OpenAI's internal push, can AI agents really become the next stage of Internet application development?

Image source: Generated by Unbounded AI

The development history of Internet applications can be seen as a process of continuous evolution and innovation. With the continuous advancement of technology, innovative Internet applications continue to appear.

The development of Internet applications can be divided into three stages:

1. Websites

In the 1990s, with the popularization of the Internet, some important Internet companies emerged, such as Amazon, Yahoo, Google, etc. These companies developed some important Internet applications, such as e-commerce, search engines, online advertising, etc.

2. Mobile application (Mobile)

In the 2000s, with the rise of the mobile Internet, some important mobile applications appeared, such as smart phones, mobile application stores, etc.

3. AI agent application

In the 2020s, with the development of AI technology, some important artificial intelligence applications have emerged, such as speech recognition, image recognition, and natural language processing. Especially after OpenAI launched ChatGPT, the autonomous AI agent application driven by the large language model (LLM) will bring the AI agent application to a new stage of development.

AI agent development map

Image source: e2b

What is an AI agent

AI agent (AI agent) refers to a computer program designed and programmed using AI technology, which can independently perform certain tasks and respond to the environment. An AI agent can be viewed as an agent that perceives its environment, changes it through its own decisions and actions, and improves its performance by learning and adapting. Using both short-term memory (contextual learning) and long-term memory (retrieval of information from external vector stores), the agent has the ability to plan by "thinking" step by step, break down goals into smaller tasks, and reflect on its own performance.

AI agents usually incorporate multiple technologies, such as machine learning, natural language processing, computer vision, planning, and reasoning, that enable agents to process information and make decisions autonomously.

OpenAI has repeatedly expressed its enthusiasm for AI agents. OpenAI co-founder Andrej Karpathy recently said in an offline event for developers that if a paper proposes a different model training method, OpenAI will scoff at it internally, thinking that it is all left over from their play. But when the new AI Agents paper comes out, they will discuss it seriously and excitedly.

What is an autonomous agent supported by LLM

Lilian Weng, director of AI application research at OpenAI, recently published a 10,000-word long article on AI agents: "Autonomous Agents Supported by Large Language Models (LLM)", which provides an in-depth interpretation of what is an AI agent application built by LLM training. There are many excellent applications of AI agents supported by LLM, such as AutoGPT, GPT-Engineer, BabyAGI, and SuperAGI.

In an LLM-powered autonomous agent system, the LLM acts as the brain of the agent and is complemented by several key components: Planning, Memory, and Tool Use.

This agent breaks down large tasks into smaller, manageable sub-goals, enabling efficient handling of complex tasks. It also allows for self-criticism and self-reflection about past actions, learning from mistakes and refining for future steps, thereby improving the quality of the end result.

A special feature of the LLM autonomous agent is that it is like having a "memory", which is capable of short-term (long-term) remembering what it has learned during training. In addition, LLM autonomously brings the ability to learn to call external APIs to obtain additional information missing in model weights (usually difficult to change after pre-training), including current information, code execution capabilities, access to proprietary information sources, etc.

As exciting and promising as AI agents are, there are still many challenges surrounding the hype around AI agents. AI agents are becoming the future of software applications and will become more and more common.

As mentioned by Lilian Weng, there are also some common limitations of LLM autonomous agents, including limited context length, challenges of long-term planning and task decomposition, stability of LLM, etc.

But there is no doubt that these problems and challenges will be overcome or alleviated. AI agents have brought changes to our work and life, and this change is difficult to reverse. After trying something good, do you put up with something really bad?

References: