Shulou(Shulou.com)11/24 Report--

Some time ago, Google engineers claimed that their AI chat robot LaMDA had become conscious, causing chaos.

LaMDA is a large language model (LLM) that can predict the possible next word based on any given text. Many conversations are predictable to some extent, so such a system can promote and keep the dialogue going smoothly. LaMDA was so good at this that the engineer, Blake Lemoine, began to suspect that it had a humanoid perception.

As LLM becomes more and more common and powerful, human views on LLM become more and more different. It is true that today's systems have exceeded many benchmarks for "common sense" language reasoning, but many systems still lack common sense and are prone to nonsense, illogical and dangerous advice. So this leads to a paradoxical question: why are these systems so intelligent and so limited?

Recently, Yann LeCun, a Turing Award winner, and Jacob Browning, a postdoctoral fellow at New York University, published an article in Noema magazine to answer this question:

The fundamental problem is not the AI itself, but the limitation of the language.

He made the following argument:

1. Language carries only a small part of all human knowledge.

two。 Most human knowledge and all animal knowledge are non-verbal (non-symbolic).

3. Therefore, large language models can not approach the level of human intelligence.

Obviously, LeCun believes that we should abandon the old hypothesis about the relationship between language and thinking, that is, there is identity between language and thinking. In his view, these language systems are inherently "superficial", and even the most advanced AI on earth will never be able to acquire all the thinking that human beings have.

This view actually denies the validity of the Turing test. The basis of the Turing test is that if a machine says everything it wants to say, it means it knows what it is saying, because knowing the correct sentences and when to use them will exhaust their knowledge.

LeCun first talks about it from the perspective of philosophy of language to explain the ideological basis and irrationality of Turing test.

AI does not understand what he is saying. The 19th-and 20th-century philosophy of language holds that "knowing something" means being able to come up with the right sentence and knowing how it relates to other sentences in the great network of truth as we know it. According to this logic, the ideal language form is a pure formal language of mathematical logic, which is composed of arbitrary symbols connected by strict reasoning rules. But if you are willing to make some effort to eliminate ambiguity and inaccuracy, you can also use natural language.

The philosopher of language Wittgenstein once said: "the sum of true propositions constitutes the whole of natural science." For a long time, people have been convinced that logical mathematics and formalization are the necessary foundation of language. In the field of AI, the embodiment of this position is symbolism: everything we can know can be written in an encyclopedia, so just reading everything can give us a comprehensive understanding of everything. In the early days, the operation of binding arbitrary symbols together in different ways according to logical rules became the default paradigm of artificial intelligence.

In this paradigm, the knowledge of AI is composed of a large database of real sentences, which are connected to each other by artificial logic. The criterion for judging whether AI is intelligent or not is whether it can "spit out" the right sentence at the right time, that is, whether it can manipulate symbols in an appropriate way. This is the basis of the Turing test.

But LeCun believes that refining human expertise into a set of rules and facts has proved to be very difficult, time-consuming and expensive. Although it is simple to write rules for mathematics or logic, the world itself is very ambiguous.

So when it comes to LLM, LeCun doesn't agree with the idea on which the Turing test is based. He believes that just because a machine can talk about anything doesn't mean it understands what it's talking about. Because language does not exhaust knowledge, on the contrary, language is only a highly specific and very limited representation of knowledge. Whether it is programming language, symbolic logic or natural language, they all have specific types of representation patterns and are good at expressing discrete objects and attributes and their relationships at a very high level of abstraction.

All representation patterns involve the compression of information about something, but the contents left and left out by compression are different. The representation mode of language deals with more specific information, such as describing irregular shapes, the motion of objects, the functions of complex mechanisms, or detailed strokes when painting, if you want to describe a surfing event, consider the actions in a particular context.

There are also non-verbal representations that can convey information in a more understandable way, such as symbolic knowledge, including images, recordings, charts, maps, and so on. The same is true of distributed knowledge found in trained neural networks.

According to LeCun, the limitation of language, the characteristic of language representation schema is that it conveys very little information, which is the reason for its limitation.

From the perspective of information transmission, the bandwidth of language transmission information is very low: isolated words or sentences, no context, very little content. In the view of linguists, natural language has never been a clear communication tool, because of the large number of homonyms and pronouns, many sentences are very ambiguous.

So, does natural language hinder us from expressing our thoughts? Apparently not. LeCun points out that humans do not need perfect communication tools because we have a common understanding of many non-words. Our understanding of a sentence usually depends on our deeper understanding of its context, so as to infer the meaning of the sentence.

In conversation, interlocutors usually have the same knowledge background. For example, you talk to a friend about the football game that is playing right in front of you, or a person has some kind of targeted communication under his or her specific social role, such as a consumer ordering from a waiter.

The same is true in reading situations, where studies have shown that children's background knowledge of the current topic is a key factor in understanding a sentence or paragraph. However, AI does not perform well in this kind of common sense language test.

LeCun thus points out that the context of words and sentences is the core of LLM. Neural networks usually represent knowledge as an ability called "know-how", that is, the ability to master highly context-sensitive patterns and find rules (concrete and abstract). These rules need to be applied to handle input in different ways in a particular task.

Specific to LLM, this involves the recognition pattern of the system at multiple levels of existing text, where you can see not only how a single word is connected in a paragraph, but also how sentences are connected to form a larger paragraph. Therefore, LLM's mastery of the language must be context-sensitive. It understands each word not according to its dictionary meaning, but according to its role in various sentences.

So, what should LLM look like? LeCun's view is that the training goal of LLM should be to enable it to understand the background of each sentence and observe the surrounding words and sentences to piece together what is happening. In this way, it is infinitely possible to use different sentences or phrases as input, and to continue the dialogue or continue the article in a reasonable way. Systems that train on paragraphs written by humans often talk to each other, so they should have the general understanding needed to start an interesting conversation.

LLM understands that many people do not want to say that LLM's behavior is "understanding" or that LLM is "intelligent". Critics are right to think that LLM is just doing some kind of imitation. Because LLM's understanding of language is great at first glance, but it is actually very superficial. This superficial understanding is familiar: all the students in the classroom are talking, but they don't know what they're talking about-they're just imitating the text that the professor or themselves are reading.

This reflects the nature of the real world: we often don't know how little we know, especially when we don't get much knowledge from the language.

LLM has this superficial understanding of everything. Systems like GPT-3 screen out possible words in sentences / articles, let the machine guess the words that are most likely to appear, and then correct the wrong guess. The system will eventually be trained to guess the most likely words, making it an effective prediction system.

However, the ability to explain a concept in language is different from the ability to actually use it. The system can explain how to perform long division, but it cannot complete long division; the system can also explain which words are offensive words that should not be spoken, and then say them without pressure. Contextual knowledge is reflected in the ability to recite language knowledge, but not in the ability to deal with problems.

For language users, the ability to deal with problems is essential, but being able to deal with problems does not mean that they have relevant language skills. This situation is reflected in many places, such as science classes will require students to give speeches, but students' scores are mainly based on their experimental results. Especially outside the humanities, being able to talk about something is often superficial, or the skills that make things work smoothly are more useful and important.

Once we get below the surface, it's easier to see the limitations of the system: their attention duration and memory are roughly enough for a paragraph of text. If we are talking to LLM, this is easy to overlook, because people tend to focus on the last one or two responses and the ones they are about to get.

However, skills for dealing with more complex conversations, such as active listening, recalling previous conversations, insisting on talking about a topic to make a particular point, while avoiding interference, and so on, these skills require the system to have greater attention and memory capacity. This further weakens the understanding of the system: we can easily deceive the system by changing our point of view or speaking another language every few minutes. If you have to go back too much, the system will start from scratch, accept new ideas consistent with the old comments, change your language or admit that you believe anything you say. The understanding necessary to form a coherent worldview is far beyond the capabilities of the system.

Apart from language, giving up the wrong view that "all knowledge is linguistic knowledge" can make us realize how much knowledge is non-verbal knowledge. Books contain a lot of information that we can use, as well as instructions, papers, charts, and city maps. In addition to the information reflected in words, natural characteristics, man-made products, psychological and physiological characteristics of animals and human beings are full of information that can be used by human beings.

This shows that in addition to language, the world itself shows a lot of information for human beings to explore and use. Similarly, social customs and separation rituals can be passed on to the next generation only through imitation. A lot of human cultural knowledge is iconic and can be passed on just by watching. These subtle information patterns are difficult to express in words, but the person who receives the message can still understand it. From this we can see that non-verbal understanding is very important for the development of human beings.

LeCun points out that there is not much human knowledge recorded in words, and the knowledge of non-human primates can hardly be captured in primate communication.

We think that language is important because it can convey a lot of information in small formats, especially after the advent of printing and the Internet, language can copy information and spread widely. But it is not without cost to compress the information in the language, which requires us to spend a lot of energy to decode the information-intensive paragraphs. For example, humanities courses may require students to do a lot of extracurricular reading, or a lot of class time may have to be spent reviewing difficult articles, and so on. Although the information is provided, it is still time-consuming to have an in-depth understanding of the information.

This explains why language-trained machines know so much but know so little. Machines acquire a small part of human knowledge, but this small part of human knowledge can be about anything. This is a bit like a mirror. The system gives people a deep illusion and can reflect almost anything. But the problem is that the mirror is only one centimeter thick, and if we try to explore it, we will hit our heads.

It can be seen that human beings have a profound non-verbal understanding, so that language has the opportunity to give full play to its talents. It is precisely because we have a deep understanding of the world that we can quickly understand what others are talking about. This broader, context-sensitive skill is a basic knowledge that human beings have had since ancient times.

Non-verbal understanding enables perception to emerge and enables it to survive and prosper. So for AI researchers, finding common sense (common sense) in artificial intelligence is a more important task than focusing on the AI language.

Finally, LeCun concluded that LMM does not have a stable body and cannot perceive the world permanently, so they can only pay more attention to language, so the common sense of LLM is always superficial. Our goal is to get artificial intelligence systems to focus on the world they are talking about, not the language itself-although LLM does not understand the difference between the two. We can not achieve a deep understanding only through language. Through the study of LLM, we can see how little we can understand only from the language itself.

Reference link:

Https://www.noemamag.com/ai-and-the-limits-of-language/

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.