Liu Yongmou
He is a researcher at the National Institute of Development and Strategy, Renmin University of Chinese
Professor, Faculty of Philosophy
OpenAI's "palace fight" drama has made the word "alignment" of AI go far and wide. Many people believe that the spokesman of evil capital like Altman suppresses Sutskevi, the spokesman of justice science who insists on alignment, and we must stand on the side of justice. Is this really the case? By aligning operations, AI development can meet people's needs and not deviate from human intentions? Is it really so amazing?
1
Question the possibility of AI alignment
What is alignment? It emerged in the development of machine learning, especially large model technology. According to the book Human-Machine Alignment, "how to prevent this catastrophic divergence – how to ensure that these models capture our norms and values, understand what we mean or intent, and, most importantly, act the way we want – has become one of the most central and pressing issues in computer science." This problem is called the alignment problem." That is, alignment means having a machine learning model "capture" human norms or values.
"Capture" is the opposite of "indoctrination", where AI follows a specification that comes from machine learning, rather than programming input from engineers. By learning a lot about human behavior, the AI "figurs out" the rules of human behavior and then acts in accordance with them. Therefore, the alignment problem can be divided into at least two, i.e., what to align and how to align.
In the opinion of many, including the "limitedism of AI development" (emphasizing the limited and controlled nature of AI development), the question of "what to align" cannot be fully clarified.
First of all, there is no universal value for human beings. People of different genders and classes living in different countries and regions, with different traditions and cultures, have different value judgments about the same phenomenon. For example, in the face of the coronavirus pandemic, some people believe that preserving life is the most important thing, while others think that freedom of movement is more important. Whose rules of action does the big model have to learn?
Second, the prevailing values of humanity are constantly changing. For example, polygamy was popular in China more than 100 years ago, and now it is a crime of bigamy. So, what time period do we want to input data to the large model for learning?
Thirdly, there is a deviation between what should be and what is actually in the rules. For example, equality between men and women is a value promoted by society, but in reality there is no shortage of gender discrimination. If the AI learns from real cases, it is likely to become sexist. This kind of problem is called the representative problem of large models, and it is not uncommon in practice.
Finally, there are AIs like robot pet dogs, which should be aligned with pet dogs, not people. Otherwise, it becomes a dog-like person, and having it doesn't have the pleasure of having a pet. In other words, not all AI needs to be aligned with humans.
As a result, the question of "what to align" is "a human, social, and political problem that machine learning cannot solve on its own." The question of alignment is essentially a question of clarifying complex human rules and values with data methods or statistical methods.
2
AI alignment is of very limited use
Fundamentally, the question is whether big data or statistical techniques can solve problems that have not been fully solved by moral philosophy or ethics? Indeed, the answer is no. However, just as ethics has solved some of the values problem, big data technology is not useless for learning human rules. In everyday situations, it is not entirely clear that human values are fully articulated, and the agent "knows" how to act.
Most of the time, AI simply needs to respond to common situations in a given situation in a common way. In autonomous driving research, it is often analyzed with the example of the "tram problem". However, human drivers rarely face the need for such difficult decisions. Whether it's "indoctrination" or "learning", self-driving cars can solve the problem with random solutions or direct braking. It is important to take responsibility for the accident rather than dwell on how autonomous driving can solve the "tram problem".
At present, machine learning models mainly use imitation and inference to align AI. The former is to see what humans do, and AI will follow suit. There are many problems with imitation, such as excessive imitation, many people will roll up their sleeves before stir-frying, and AI may imitate this unnecessary action. What's more, the situations of imitation are roughly the same, but they can't be absolutely the same, at least the time, place, and object are different. At this point, the AI needs to make some kind of inference about human behavior and then draw conclusions about how to act. Obviously, such inferences are prone to error, because AI inferences are based on data and logic, while human behavior is adulterated with irrationality, especially emotional factors.
As a result, limitalists argue that AI alignment, while not completely useless, is of very limited use.
What's more, in human society, a large number of situational responses are uncertain, and it is impossible to extract some kind of consistent social rules. At this point, there is no alignment at all, and AI should not handle it, but it should be left to humans to make decisions. If the AI is left unexplained, it can lead to serious and irreversible consequences. Moreover, AI cannot be held accountable for its own actions, which ultimately leads to the absurd situation of "no one is responsible".
In conclusion, it is important to keep in mind that AI alignment is very limited and should not be expected too much. Many researchers believe that alignment is basically useless, but just another high-sounding cover thrown by the AI industry.
3
It is up to people to make the rules
In specific occasions and specific tasks, whether it is indoctrination or learning, it is not difficult to make AI actions meet human needs. The difficulty is that the so-called "general-purpose AI" cannot predict the "general" scenario, so it can neither "inculcate" all the coping rules in advance, nor allow it to "learn" reliable coping rules in time. It is the attempt to make machine learning models "universal" that gives rise to the so-called AI alignment problem. Many people believe that AI cannot be universally used, and that it is nothing more than a dedicated alternative labor tool.
As a result, limitedists argue that general-purpose AI is difficult to align and that it is very dangerous to make AI universal. Obviously, the danger lies not only in the fact that, like ChatGPT, it can generate false thoughts and lead humans into the chaos of "post-truth" thoughts, but also that its combination with bots is likely to lead to a large number of wrong, dangerous and even irreparable consequences of action. There are concerns that super AI may rule over humanity, and perhaps we should be more concerned that relying on AI that is not aligned will make a mess of the world.
Thinking further, machine learning models summarize human rules and let robots act according to these rules, which in turn requires humans in the so-called "AI-assisted survival society" to adapt to machine actions. As a result, the rules of the machine have become the rules of man, and man has to live according to the requirements of the machine. As a result, "we must be careful not to allow a world where our systems do not allow things to happen beyond their understanding, and they are actually enforcing their own limited understanding." ”
If the power to make rules is completely handed over to machines, and AI is aligned with humans, and humans are aligned with AI, the final result will inevitably accelerate the "mechanization of humans", that is, human beings lose their spirituality and autonomy, and increasingly become some kind of accessory of intelligent machines.
The selectorists of technological control believe that at all times, human beings must strive to control all new technological developments, including AI, to make them beneficial to human welfare. If we are not sure whether a particular development of AI will be truly beneficial, we should stop and change this approach to AI development, which is what I call the "finitarian approach to AI development." According to this view, rule-making is the exclusive right of human beings to bear the responsibilities and consequences of the rules they make, while AI is only responsible for obeying human orders and carrying out human instructions, and cannot allow it to act "without permission".
In short, AI alignment is not a confrontation between capital and science, justice and sinisterity, but a very dangerous attempt. In this sense, OpenAI's "Gong Dou" drama is another high-quality "AI propaganda" operation in the AI circle. While there are concerns that the wild growth of AI may deviate from the goal of satisfying human needs, "AI alignment" gives the public the impression that the problem can be solved entirely through alignment.
Source: Social Science Journal
WeChat editor: Zhang Jingjing
The National Development Institute of the People's Republic of China is a new type of university think tank with Chinese characteristics that the Chinese University focuses on building with the efforts of the whole university, and the current chairman is Zhang Donggang, secretary of the party committee of the university, and the current president is Lin Shangli, the current president. In 2015, it was selected as one of the first batch of pilot units for the construction of "National High-end Think Tanks" in China, and was selected as one of the top 100 think tanks in the world, and ranked first in the "Top 100 Think Tanks of Chinese Universities" in early 2018. In 2019, it was selected as the first echelon in the comprehensive evaluation of national high-end think tanks, and was the only university think tank selected for the first echelon.
The National Development Institute of the People's Republic of China is actively building a high-end think tank platform with "new platform, large network, interdisdisciplinary, cross-cutting, promoting innovation and high output". Focusing on the four major research fields of economic governance and economic development, political governance and rule of law construction, social governance and social innovation, public diplomacy and international relations, it has gathered high-quality resources of first-class disciplines in the university, and has achieved remarkable results in infrastructure construction, decision-making consultation, public diplomacy, theoretical innovation, public opinion guidance and internal governance. With the goal of becoming a "leader of a new type of university think tank with Chinese characteristics", the National Development Institute is rooted in the land of China, adheres to the national strategy, adheres to the mission of the times, and is committed to building a world-class university think tank with global influence.