From Monolithic Models to AI Agents: A New Era of Generative AI

(Images made by author with Microsoft Copilot)

Large Language Models (LLMs) have demonstrated impressive capabilities, yet their tendency to produce incorrect or misleading information, known as hallucinations, limits their reliability for tasks requiring high accuracy. To address this challenge, agentic AI has emerged as a promising solution. Recently, IBM Technology shared an insightful discussion on AI agents in a YouTube presentation. In this post, we will share the key insights from that presentation.

Table of Contents

The Shift from Monolithic Models to Compound AI

Monolithic AI models, such as LLMs, are limited in their capabilities because they are restricted by the data they were trained on. This limited data restricts their knowledge and problem-solving abilities. While these models excel in tasks like summarizing documents or drafting emails, they struggle with more complex or personalized requests. The true potential is unlocked when systems are built around these models, integrating them into existing processes.

Compound AI systems offer a solution by integrating AI models with components like databases, search engines, and programmatic elements. This approach enables systems to access external information and perform actions, making them more versatile and adaptable. For instance, a compound AI system can combine a language model with a database to accurately retrieve and provide information in response to user queries.

One type of compound AI system is Retrieval-Augmented Generation (RAG). It performs well with specific queries but struggles with questions that fall outside its predefined scope. This limitation stems from its fixed control logic programmed by humans, which restricts the system’s adaptability.

Introducing AI Agents: A More Autonomous Approach

A more autonomous solution is AI agents, which leverage the improved reasoning capabilities of LLMs to dynamically control the system’s logic and decision-making process. Instead of following rigid programmatic instructions, as seen in compound AI systems, these agents can “think slow,” break down complex problems, and use external tools to arrive at a solution.

AI agents possess three primary capabilities: reasoning, acting, and accessing memory. They reason by planning and evaluating steps, act by utilizing external tools such as web searches or calculators, and access memory by storing previous interactions or internal thought processes.

One common agent configuration is ReACT (Reason + Act), which enables LLMs to tackle complex problems through iterative planning and execution. For instance, calculating the number of two-ounce sunscreen bottles needed for a Florida vacation would involve these steps:

Reasoning Steps

Determining the problem: Calculating the number of sunscreen bottles needed for a Florida vacation.
Identifying relevant factors: trip duration, expected average daily sun exposure, recommended sunscreen dosage, and bottle size.
Breaking down the problem: Understanding the relationship between these factors and the desired outcome (number of sunscreen bottles).

Acting Steps

Accessing the Florida weather forecast for the planned vacation dates to determine average daily sun hours expected.
Consulting a knowledge base for sunscreen dosage recommendations.
Using a calculator tool to perform computations based on the gathered data and identified factors.
Providing a precise recommendation for the number of sunscreen bottles.

Choosing the Right AI System : Matching System Complexity to Problem Complexity

The choice of the right AI system depends on the problem’s complexity. Simple tasks are well-suited to pre-programmed logic, while complex and varied tasks are better handled by AI agents with more flexibility. This adaptability marks a significant step toward more autonomous and sophisticated AI systems.

Conclusion

By understanding the transition from monolithic models to compound AI systems and the rise of AI agents, we can better appreciate the advancements in generative AI and their potential applications. Agentic AI offers a promising path forward, enhancing accuracy and enabling more complex problem-solving capabilities.

Markmap of key ideas

click to open

Reference

Murad, Maya. “What are AI Agents?” IBM Technology, July 15, 2024. YouTube Video.

This post was researched and written with the assistance of various AI-based tools.

AI Talks