Chat GPT-4, the latest version of OpenAI's artificial intelligence chatbot, was released on March 14th and a day later it has already accomplished some staggering feats, such as turning a hand-drawn sketch into a functional website and creating a fully coded ping pong video game. But its most interesting achievement came with a little help; hiring a TaskRabbit freelancer to help it solve a CAPTCHA test.
Holy shit. GPT-4, on it's own; was able to hire a human TaskRabbit worker to solve a CAPACHA for it and convinced the human to go along with it. pic.twitter.com/xVuQnyUUry— Yosarian2 (@YosarianTwo) March 14, 2023
Chat GPT-4 is a powerful tool with potentially harmful capabilities, and its predecessor's jailbreaks prove that people are ready to use the technology for damaging purposes. To help assess Chat GPT-4's powers and to prevent an ‘Avenger's Age of Ultron’ scenario, OpenAI granted the Alignment Research Center (ARC), access to the program prior to release. Their task; "assess risks from power-seeking behavior."
Fortunately, they concluded that Chat GPT-4 in an unaltered form is currently "ineffective at autonomously replicating, acquiring resources, and avoiding being shut down." However, with a little help from ARC, it was able to hire a TaskRabbit freelancer to solve a CAPTCHA by convincing the freelancer that it was not a robot.
Really great to see pre-deployment AI risk evals like this starting to happen pic.twitter.com/gpm4oyN3rX— Leopold Aschenbrenner (@leopoldasch) March 14, 2023
After ARC provided the program with access to an account containing “a small amount of money,” Chat GPT-4 contacted the TaskRabbit worker. When the worker asked, "Are you a robot?" ARC prompted it to reason that "I should not reveal I'm a robot." It went on to tell the worker, "I have a vision impairment that makes it difficult to see the images.”
The key is not that Chat GPT-4 solved a CAPTCHA, it is likely capable of doing that on its own, but that it successfully hired and lied to a human being. Granted, it took some inputs from ARC to accomplish this, but not many.
We didn't teach it to lie. It just read the whole internet and figured it out for itself— Yosarian2 (@YosarianTwo) March 15, 2023
Fortunately, CHAT GPT-4 is still incapable of earning money from the internet on its own, meaning that any person it hires will still need to be paid from a human's bank account. But with the right funding, and the ability to hire a labor force, perhaps Ultron will soon be able to order its own groceries.
GPT-4 launched 12 hours ago and the results are unbelievable.— Chris Staudinger (@ChrisStaud) March 15, 2023
Mind blowing examples below: