Write a book, build a website, or develop a business model – AIs like ChatGPT are already capable of this type of creation. But what if they could create all of this on their own? This is the promise of Auto-GPT, a new type of artificial intelligence capable of carrying out large-scale projects without human intervention… or almost.
Let the AI create its own tasks
AIs like GPT-4 are capable of creating many things, but from the initial idea to the finished product, the process can be long and require a lot of human input. For example, if you want to create a website or a video game, it is not enough to give ChatGPT great ideas, as Tech&Co had seen: you have to give it the beginning, copy the code yourself, download or create the graphic elements. yourself…
You need to go back and forth, flag errors to the software for correction, and remind it of important project milestones, as ChatGPT’s limited “memory” is not well suited for large projects. .
This is where Auto-GPT comes in. This model, available for free on GitHub, falls under the category of “agents”: all you have to do is describe their role and give it some general goals, and the AI itself will create a list of intermediate tasks to complete in order. to achieve your goal. A type of model that has been used in particular by researchers to create a virtual village à la Pokémon fully populated by AI.
No need for human intervention
An example? Its creator, Toran Bruce Richards, published on his Twitter account a demo video asking Auto-GPT to search the internet for upcoming events and create a suitable cooking recipe. Auto-GPT then launches Internet searches, studies the content and assesses its relevance, before finding the event and a dedicated recipe, all without human intervention.
To do this, Auto-GPT can search the Internet (like the chatbot built into Bing), and connects several AIs: GPT-4 to generate text and code, as well as GPT-3.5 (the version used in ChatGPT) to store and summarize the accumulated information.
Another netizen explains that he used it to make fake market research. From this simple goal, Auto-GPT generated a list of questions it needed to answer, searched the Internet for related information, suggested checking the credibility of reviews left by buyers, created its own “sub-agent to analyze websites, and generated a summary of this information.
Results still limited
But Auto-GPT is still far from being an autonomous AI in the I robot. Right, because you don’t really “think” about the action to take. Therefore, Auto-GPT can get stuck in a loop of infinite actions without “thinking” to interrupt the process to get out of it, warns its creator on GitHub. Auto-GPT can also perform tasks that would seem useless to a human being, because the program has judged them to be the most likely.
And AutoGPT’s final creations are still not professional or human grade, as pointed out in some comments on the fake market research. One of the OpenAI co-founders recalls On twitter that models are not ‘aware’ of their own capabilities and limitations, so they can get into things they can’t, like low-complexity math.
Therefore, the creator of Auto-GPT believes that human supervision is always recommended. “Ensuring that these models operate within an ethical and legal framework, while respecting privacy and security concerns, should be a priority,” Toran also tells Bruce Richards in Vice.
Source: BFM TV
