Silicon Valley's quest to automate everything is relentless, which explains its latest obsession: Auto-GPT.
In essence, Auto-GPT uses the versatility of the latest OpenAI AI models to interact with software and online services, allowing it to “autonomously” perform tasks like X and Y. But as we are learning with the big models Linguistically, this capacity appears to be as wide as an ocean but as deep as a puddle.
Auto-GPT - which you may have seen blowing up on social media recently - is an open source application created by game developer Toran Bruce Richards that uses the OpenAI text generation models, mainly GPT-3.5 and GPT-4. , to act “autonomously”.
There is no magic in that autonomy. Auto-GPT is limited to managing the follow-up of an initial query of the OpenAI models, both asking and answering until completing a task.
Auto-GPT, basically, is GPT-3.5 and GPT-4 paired with a companion bot that gives GPT-3.5 and GPT-4 instructions on what to do. A user tells Auto-GPT what their goal is, and the bot, in turn, uses GPT-3.5 and GPT-4 and various programs to carry out all the steps necessary to reach the goal it has set.
What makes Auto-GPT reasonably capable is its ability to interface with both online and local applications, software, and services, such as web browsers and word processors. For example, given a question such as "help me grow my flower business," Auto-GPT can develop a somewhat plausible advertising strategy and create a basic website.
As Joe Koen, a software developer who has experimented with Auto-GPT, explained to TechCrunch via email, Auto-GPT essentially automates multi-step projects that would have required a roundtrip with a chatbot-oriented AI model like, for example, OpenAI's ChatGPT.
"Auto-GPT defines an agent that communicates with the OpenAI API," explains Koen. “The objective of this agent is to carry out a series of commands that the AI generates in response to the agent's requests. Before the agent starts executing the orders, the user is asked to specify the role and objectives of the AI.
In a terminal, users describe the name, role, and goal of the Auto-GPT agent, and specify up to five ways to achieve that goal. For example:
- Name: Smartphone-GPT
- Function: An AI designed to find the best smartphone
- Objective: Find the best smartphones on the market
- Objective 1: Carry out a market study of the different smartphones that currently exist on the market.
- Goal 2: Get the top five smartphones and list their pros and cons
Behind the scenes, Auto-GPT relies on features like memory management to execute tasks, along with GPT-4 and GPT-3.5 for text generation, file storage, and summarization.
Auto-GPT can also connect to speech synthesizers, such as the one in ElevenLabs, to “make” phone calls, for example.
Auto-GPT is publicly available on GitHub, but requires some setup and technical knowledge to get up and running. To use it, Auto-GPT must be installed in a development environment such as Docker, and must be registered with an OpenAI API key, which requires a paid OpenAI account.
It might be worth it, though the jury is still out. Early adopters have used Auto-GPT to perform mundane tasks that would be better delegated to a robot. For example, Auto-GPT can handle things like debugging code and writing an email, or more advanced things like creating a business plan for a new startup.
"If Auto-GPT encounters any obstacles or an inability to complete the task, it will develop new prompts to help you navigate the situation and determine the appropriate next steps," Adnan Masood, chief architect at UST, a technology consultancy, told TechCrunch in a blog post. email. “Great linguistic models excel at generating human-like responses, but rely on user prompts and interactions to get the desired results. Instead, Auto-GPT takes advantage of the advanced capabilities of the OpenAI API to work independently without user intervention."
In recent weeks, new applications have appeared that make Auto-GPT even easier to use, such as AgentGPT and GodMode, which offer a simple interface where users can enter what they want to achieve directly on a browser page. Note that, like Agent-GPT, both require an OpenAI API key to unlock their full capabilities.
However, like any powerful tool, Auto-GPT has its limitations and risks.
Depending on what the tool is intended for, Auto-GPT can behave in very... unexpected ways. A Reddit user claims that, with a budget of $100 to spend on a server instance, Auto-GPT created a wiki page about cats, exploited a bug in the instance to gain admin-level access, and hijacked the Python environment. in which he was executed, and then "committed suicide".
There's also ChaosGPT, a modified version of Auto-GPT with goals like "destroy humanity" and "establish global dominance." Unsurprisingly, ChaosGPT hasn't come close to causing the robotic apocalypse, but it has tweeted unflattering things about humanity.
Arguably more dangerous than Auto-GPT's attempt to "destroy humanity" are the unforeseen problems that can arise in perfectly normal scenarios. Because it is based on OpenAI's linguistic models - models which, like all linguistic models, are prone to inaccuracies - it can make mistakes.
It's not the only problem. After successfully completing a task, Auto-GPT often doesn't remember how to do it for later use, and even when it does, it often doesn't remember how to use the program. Auto-GPT also has a hard time breaking complex tasks down into simpler sub-tasks effectively, and has a hard time understanding how different goals overlap.
“Auto-GPT illustrates the power and unknown risks of generative AI,” Clara Shih, CEO of Salesforce Service Cloud and Auto-GPT enthusiast, says via email. "For enterprises, it's especially important to bring a human approach into the loop when developing and using generative AI technologies like Auto-GPT."
