscreen.
  • Modular system allows handling complex queries with multiple data sources.

AI Autonomy

• Scale of AI Autonomy

  • Trade-offs between autonomy levels and problem types.
  • Narrow problem sets are efficiently handled by programmatic approach.
  • Complex tasks benefit from agentic approach.

• Future Outlook

  • Compound AI systems and agentic behavior evolving rapidly.
  • Human oversight will continue as accuracy improves.

Conclusion

• Anticipation of significant progress in AI agent systems.
• Encouragement to subscribe to learn more.