Artificial intelligence (AI) has the potential to surpass human intelligence someday, a notion that some people may consider science fiction. However, some researchers believe that there are already early signs of this possibility in GPT-4.
Can an artificial brain truly comprehend? To some researchers, this no longer seems absurd.
AI could become very dangerous for us humans. There is a surprising consensus on this matter. The hype surrounding AI and chatbots has recently turned into a frenzy of warnings.
 |
| How Close Are We to Superintelligence? |
However, upon closer inspection, one can see that there is a fierce debate about what the dangers are that come from machine learning systems. Some experts warn of the risks in the here and now. Chatbots could be abused, for example, to spread false information en masse. Or they could reproduce racist and sexist distortions in their training data. Others are looking further into the future. They warn about what could happen if AI continues to be developed at its current pace. Then it could, in their view, become just as intelligent as humans or even more intelligent. Researchers call this artificial general intelligence (AGI), while some refer to it as strong AI or superintelligence.
Such intelligence could "someday surpass us, make us redundant, and replace us." At least that's what a group of AI scientists and tech leaders wrote in an open letter last week. They therefore called for a pause in the development of such systems and instead demanded a six-month break to develop security measures to ensure that any potential superintelligence helps humanity solve its problems, rather than getting out of control and harming us.
Whether one finds this sensible or alarmist depends largely on whether one assumes that strong AI is really just around the corner. Researchers from Microsoft Research now claim to have found evidence for this. In a publication (PDF), they report that they have found "sparks of General Artificial Intelligence" in a comprehensive study of the language model GPT-4. The system shows "remarkable abilities" in a variety of areas, including abstraction, programming, mathematics, medicine, law, understanding human motives and emotions, and much more. In many of these areas, it already surpasses humans today.
This directly contradicts what other researchers have been saying for a long time and repeatedly emphasize. They believe that it is impossible for general intelligence to emerge solely from statistical models. The fear of superintelligence is "nonsense."
While it is uncertain whether or not we are truly on the brink of developing superintelligence, the debate surrounding this possibility is an important one. As AI continues to advance, we must remain vigilant and ensure that it is developed with the interests and safety of humanity in mind.