The field of generative AI is undergoing a significant transformation, moving away from monolithic models to more sophisticated, compound AI systems. Traditional models are limited by the data they’ve been trained on, impacting their knowledge and ability to adapt to new tasks. These models can perform tasks like summarizing documents or creating drafts of emails. However, their capabilities are enhanced when integrated into larger systems that can access external data and tools.
Compound AI Systems: Combining Models with System Design
Compound AI systems are built by combining models with other components, such as databases and external tools. They solve problems by applying principles of system design.
Key features of compound AI systems include:
– Modularity: Systems are made up of multiple components, including various models and programmatic components.
– Adaptability: Compound systems are easier to adapt and faster to solve problems than tuning models.
– Programmatic Control: The path to answering a query is often defined by a human using programmatic control logic.
Retrieval Augmented Generation (RAG) is an example of a popular and commonly used compound AI system. However, these systems have limitations. For example, a system designed to access vacation data would not be able to answer a question about the weather.
The Emergence of AI Agents: Putting LLMs in Charge
AI agents take compound AI systems a step further by placing a large language model (LLM) in charge of the control logic. This is possible due to the advancements in the reasoning capabilities of LLMs. Instead of following a pre-defined path, an agent can analyze a complex problem, develop a plan, and execute it.
Key capabilities of LLM agents:
– Reasoning: LLMs analyze problems and create plans to solve them.
– Action: Agents use external programs, or “tools”, to carry out tasks. These tools can include search engines, databases, calculators, and other language models.
– Memory: Agents can store and retrieve inner logs and conversation histories, enhancing personalization.
ReACT: A Popular Approach to Configuring Agents
A common way to configure AI agents is through the ReACT framework, which combines reasoning and action. In a ReACT agent:
– A user query is fed into an LLM.
– The LLM reasons and plans how to address the query.
– The agent then acts by using external tools, as needed.
– The agent observes the results, and iterates on the plan until it reaches the correct answer.
The Spectrum of AI Autonomy
There’s a sliding scale of AI autonomy when it comes to designing AI systems. For narrow, well-defined problems, a programmatic approach may be more efficient. However, for complex tasks and a wide variety of queries, an agentic approach is more beneficial, due to the fact that it would be too much work to configure every single path in the system.
The Future of AI Systems
The trend in AI is towards more agentic systems. These systems combine system design principles with agentic behavior. While still in the early stages of development, AI agents show promising progress, with humans remaining in the loop to ensure accuracy.
In summary, AI is advancing from monolithic models to modular, adaptable, and eventually autonomous agents. The coming years will see more AI systems incorporate agentic capabilities, enhancing their ability to solve complex problems and tackle a broad range of queries.
