The most sci-fi episode: Tsinghua professor used AI to write a novel and won a science fiction award

[Text/Observer.com Zhou Yi, Editor Zhang Guangkai]

Shen Yang, a professor at Tsinghua University, has recently become "popular" again.

Not long ago, the 5th Jiangsu Youth Science Fiction Contest announced the list of winners, and "The Land of Memories" signed "@硅禅" won the second prize.

@硅禅和 origin of "The Land of Memories" is not ordinary: as one of the earliest scholars to start the study of the metaverse in China, Professor Shenyang specified the Kafka style, and after 66 dialogues, about 43,061 characters were output by AI, and he copied 5,915 characters from it, completing the version of the work - @硅禅这个名字, also by AI.

This is probably the most unusual second prize: for the first time, AICG (Generative Artificial Intelligence) works are entered and won alongside human works.

Image reproduced from Modern Express

The biggest "spoiler" in the technology industry this year is the big model.

On the one hand, the popularity of ChatGPT has made people aware of the disruption brought about by the progress of AI technology, and various general models and industry models have sprung up, and on the other hand, AIGC (generative artificial intelligence) is becoming the "standard" of more and more products, which are being integrated into all aspects of information search, processing and output, and to all aspects of life.

The discussion about the capabilities of large models at home and abroad has never stopped.

Whether to grasp the basic ability of model research and development, or to grasp the application of actual combat, China's large model industry also has the path of "cities surrounding rural areas" and "rural areas surrounding cities". Many people believe that China's large model should develop its strengths and not avoid its weaknesses: the ability of the large model base should catch up, but the combination of large models and intelligent search, intelligent manufacturing and characteristic cultural tourism products is also explosive.

AI authors, blending in with humans

At a previous event, Professor Shen Yang shared the creative process of "The Land of Memories" with the Observer Network and other media.

"Before, I used artificial intelligence to write the beginning of a novel and post it to my friends. A reporter left a message suggesting that I say: Mr. Shen, you can use this to submit your manuscript in the Human Fiction Contest to see if you can win the award - and let the judges not notice it?" According to Shenyang, the novel does not even have an artificial outline, from the outline, text, pictures, and even the pen name is generated by AI itself.

According to a previous report by the Yangtze Evening News, "The Land of Memories" was invited by Fu Changyi, director of the Science Fiction Committee of the Jiangsu Science Writers Association, to contribute from Shenyang. The two met, and there was no note in the call for papers that AI was not allowed. Fu Changyi recalled that three of the six judges voted to recommend the novel, but some people still noticed something unusual, "One of them told me after the selection that it felt a bit like the article was written by AI, and it still felt like there was a gap in reading."

In Fu Changyi's view, there is still room for improvement in this work. "I read the first version when it came out, and it didn't feel like a novel, like a script outline, and then it was revised to be much better than the first draft, but it was still lacking, and it can only be said that it has reached the state of high school students or first-year college students writing at the beginning of their careers, and there is still a gap between them and real writers in terms of technique. ”

When ChatGPT first came out, it demonstrated the ability to write novels in April 2023 test screenshots

What will be the impact of AI writing novels and winning awards?

"AI creates literary works, I think it can only be said to be a starting point now. In Fu Changyi's view, it is difficult to say what impact the award of AI creations will have in the future. Fu Changyi said, "The progress of AI is also to promote our human innovation, the biggest advantage of human beings over AI is that human beings have stronger innovation ability, curiosity and imagination, which may become a beginning, what will happen in the future we remain in a wait-and-see state." ”

Interestingly, it was long ago that there would be new "writers".

For example, there is a discussion of the Infinite monkey theorem – an infinite number of random keystrokes on a typewriter that can end up typing any work – even the complete works of Shakespeare. The idea dates back to 1913, when the French mathematician Emile Borrell discussed probability. Later, someone wrote a "monkey simulator" specifically to simulate the probability of this happening.

This probability is quite small.

Screenshot of "Monkey Simulator".

Including lowercase letters (26), spaces, commas, periods, semicolons, and hyphens, Monkey Simulator can generate 31 random characters. Regardless of letter case sensitivity and special punctuation, the probability that the first generated character will match the original text is 1/31, and the probability that the first two characters will match is 1/961 (31x31)...... By analogy, the complete works of Shakespeare have a total of about 5 million characters, so the probability of a monkey knocking out the complete works of Shakespeare is 1/31^5000000.

The article published in "Math Plus Magazine" calculated: take the lottery as an analogy, the probability of winning the jackpot is about 1/14000000≈10^7.1, then the probability of the monkey knocking out Shakespeare's complete works is equivalent to winning about 1 million lottery jackpots in a row - assuming that you win the lottery jackpot every week, it will take about 20,000 years to win in a row; in time, if the simulator generates characters at a rate of 50 characters per second, and keeps generating characters, then it is expected that in about 13.2 billion years, human beings will have 99% Read Shakespeare's complete works - but the universe has only been born for 13.8 billion years.

More than just copying and reproducing, artificial intelligence has now shown a certain creative ability. Technological progress now seems to be blurring the boundaries between literature and technology in another way. When ChatGPT and other large models pass the Turing test, human life will face a real reconstruction.

From 1 to 100: Develop large models in your application

Those who are familiar with Professor Shenyang will find that he has become more and more "homely" recently.

The last time observers met Professor Shenyang online was at an event in October. Shen Yang doesn't like to talk to people these days, but he talks to AI for 4 hours a day - he even talks to AI about the future, and what the future holds for GPT5, GPT6, and GPT7.

However, Shenyang has also maintained a long-term view that the basic principle of artificial intelligence is probabilistic models, and probabilistic models are bound to make mistakes.

In Shenyang's view, the errors of large models are mainly divided into several categories:

One is the AI hallucination, which is "talking nonsense in a serious way", and the other is that the Xi is not enough to learn, forming a blind spot in knowledge, such as asking GPT4 what is the effect of a certain drug, and it can't answer it if it hasn't Xi learned it—and the other is the quotation fallacy, "You ask who the hero and heroine of a movie are?" The AI says that the man is called Xiaoshuai, and the heroine is Xiaomei. Where did this information come from? From a large number of short videos. This information misleads the large model. ”

How to reduce the error rate of large models? In Shenyang's view, the combination with search engines is an opportunity.

First, search engines have a large amount of data and naturally have all kinds of public data; second, they have been tested for a long time in terms of content security; third, search engines provide users with data queries, and a large number of clicks and feedback are themselves a kind of interaction. Large models can aggregate search results - many of the top players in large models are also companies with search businesses such as Google and Baidu.

Since the beginning of this year, leading companies, including Microsoft, have tried to open up the link between search and AI intelligent dialogue. In Shenyang's view, this will also force the progress of search engines, "According to my judgment, after 5 to 10 years, there may be no concept of search engines, only the concept of large models." Because if you have any questions, you will be Xi to asking large models, and no longer use search engines. This is also the part that we have been focusing on a lot lately. ”

Microsoft adds AI large model capabilities to the Bing browser

In Shenyang's view, China's large model should integrate text, pictures and videos to achieve multimodal capabilities. "From picture recognition to raw map, there will be a lot more things that can be done. ”

In addition, search engines and large models should be deeply integrated, especially the advantages of being search engines, fully internalized into the research and development of large models - this may allow Chinese to take a different path from Americans. Even if the ability of Chinese 0 to 1 is weaker, the ability of 1 to 100 is not weaker than that of Americans.

"In the mobile Internet, China is the first in the world in short videos, and China is also the first in the world in live broadcasting. In the wide application of large models, we have done research, especially hoping that domestic large models can be deeply integrated with China's intelligent manufacturing and cultural digitalization, and with China's existing mobile Internet world-leading products, so that Chinese characteristics can be reflected in this process. ”

(Speeches are recorded on the spot and have not been reviewed by the parties)