In 2023, ChatGPT initiated a technological revolution. Interactive AI agents quickly evolved to index documents, connect to data sources, and even perform data analysis with simple sentences. Despite promises from many to deliver large language models (LLMs) soon, few have delivered due to several challenges:
1. We’re building AI agents, not LLMs.
2. The problem is being treated as research, not engineering.
3. There’s too much bad data.
This article explores the role of AI agents in linking LLMs with backend systems, setting them up as the next generation of user interface and user experience (UI/UX). It also highlights the need to revive software engineering principles that have been overlooked.
Intuitive User Interfaces with LLMs
LLMs provide a more intuitive and streamlined UI/UX than traditional point-and-click interfaces. For example, ordering a “gourmet margherita pizza delivered in 20 minutes” through a delivery app with normal UI/UX could involve multiple complex interactions:
– Selecting the “Pizza” category
– Browsing listings and photos
– Checking menus for margherita pizza
– Confirming quick delivery
– Backtracking if criteria aren’t met
Beyond LLMs
LLMs like GPT-3 excel in natural language processing (NLP) and generating relevant responses. By connecting LLMs to external data sources, algorithms, and specialized interfaces, their flexibility and analytical capabilities increase, enabling tasks not yet possible. However, LLMs alone can’t handle all necessary connections, such as restaurant databases, inventory management, and delivery tracking, to complete a seamless order.
AI Agents
AI agents, built on LLMs, respond to a wide range of queries by leveraging several key components:
– The agent core uses the LLM and manages overall functionality.
– The memory module allows for context-aware decisions.
– The planner determines the agent’s actions based on available tools.
These components, including data sources, algorithms, and visualizations, enable AI agents to process data, reason, and generate appropriate responses effectively.
Engineering Over Research
Adding LLM-based AI agents to your data is an engineering challenge, not a research problem. Natural language helps specify use cases for software development, but the ambiguity of English can complicate system specifications. Fred Brooks’ principles from “The Mythical Man-Month” remind us that:
– No silver bullet can replace proper software engineering practices, not even LLMs.
– Formal documentation is essential, requiring detailed specifications for use cases, backend systems, visualizations, and system limitations.
Addressing Bad Data
Effective LLM-based AI agents require well-organized data and high-quality documentation. Training these models necessitates vast amounts of high-quality text, often from copyrighted works. This need is even more critical for RAG-based technologies, which index document chunks using embedding technologies in vector databases and use top-ranking documents to generate responses.
Conclusions
Despite the promises of LLM-based systems, few have been realized. To build intelligent AI systems, we must acknowledge the complexity of these software engineering systems. Proper specification and testing are crucial, as is treating data as a first-class citizen, given the susceptibility of intelligent systems to bad data.