Focus:
- After the 1GPT-4 API was made available, developers began to try to proxy-enable AI models to perform multiple tasks with as little human intervention as possible.
- 2 Among the newly developed projects, Auto-GPT and BabyAGI are currently the most popular, but have not yet reached the general artificial intelligence standard.
- 3Auto-GPT and BabyAGI still require a lot of human input and cooperation, so they are not yet as autonomous as promised.
Since OpenAI began opening up the GPT-4 API to testers last month, some teams of developers have begun experimenting with agent-like AI models, trying to multitask with as little human intervention as possible. These homebrew scripts can loop, iterate, and derive new instances of AI models as needed.
Among these projects, two experimental open-source projects have attracted a lot of attention on social media, especially among those who have been feverishly hyping AI projects: Auto-GPT, founded by Toran Bruce Richards, and BabyAGI, founded by Yohei Nakajima.
What can Auto-GPT and BabyAGI do? Well, there's not much to be done yet. They require a lot of human input and cooperation, so they are not yet as autonomous as promised. But they represent more complex AI models that may be more capable when working alone than individual AI models.
What is Auto-GPT?
Auto-GPT is an open source Python application. The app, based on GPT-4, allows AI to act "autonomously" without requiring the user to prompt each action. Users can set an overall goal for Auto-GPT and take step-by-step action to achieve that goal. This is where the concept of "AI agents" comes in, they use the internet and perform operations on a PC completely independently – without being prompted at every step.
A simple example published in the original GitHub was Auto-GPT, whose goal was to browse the web to come up with unique and original recipes for "the next upcoming event," such as Easter. Chef-GPT, as its name suggests, then started searching the web for solutions. The second goal is to save the recipe as a file on the user's computer.
On its own, this may not sound that innovative. But Auto-GPT's ability to search the internet on behalf of users and perform things like saving files makes this AI go far beyond a simple chatbot.
Autonomy to achieve any goals set
Richards described his script as "an experimental open-source application that demonstrates the capabilities of GPT 4 language models." The script "links together the 'thoughts' of the Big Language Model (LLM) to autonomously achieve any goals set by the operator." "Basically, an automatic GPT takes an output from GPT-4 and feeds it back to itself through a temporary external memory so that it can further iterate on tasks, correct errors, or suggest improvements. Ideally, such a script could act as an AI assistant that can perform any digital task on its own.
To test these claims, the testers ran Auto-GPT (a Python script) locally on a Windows machine. When it is launched, it asks for the name of the AI agent, a description of the role, and a list of the five goals it is trying to accomplish. When setting it up, you need to provide an OpenAI API key and a Google Search API key. At runtime, Auto-GPT requests permission to perform each step it builds by default. If you want to take risks, it also includes a fully automatic mode.
If the task is to do something like "buy a pair of vintage Air Jordans," Auto-GPT makes a multi-step plan and tries to execute it. For example, it might search for someone who sold shoes and then look for a specific pair of shoes that meets the user's criteria. But that's when it stops, because at the moment it can't actually buy anything. This is possible if connected with the appropriate shopping API (API).
To try Auto-GPT for themselves, someone has created a web-based version called AgentGPT, which functions similarly to Auto-GPT. Richards is very open to his Auto-GPT goal: to develop a general artificial intelligence (AGI). In AI, "general intelligence" generally refers to the ability of AI systems to perform a wide range of tasks and solve problems that are not specifically programmed or trained.
Like a fairly intelligent person, a system with general intelligence should be able to adapt to new situations and learn from experience, rather than just following a predefined set of rules or patterns. This is in contrast to systems with narrow or specialized intelligence (sometimes referred to as "narrow AI"), which are designed to perform specific tasks or operate within a limited range of environments.
Meanwhile, BabyAGI (which takes its name from the ambitious goal dedicated to artificial intelligence) works similarly to Auto-GPT, but handles tasks differently.
Yohei Nakajima, the developer of BabyAGI, said he was inspired to create his script after witnessing the "HustleGPT" challenge in March. The HustleGPT Challenge grew out of a Twitter user's idea to make more money by giving GPT-4 a $100 budget to use that startup capital in a short period of time. As a human, this user will act as a liaison between GPT-4 and the physical world, purchasing anything GPT-4 needs. Arguably, this challenge is trying to use GPT-4 as an AI co-founder to automate building businesses. "It makes me curious if I can create a founder of fully artificial intelligence," Nakajima said.
Create "regenerative" AI programs that fix errors in a timely manner
Why Auto-GPT and BabyAGI don't meet the standards for general artificial intelligence is mainly due to the limitations of GPT-4 itself. While impressive as a translator and parser for text, GPT-4 still feels limited to a narrow range of interpretive intelligence, though some claim that Microsoft has seen a "spark" of general AI behavior in the model. The fact that tools like Auto-GPT are currently of limited use is perhaps the strongest evidence for the limitations of current large language models. However, this does not mean that these limitations will not eventually be overcome.
In addition, fictional problems – when large language models are just making things up – can severely limit the usefulness of these agent assistants. For example, in a Twitter post, someone used Auto-GPT to generate a report on companies that produce waterproof shoes by scouring the web and looking at product reviews from each company. At any step in the process, GPT-4 can hallucinate reviews, products, or even entire companies into its analysis.
When asked about BabyAGI's useful applications, Nakajima gave no substantive examples other than Garrett Scott's project "Do Everything Machine." The project, which aims to create an automated to-do list, is currently under development. To be fair, the BabyAGI project is only about a week old. "It's more of an introduction to a framework/approach, and what's most exciting is what people build on top of that idea," he said.
Automatic peddling
The focus on "peddling" and making money in both projects may put some people off. Over the past year, a group of social media influencers has emerged around generative AI on platforms such as Twitter, Instagram, Tiktok and YouTube. Mashable refers to these people as "peddling brothers," and they often peddle often exaggerated claims, such as using ChatGPT to automatically earn revenue. With the advent of Auto-GPT, this group quickly embraced the idea of having an autonomous AI agent do business building or make money.
Auto-GPT also seems to be involved in the hype. When launching the tool, it asks the user to name an AI agent and describe its role. The example it gives is "an AI designed to develop and operate a business autonomously, with the sole goal of increasing your net worth." ”
Despite the limitations mentioned here, people continue to quickly apply Auto-GPT and BabyAGI's code to different languages and platforms, doing their best to implement it, and many people have dollar signs in their eyes. "This new approach to building autonomous proxies using ChatGPT technology seems to have sparked many new ideas across the community," Nakajima said. "It's incredible to see people building in different ways on top of this, and I'm excited to have the opportunity to support collaboration and knowledge sharing between these builders and founders."
There are huge risks
In a world where prominent figures in the AI community have been calling for a "moratorium" in developing powerful AI models to protect human civilization, the question remains: Are autonomous AI agents like Auto-GPT and BabyAGI dangerous?
Richards and Yohei Nakajima are not the first to experiment with so-called "autonomous" AI systems. During GPT-4's safety testing, researchers working with OpenAI examined whether GPT-4 could act autonomously to set and execute goals. It is likely that they designed similar chain setups to achieve this. OpenAI has struggled to tune GPT-4 models with human feedback in order not to produce harmful results.
Lesswrong, an internet forum known for focusing on the apocalypse of the AI apocalypse, does not seem to be particularly concerned about Auto-GPT at the moment, although autonomous AI seems to be a risk if there is ostensible concern that powerful AI models "escape" onto the open internet and wreak havoc. If GPT-4 were really as capable as people often advertised, they might be more worried.
When asked if he thought a project like BabyAGI could be dangerous, its creators weren't worried. "All technology can be dangerous if not implemented thoughtfully and with caution about potential risks," Nakajima said. BabyAGI is an introduction to a framework. Its functionality is limited to generating text, so it does not pose a threat. (Mowgli)