Imagine you are at work, any day. You are assigned a task, such as writing a market research on a specific product, which perhaps your company intends to develop. It is a process that goes on gradually: first you get an idea of the topic, then perhaps you look for examples of similar documents, you start writing a few paragraphs, if necessary you ask for help from some colleagues.
In other words, when a human being is given a task, the answer goes through a series of steps. And it’s exactly this kind of process that a new wave of AI-powered applications is trying to replicate. These are open source systems, currently available almost exclusively in code form on GitHub, based on the GPT-4 model: the best known is called AutoGPT. They are software that allow you to create autonomous agents based on AI, able, given a wide-ranging objective, to arrive at a solution through a series of steps, which can also include actions such as web search or code execution.
“Today ChatGPT allows you to perform one task at a time, always starting from a stimulus from a human being (the so-called prompt, ed) – he explained to us Pietro Schirano, head of artificial intelligence at the US fintech Brex – Systems such as AutoGPT are instead capable of reaching the desired result for intermediate steps, autonomously, among other things by exceeding the token limits of GPT-4 (about 24 thousand words in the developer version, ed). If we ask him to write a book, he’ll proceed to character creation first, then a plot outline, and only it will eventually come down to the writing job real. There are many possible uses: someone used it to do market research, others to have an assistant in scientific research”.
The process explains it well, with an analogy, Andrej Karpathy, one of the founding members of OpenAIin a thread on Twitter: “Every generation of GPT can be compared to a thought – he wrote – If these are connected and linked, it is possible to create agents capable of acting autonomously on the basis of a defined objective”.
From a technical point of view, AutoGPT is actually software based on GPT versions 4 and 3.5, which is used to interpret user goals, establish a course of action, and create intermediate instructions. It has a medium-long term memory, which allows it to remember the passages of reasoning, and can access the Web, as well as files and everything the user makes available. A bit like with the ChatGPT pluginswith the substantial difference that in this case it is the artificial intelligence itself that autonomously decides which systems to interrogate.
There are already many versions of it, currently available for developers: besides AutoGPT, there are BabyAGI and Jarvis, by Microsoft. There is also a first variant open on the Web, for an audience of non-experts: it is called AgentGPT and can be accessed from a browsereven if at the moment the most advanced functions are reserved for those who have a subscription to the OpenAI API.
The system is still in an embryonic stage and, net of the hype that brought it to first place among the trends on Twitter in the US, it is probably still far from representing a credible automation resource for the general public: “It is a resource still in its infancy – is Schirano’s conclusion – Sometimes it gets stuck, doesn’t finish the task it has been assigned or gets lost in too complex passages. That said, it’s improving at impressive speed.”
Artificial intelligence
The AI between racism and discrimination: two years later, we are worse off than before
by Emanuele Capone
A city populated by artificial intelligences
In short, it is still too early to imagine a future of AI-based autonomous agents at the service of all Web users. However, this could be a long-term direction. It proves it an experiment conducted by scientists from Google and Stanford and published in pre-print (it has therefore not yet had a review by the scientific community) in recent days.
It’s about the simulation of a virtual city, called Smallville and inhabited by 25 artificial intelligences, programmed to interact with each other based on a personal story, a ChatGPT prompt. For example, in the case of John Lin, one of the characters: “he IS a pharmacist at Willow Market and Pharmacy who loves helping people. He lives with his wife, Mei Lin, who is a university professor, and his son, Eddy Lin, who is a music student”.
Starting with this mini-plot, the ChatGPT-based agents were asked to interact, in relation to the time of day and the information provided in real time on the surrounding environment. There is online a video of this simulation, which has transformed what are actually only textual interactions into images. Well, the experiment tells us that, within the virtual city, social interactions and particularly suggestive developments have arisen: Sam, one of the characters, has decided to run for mayorwhile Klaus decided to start a study on the gentrification of low-income areas.
In particular, the researchers tell, the agents have shared information with each other, formed relationships and were able to coordinate to organize events, such as a Valentine’s Day party. They simulated a community of people, just as the study intended. Who sees a future for this kind of agents “in forums, in the metaverse or, even, in physical space, thanks to multimodal robots”. Obviously always keeping in mind a series of ethical issues and risks, such as that humans can form parasocial relationships with AIs, or that the latter, placed in a different context from the closed one of the experiment, they can deviate from the predefined objectives and generate disinformation or represent a danger for the human users who interact with them.