Quote:

In 1857, a Russian artist, Alexander Ivanov, left his last brushstrokes on a huge oil painting.

This painting, titled “Appearance of Christ”, has been devoted to his efforts for more than 20 years, and it is also the longest time-consuming work of art in the world.

Just a year later, the artist passed away at the age of 49.

More than a hundred years later, an artist created a delicate and beautiful painting in just a few minutes, and was called out at Christie’s for a high price of 430,000 US dollars.

AI work “Portrait of Edmond Belamy”

However, on the other side of the canvas, the writer is not a “human”, but an AI composed of algorithms and programs.

Many years ago, there has been a wide-ranging debate about AI and humans “robbing jobs”. In fact, this issue does not need to be discussed. China has experienced a decline in the working-age population for 11 consecutive years, and in 2021, the net population growth will be almost zero. In the future, we will worry about whether AI can take over enough human jobs, rather than worrying about human beings being robbed of their jobs.

Vanke, Sequoia China, Shanghai Pudong Development Bank and other large companies have also begun to try to introduce digital labor. From the data point of view, the productivity has improved a lot.

But on the other hand, human’s feelings about AI are complicated. For example, in people’s hearts, there is always an idea that AI cannot replace humans in creating art and content that requires a high degree of spirituality.

For a long time, we have believed that it is the pursuit of beauty in human nature that enables our civilization to produce a large number of shocking content works. After all, artificial intelligence is not human, and aesthetics are difficult to quantify. Wouldn’t it be foolish to want AI to replace humans to create content?

However, as technology evolves, creating content will soon cease to be a human “privilege”. However, unlike the creation of consumer-level content, in the sacred hall of art, can AI artists slowly move from “science fiction” to “art”?

1. “Assistant Era”: Tool-based AI-assisted creation efficiency improvement

In the traditional impression, content creation is nothing more than two modes – PGC (producer created content) and UGC (user created content).

And both of these models have weaknesses, which Leifeng.com discussed in previous articles.

To put it simply: it is too expensive to do PGC, and it is too excessive to do UGC.

For example, if only painters are allowed to draw, the content will be difficult to mass-produce, and users will be prone to aesthetic fatigue; on the other hand, if every user is allowed to draw, on the one hand, there must be a lot of uneven content, which is a waste of time. and resources.

In this way, the AIGC (Artificial Intelligence Autonomously Generated Content) model has soil.

From the perspective of audiobooks, voice assistants, and intelligent creation assistance, AI participation in creation can help humans to output faster, cheaper, and more controllable, and has a huge empowerment for content creation and production.

In the next step, in the pre-metaverse era where the virtual and the real coexist, the content created by AI will appear more and more. It can be said that in the consumer-level content, AI will have a blowout in the next five years, which seems to be an indisputable fact.

At the 2022 Baidu World Conference, Baidu founder Robin Li specifically mentioned AIGC in his speech. He became the first person to openly discuss this topic among the domestic BAT-level leaders, and even the first person to discuss this topic in depth on the entire network.

He believes that in another ten years, existing content creation will be subverted by AI. However, this process is by no means achieved overnight, and the development of AIGC has also followed a “syllogistic” development trajectory.

When AI technology was first applied to the field of content production, artificial intelligence could only be used as an “assistant” for human beings – using auxiliary means to assist content production, such as AI speech synthesis, computing images, etc., can be regarded as AI Auxiliary typical.

In the “assistant era”, AI is still just a “ruthless content production tool” that performs simple and repetitive labor to liberate the productivity of creators.

As an example, Robin Li’s new book “Intelligent Transportation” last year received a good response from readers. In April this year, the audiobook version of “Intelligent Transportation” was launched. At first glance, it sounds like Li Yanhong’s personal narration, and the volume of this book is 200,000 words. If Baidu boss personally recorded it, it would take unimaginable time.

Here is Baidu’s AIGC technology. After collecting hundreds of sentences, it analyzes the characteristics of the sound and uses AI technology for high-fidelity synthesis, which can almost make the voice “fake the real”, saving Li Yanhong hundreds hours. .

“The technology is so powerful that it can even simulate nasal sounds,” one user reported.

In addition to the conversion of text and speech, AIGC, as an assistant, also plays a role in many fields of content production.

In many film and television works, AI face changing is a commonly used auxiliary means to assist stunt actions or other special effects scenes. On the eve of the launch of the movie “Fast and Furious 7”, the starring Paul Walker died in a car accident. The classic parting scene that was finally re-shot used AI face-changing technology to explain a complete ending to the characters in the play.

Behind the scenes of the movie “Fast and Furious 7”

Art restoration has always been a time-consuming and labor-intensive technical task. To repair the flaws of ancient artworks due to various reasons, restorers often consume a lot of time and effort.

At the Baidu 2022 World Congress, Baidu AI performed the AI ​​restoration of the fragments of the ancient painting “Fuchun Mountain Residence”.

Baidu AI Repairs Fragment of “Fuchun Mountain Residence”

With the help of AIGC technology, the ill-fated “Fuchun Mountain Residence” can finally return to its “original appearance”. Before the final announcement, not only the predecessors, but also today’s people can hardly imagine that AI has such a magical effect.

In addition to face replacement and repair, AI has other uses as an “assistant”. Built on the technical base of the Baidu Wenxin model, the Baidu AI Open Platform also supports practical assistance functions such as intelligent auditing, increased image clarity, and text error correction, liberating a lot of productivity for creators.

2. “Era of Collaboration”: AI completes “proposition composition” under the guidance

Robin Li refers to the second stage of AIGC development – which is the stage we are now in – as “AIGC’s collaborative stage”.

If we say that in the “assistant era”, AIGC is only a creation tool; in the “collaboration era”, AI is slowly becoming the main body of content creation, which can independently create content according to given conditions.

Let’s talk about the technical level first: the evolution of the AIGC model reflects the progress and development of AI from “weak” to “strong”.

In the “assistant era”, AIGC tools mostly use “weak AI technology” – also known as applied artificial intelligence technology. Without talking about those difficult terms, in short, “weak AI” can only solve a limited range of problems.

Google’s AI “AlphaGo” is the world’s best player at Go, but you want “AlphaGo” to tell you where the nearest restaurant is? Sorry, can’t do it.

The “strong AI technology” is different. It aims to face more general fields and solve users’ problems in many fields. It has the ability to learn, solve problems, plan for the future, and even think and even have self-awareness.

Of course, the current technology is far from reaching the standard of “strong AI”, and no one predicts how many years it will take for strong AI to arrive. However, in the field of content production, the generalization of AI technology is developing rapidly.

The Baidu World Congress in 2020 is the first time Du Xiaoxiao appeared. At that time, it was just a virtual AI assistant, an upgraded version of “Xiaodu”. More like “assistant” than “collaboration”.

In the past two years, Du Xiaoxiao’s boundary is still expanding, and is slowly changing from “it” to “she”.

During the college entrance examination period in 2022, Baidu’s AI virtual person Du Xiaoxiao performed a “unique skill”. At a speed of 40 seconds and 40 essays, she answered the composition of the Chinese college entrance examination national paper. After being scored by a former grader, she finally scored 48 points, which is far ahead of most candidates, causing heated discussions on the Internet.

Not only writing, but also drawing, writing lyrics, and composing music, Du Xiaoxiao is proficient in everything. Nearly 9,000 copies of her paintings were sold in 24 hours. Du Xiaoxiao was also invited to participate in the 22-year undergraduate graduation exhibition of Xi’an Academy of Fine Arts, and was evaluated as “having the level of graduates of the Academy of Fine Arts”.

And she and Gong Jun sang “every minute, every second, every day”, and all the lyrics and music were written by AI, which made many people exclaim: “The future of AIGC has come.”

In addition, Du Xiaoxiao has worked as a reporter in the Workers Daily, and can do interviews and reports. The timeliness and professionalism of the content are quite good. According to media evaluations, Du Xiaoxiao is not even inferior to professional anchors in terms of the quality of content output.

At the Baidu World Conference in 2022, host Sa Beining asked Du Xiaoxiao’s “sister”, Baidu AI virtual person Xijia, to draw a “postmodern colorful cat with hazy colors”. “. It only took a few seconds for Xigaga to create a commendable work, which also comes from the strong technical support behind Baidu’s AI basic technology.

When it comes to technical support, the AIGC era of collaboration is the era when deep learning of large models shines.

“Familiar with 300 Tang poems, even if you don’t know how to write poems, you can sing.” Li Yanhong described the key role of Baidu Wenxin model to AIGC. If deep learning is compared to a student, then the big model is like a well-informed “learner” who is stronger than ordinary students in terms of comprehension and creativity.

With the aid of Baidu Wenxin’s large model, AI can integrate learning in large-scale knowledge and massive unlabeled data. The more things you learn, the higher the natural efficiency and the better effect, which can assist AI to create more valuable and popular works.

And “Xueba” is not only stupid, they also have more efficient learning methods. Not only does it have a large amount of knowledge, but Baidu Wenxin’s large model can also learn in multiple languages ​​and across modes, drawing nutrients from various forms of materials such as other languages, audio and video, and pushing the level of AIGC to a new level.

Since its birth in 2019, the Wenxin model has always achieved excellent results, and its ability has topped the two international authoritative lists of GLUE and SuperGLUE, surpassing many international pioneers such as Google and OpenAI.

In May of this year, at the Wave Summit Deep Learning Developer Summit in May 2022, Baidu announced new progress in the development of Wenxin large models. Integrating the knowledge of learning task knowledge to enhance the 100 billion large model, the multi-task unified learning visual model, the cross-modal large model and other large model solutions can form new empowerment for the development of AIGC, allowing AI to produce Better and more valuable unique work.

In the collaborative era of AIGC, the theme is human-machine coexistence and virtual-real symbiosis. Although AI has not yet reached the level of completely autonomous creation, with the training, guidance and assistance of humans, AIGC has been able to efficiently produce high-quality content that is recognized by people.

3. “Original Era”: Unique Value and Independent Perspective

Is it possible for AIGC to be independent and original? Li Yanhong believes that this is possible.

The ultimate form of AIGC development, Robin Li summed it up as the “original era” of AIGC – that is, it can independently produce works with unique value and independent perspective without the assistance of human beings.

If this trend is understood in terms of “weak AI” and “strong AI”, then in the original era, the required AI technology must at least reach the entry stage of strong AI. In addition to being more general and smarter, the key It should be able to have some basic conditions for originality, such as aesthetic ability, emotional perception ability, and a certain degree of personality and personality.

Is AI still far from the emergence of personality? Whenever this question is asked, how many people will have scenes from science fiction films such as “A Space Odyssey 2001”, “I, Robot”, and even “Westworld”.

And the personification of AI received a lot of attention earlier this year. On June 11, according to the Washington Post, Google’s intelligent chatbot LaMDA already has a personality. Blake Lemoine, a Google AI engineer, pointed out in his research report that in the year he observed LaMDA, the latter had insight into emotions and began to have his own personality.

When this article came out, the whole world was shocked. Many people held high the banner and shouted that “the era of AI has come”; many people questioned the authenticity of the research report, thinking that the engineer was just grandstanding.

After the engineer released his research report, Google placed him on paid leave, citing a breach of confidentiality. The outside world is still arguing about AI’s generation of personality, and there is no authoritative conclusion yet.

Aside from the concept of AI personality that “sci-fi” is greater than “technology”, another more likely landing scenario for AIGC is the metaverse that is considered to be the “ultimate form of the next-generation Internet”.

If the current Internet is more dominated by PGC and UGC, then the future metaverse may be dominated by AIGC.

The metaverse can be said to be the content universe, and anything you touch is content. For example, you can set your window to be a small building with spring rain, and some people hope that the sun is shining brightly; you hope that the virtual person you are dealing with (though you may not be able to tell the difference) is Linghui Neixiu, and some people hope that it is bold and hot… In the immersive world, the five senses involving the eyes, ears, nose, body and further inner spiritual needs require AI originality, because no content production system that relies on human labor in the world can produce so much personalized content.

Therefore, whether it is the projection from the real world to the virtual world, or giving joy, sorrow and joy to the virtual characters in the metaverse, it is necessary to give full play to the imagination of AIGC.

However, in order to realize AIGC’s independent originality, AI scientists are still working on key problems.

Where is the difficulty? Many problems may still be faced by AI researchers all over the world.

The first problem is the “partiality” of AI technology. For example, Du Xiaoxiao, a Baidu AI virtual person, used a large number of poems and idioms “government” to judge the people before the college entrance examination.

However, as everyone knows, this is Du Xiaoxiao’s strengths and weaknesses – there is media analysis, with the current development of NLP (natural language processing technology), sorting out and extracting text semantics is the strength of AI; and in terms of abstract concepts, causal judgment, and trying to understand The ability of AI can often only raise the white flag.

It may be biased to say that it is “biased”. After all, the difficulty of AI processing logic and causality is much higher than semantic analysis and understanding. It can only be said that at this stage, AI does not yet have such capabilities.

Another difficulty is that AI has difficulty generating opinions and truly creative thinking. It is naturally difficult for an AI without personality to produce good or bad, and it is difficult to produce native opinions about “beauty” and “ugly”, “good” and “bad”.

Although AI engineers or work consultants will mark the samples during training (that is, to indicate what is beautiful and what is ugly), if AI is to truly achieve originality and AI-native aesthetics, AI must have the ability to natively understand the work. judgment.

But this is super hard. The so-called AI is an in vitro simulation of the operation of human intelligence, but for typical human thinking phenomena such as “inspiration”, “beauty”, “style”, and “spiritual resonance”, even modern psychology and literature and art have not found a clear proof, let alone Instruct and train machines to do these things.

However, human beings are smart after all. Scientists have noticed that even if an artistic genius like Mozart can compose music at the age of 7, there is a process from imitation to learning and then to originality. Imitation and learning are the strengths of machines.

Some people have put forward the point of view-why poetry creation has been on the decline since the Tang Dynasty, a conjecture is that under strict rules (such as rhythms), human beings have more and more repeated innovations in the combination of thousands of words, so excellent It’s also getting harder to come up with new titles.

But the strength of AI is that it can try tens of millions of combinations in an instant. Then, maybe after enough overlaps, there is an AI program that can leap beyond the singularity and begin to enter the original world. And its unique content creation mode may also form a unique genre – AI-style content.

Therefore, moving towards the “original era”, the big model that shines in the AIGC collaboration era, is about to become more critical.

At present, the scale of data acquired by AI deep learning has far exceeded the amount of reading that ordinary people can access in a lifetime, and this scale is still expanding rapidly today.

Empiricist philosophers believe that human cognition comes from constant experience-induction-summarization of the real world. Compared with humans, the learning model of AI is closer to this paradigm.

Looking at it this way, as long as the more samples that AI can contact, the more opportunities AI has to achieve the ideal form in the minds of scientists, and AIGC can also produce more works with unique perspectives and unique value.

The key role of large model learning for AIGC is also reflected here – relying on a large amount of data and pre-trained models, on the one hand, it ensures that a large amount of multi-modal data can be input into the AI ​​sample library, and assists the AI ​​to form a more complete general knowledge The knowledge system realizes AI technology from “weak” to “strong”; on the one hand, it bypasses a large amount of data labeling, making AI training more effective with less effort, and improving the speed of AI learning and self-improvement.

On the way of AIGC to the “original era”, it needs a “singularity” – a key point where AI technology changes from quantitative change to qualitative change. Although it is out of reach, in today’s continuous evolution of AI technology, it is not an impossible goal to realize the original content of AIGC.

“The breakthrough of large model technology is accelerating this development trend.” Li Yanhong said at the World Congress.

Conclusion:

In 1950, British mathematician Alan Turing argued in his famous The Imitation Game that instead of thinking about whether machines can think, it is better to focus on “whether it is possible for machines to behave intelligently.”

Over the past seventy years, the application of AI has penetrated into every corner of social life. In the field of content creation, AI also has a stronger and stronger voice.

The creation of content—especially artistic content—has long been considered an expression of human intelligence and humanity. AI has the ability to create content, and in a few decades, it may also be regarded as the beginning of AI becoming human.

At the Baidu 2022 World Conference, Robin Li specially emphasized the unique value and independent perspective of AIGC content.

Today, AIGC is only a vassal of the traditional content production chain, and many people believe that within a few years, AIGC may form an independent market and gain a group of fans dedicated to AI-driven content consumption.

Regarding the prospects of AIGC, Robin Li concluded: “In the next ten years, AIGC will subvert the existing content production model, and can generate AI original content at “one-tenth of the cost” and a hundred times the production speed.”

Leifeng Network

Leifeng Network

This article is reproduced from: https://www.leiphone.com/category/industrynews/boOKrlDyVPkRD7DI.html
This site is for inclusion only, and the copyright belongs to the original author.