Over the past week, developers around the world have begun creating “autonomous agents” that work with large language models (LLMs) such as OpenAI’s GPT-4 to solve complex problems . Although very new, these agents could represent a major advance in the productive application of LLMs.
Normally, we interact with GPT-4 by typing carefully formulated prompts in the ChatGPT text window until the model generates the desired output. However, most of us lack the skill and patience to write prompts, so developers are looking to automate much of this process.
This is where autonomous agents come in.
Autonomous agents like AutoGPT can generate a systematic sequence of tasks that the LLM performs until it reaches a predefined “goal.” They can already perform a variety of tasks such as searching the Web, writing code, and creating task lists. Some call them “recursive” agents because they operate in a loop, asking questions of the LLM, each based on the result of the previous one, until the model produces a complete answer.
What is Auto-GPT?
Auto-GPT is an open-source application that has some professionals worried that it could replace jobs such as social network managers, SEO experts and customer relations employees. GPT stands for Generative Pre-trained Transformer, a neural network designed to work like the human brain, able to learn and improve its performance independently.
Developed by Toran Bruce Richards, AutoGPT is described on GitHub as a GPT-4-powered agent that can search the Internet in a structured way. It can create subtasks and launch new agents to accomplish them. It uses GPT-4 to write its own code, and then can “debug, expand, and improve” recursively.
How Auto-GPT works
Auto-GPT is distinctive in the way it breaks down the AI steps, using GPT’s excellent text generation. Auto-GPT calls them “thoughts”, “reasoning” and “criticism”, indicating exactly what the AI is doing and why.
Other interesting features of Auto-GPT include long/short term memory and text-to-speech integration via ElevenLabs . The combination of all these features makes Auto-GPT much more like an AI designed to interact with humans.
Example with the godmode.space website which allows anyone to use AutoGPT/BabyAGI directly on the web.
AutoGPT vs. ChatGPT
Unlike ChatGPT, which performs tasks after receiving specific instructions and criticism from humans, AutoGPT can think of other relevant tasks to perform once it has completed a task.
Impact on the future of work
It’s too early to assess the impact of this development on the future of humanity and the way we work. Some big names in tech, such as Elon Musk and Jaan Tallinn (one of the brains behind Skype), advocate putting the brakes on the development of AI and thinking about its impact on our lives.
A letter signed by thousands of people, including DeepMind researchers, Musk, Steve Wozniak (co-founder of Apple) and Evan Sharp (co-founder of Pinterest), calls for a six-month pause in the AI race.
Call for caution
The letter cautions against developments beyond GPT-4 and suggests that AI labs and independent experts use this pause to develop and implement a set of shared safety protocols for advanced AI design and development. Jaan Tallinn also expresses concern that the technology is advancing rapidly without time for society to adapt and control it.
Reactions and regulations
While the Italian government recently banned ChatGPT, in the UK, MPs are reluctant to introduce strict regulation of AI for fear of stifling innovation. An anonymous MP told the BBC that tougher laws would probably not be needed for a few years.
Autonomous agents at this early stage are mostly experimental and have some important limitations that prevent them from getting what they want from LLMs. They often have trouble keeping the LLM focused on a goal. Sully Omar, a Vancouver-based developer, mentions that LLMs “get confused” and don’t understand that they are in a loop. He believes that developers will likely find new ways to put “guardrails” around the LLM so that they continue to accomplish tasks without distraction.