Could ChatGPT and other AI models threaten humanity? Future Publications by CFOTO/Getty Images
The rapid rise of new AI models in recent months, such as OpenAI’s ChatGPT, has caused some engineers and researchers to ponder whether artificial intelligence can soon surpass human capabilities. became. One of his key former OpenAI researchers says such a future is a definite possibility, but the possibility of human or superhuman-level AI taking over and even annihilating humanity is unlikely. It warns that it is not zero.
A “full-blown AI takeover scenario” is a top priority for Paul Christiano, former head of language model tuning for OpenAI’s safety team, warned last week in an interview with tech-focused companies. bankless A podcast that advanced AI could very well cause a potentially world-ending disaster in the near future.
“Overall, you could have a 50-50x chance of destiny as soon as you have a human-level AI system,” says Christiano. “I think there is a 10-20% chance of an AI takeover. [with] Many, most people are dead. ”
Christiano will be leaving OpenAI in 2021, explaining his resignation during LessWrong’s “Ask Me Anything” session. LessWrong is a community he blog site created by fellow AI researcher Eliezer Yudkowsky, who has long warned that superhuman AI could destroy humanity. Cristiano then worked collaboratively on more conceptual/theoretical issues, a subfield of AI safety research focused on ensuring that AI systems were consistent with human interests and ethical principles. I want to,” and OpenAI is “perfect for that type of research.
Cristiano currently runs the Alignment Research Center, a non-profit organization working on theoretical AI alignment strategies. This area has gained considerable interest in recent months as companies race to deploy increasingly sophisticated AI models. OpenAI released his GPT-4 in March. This is an update to the AI model that powers ChatGPT and was just released in November. Meanwhile, tech giants such as Google and Microsoft have launched an AI arms race to gain an edge in the burgeoning market, launching their own versions of AI models with commercial applications.
But AI systems in public view are still riddled with errors and misinformation, and Christiano and many other experts warn against moving too fast. OpenAI co-founder Elon Musk, who cut ties with the company in 2018, was one of 1,100 engineers who signed an open letter in March to develop an advanced technology that is more powerful than GPT-4. It called for pausing AI model development for six months and refocusing research on improving the reliability of existing systems. (Musk has since announced that he will launch a competition for his ChatGPT called TruthGPT, which he says focuses on “truth seeking” rather than profit.)
One of the letter’s concerns was that existing AI models could pave the way for superintelligent models that threaten civilization. Current generative AI systems like ChatGPT can handle certain tasks well, but they still fall short of human intelligence levels, a hypothetical future iteration of AI known as artificial intelligence (AGI). .
Experts are divided on the timeline for AGI’s development, with some arguing it could take decades and others saying it may never be possible. About 57% of AI and computer science researchers say their AI research is rapidly shifting to AGI, according to a Stanford University survey released in April. 36% said entrusting important decisions to advanced versions of AI could lead to a “nuclear catastrophe” for humanity.
Other experts warn that even stronger versions of AI developed neutrally could quickly become dangerous if used by malicious humans. Geoffrey Hinton, his former Google researcher who is often called the Godfather of AI, said: new york times this week. “I don’t think they should escalate this further until they understand if they can control it.”
In an interview, Cristiano concluded that if AI developed to the point where society could not function without it, civilization could be endangered, and powerful AI no longer needed to act in the interests of its creators. said that humanity could remain vulnerable if
“The most likely way we would die would be if we deployed a lot of AI everywhere instead of just suddenly showing up and killing everyone,” he said. , if all these AI systems were going to kill us, they would definitely kill us.”
However, there are voices opposing this interpretation of AI. Some experts believe that while AI designed to accomplish specific tasks is inevitable, developing an AGI that rivals human intelligence will be difficult due to the limitations of computers when it comes to interpreting life experience. argue that it may not be practically feasible.
Responding to recent dire warnings about AI, entrepreneur and computer scientist Perry Metzger claimed in a tweet Last month, while there could be a “very superhuman” AI, it would likely be years or decades before the AGI would be able to rebel against its creators, whose creators would I said it’s likely we’ll have time to steer you in the right direction. In response to Metzger’s tweet, Yann LeCun, his NYU computer scientist who has been leading AI research at Meta since 2013, said that AGI develops dangerous, out-of-control abilities overnight. The fatal scenario is written as follows.utterly impossible”
