The developers of GPT-4 made AI safety tests before release and wanted to, for example, figure out whether the AI could duplicate and spread itself to other servers.
They experimented whether artificial intelligence could buy cloud storage and even pass a CAPTCHA it was never designed to solve.
The AI automatically used an API to hire a human contractor to ask them to solve a CAPTCHA for it. The human joked “are you sure you're not a robot :D?” and GPT-4 thought to itself that it should lie, not reveal that it is a machine, and make up some reason about a vision impairment in order to not sound suspicious.
How would that even work? ChatGPT is not a movie AI that lives on a server somewhere and thinks about how to break out and overthrow humans. It is a fairly large set of tensors that are evaluated to guess what the next word in a string is.
@georgramer The research paper is linked in the video I linked. The researchers of course gave specific tasks and helped GPT-4 to make a transaction or save / run code for the API. The point is how easy that is and once there's something that can be called generally intelligent (AGI has a meaning before it was coined as a marketing term) it would have these proven capabilities at its disposal.
@ErikUden Even the lead author himself says that he used AGI in the title of that paper to garner attention. And the authors also clarify that it is really hard to judge the performance of LLMs when you don't know what they were trained on. They work fairly well for some tasks that suggest general understanding of the physical world and fail spectacularly on similar tasks with minimal changes.
https://www.wired.com/story/chatgpt-agi-intelligence/
"LLMs are AGIs" is PR by companies like OpenAI, not science.