Wen Chen Gen.
With the application of large AI models in all walks of life and the continuous iteration of AI technology, there are more and more discussions about whether AI threatens humanity.
In fact, there have been many such discussions in the past, and even since the birth of AI technology, some people have worried about whether AI will one day replace humans or threaten the existence of human beings.
It's just that today, the explosion of AI large models has made this issue suddenly change from an abstract discussion to a very concrete one. We have to think about how we can prepare for the coming era of AIWe must face the question of what we humans should do if the performance of AI reaches or even surpasses the human level, and whether AI will really have consciousness one day in the future, and how to resolve the conflict between man and machine at that time.
In the face of these problems, there is a concept that has also been recalled, and that is "value alignment". It's not a new concept, but it seems to be particularly relevant today. So, what exactly is value alignment?Who is aligned with whose value?And how to align?
To put it simply, value alignment is actually to align the values of AI with the values of our human beings, and the core purpose of aligning the values of AI with the values of our humans is for the sake of security.
You can imagine what happens if it's not aligned.
For example, Nick Bostrom, a philosopher and director of the Institute for the Future of Humanity at the University of Oxford, once put forward a classic case. That is to say, if there is a powerful super-intelligent machine, we humans have given it a task, which is to "make as many paperclips as possible".Yes, thisA super-intelligent machine with great capabilitiesUnscrupulous means to make paper clips,Turn all the people and things on the earth into materials for making paper clipsEventuallyDestroynowThe whole world
This story actually happened as early as ancient Greek mythology. It is said that a king named Midas saved the god of wine by chance, so the god of wine promised to fulfill one of his wishes, and Midas liked it very much, so he made a wish, hoping that he could turn stones into gold. As a result, Midas really got what he wanted, and everything he touched immediately turned to gold, but soon he found out that it was a disaster, and the water he drank became **, and the food he ate became **.
Have you found a problem from these two stories, whether it is a super-intelligent machine or Midas, they are all for their own purposes, and in the end, the super-intelligent machine completed the paperclip task, and Midas also made a little stone into gold, but the result was very disastrous. Because in the process, they lack certain principles.
That's why today the concept of value alignment is being re-emphasized. AI simply does not have the same concept of value about life as humans. In this case, the greater the AI's capabilities, the greater the potential for threat and the more damaging it will be.
Because if we don't "align" AI with our human values, we may inadvertently give AI goals that are completely opposite to our own. For example, in order to find a cure for cancer as soon as possible, an AI may choose to experiment on an entire human being as a guinea pig. To address ocean acidification, it may deplete all oxygen in the atmosphere. ThisActually, yesis a common feature of system optimization: variables that are not included in a target can be set to extreme values to help optimize that target
In November, a maintenance worker at a robotics company in South Korea's Gyeongsang Province was crushed to death by a vegetable sorting robot because the robot picked him up and squeezed him for a box of vegetables that needed to be processed, causing serious injuries to his face and chest. He was then taken to the hospital, where he died of his injuries.
In addition, a large AI model with no value alignment may also output content that contains racial or gender discrimination, help cyber hackers generate ** or other content used to carry out cyber attacks, telecom fraud, try to persuade or help users with suicidal thoughts end their lives, and so on.
The good news is that different AI teams are taking different approaches to drive value alignment in AI. OpenAI and Google's DeepMind each have teams focused on solving value alignment problems. In addition to this, there are many third-party watchdogs, standards organizations, and ** organizations that also see value alignment as an important goal. It can be said that if there is no value alignment, we will not really trust AI, and the AI era of human-machine collaboration will not be possible.