For example, the recently popular ChatGPT, which is comparable to humans, makes netizens have fun. And in these two days, the AIGC track was once again lively.

Jin Lei from Aofei Temple

Quantum bits | Official account QbitAI

AIGCh (AI generated content) , this concept can be said to be very popular this year.

, for example, Stable Diffusion, just say a word to it, and the painting can be generated in seconds.

, like the recently popular ChatGPT, the answers are comparable to humans, which makes netizens have fun.

△Left: Stable Diffusion painting; Right: ChatGPT answer

And in these two days, the AIGC track was and double were lively.

Because a domestic Internet giant company announced that it has officially entered the AIGC.

and is different from other players' single route. As it enters, it directly covers four major sub-fields: AI images, AI music, AI text and AI programming.

For example, in the AI image , you only need to enter a sentence in the WeChat applet to generate a work in seconds, and it is the kind that can be held in Chinese and English:

A cat eating noodle.

is honest, not only is it realistic in terms of image, but it even portrays the cautious and cute spirit that the kitten has.

not only that, it can also generate pictures of various other styles:

AI composition is even more troublesome, directly enjoy the work:

In the ability of AI text , it gives an upper line, it can understand its meaning in seconds, give a second lower line of synergy:

Not only that, the ability to chat, Q&A, Chinese-English translation, content continuation is also no problem.

, for example, AI programming , you only need to simply install a plug-in on VS Code, press the "Tab" key at the cursor, and a whole piece of code can be presented immediately:

"Internet overseas giant" support, and several major models are released and opened in one breath. This can be said to be a major event worth paying attention to in the AIGC track this year. Who is this player

?

does not go around the corner and makes such a big move, it is Kunlun Wanwei .

As for why it is said to be "worthy of attention", it can be summarized into two points.

The first is to have technical accumulation.

Kunlun Wanwei actually formed a team of more than 200 people two years ago to start planning AIGC. The models under "Kunlun Tiangong" launched this time include Tiangong Clever Paint, Tiangong Yuefu SkyMusic, Tiangong Miaobi SkyText, and Tiangong Intelligent Code. They have covered the ability to generate multimodal content such as images, music, text, and programming.

In addition, friends who are familiar with Kunlun Wanwei know that when it was first established in 2008, its business was mainly web games, including the self-developed " Three Kingdoms Storm ", "Martial Arts Storm", " Thousand Army Break ", etc.

and then Kunlun Wanwei chose a way to enter overseas. In addition to self-development, it also operated many popular games such as "Need for Speed: The Supreme Collection 2015", "Angry Birds 2", and "Royal War".

But when it comes to its gradual growth, it is also due to the transformation of business and strategy, that is, upgrading from a single game platform to an Internet overseas company.

From 2016 to the present, Kunlun Wanwei's business has formed five major businesses, including overseas information distribution and metaverse platform Opera, overseas social entertainment platform StarX, global mobile game platform Ark Games, leisure and entertainment platform Xianlai Interactive Entertainment , and investment sector.

It is also in the process of this diversified business transformation that Kunlun Wanwei has achieved "quantity accumulation" in the technical scope involved in today's AIGC.

As we mentioned just now, the content covered by Kunlun Tiangong is aimed at the four popular areas of AIGC, namely images, music, text and programming.So in terms of the various sub-fields of AIGC, what level is the open source released by Kunlun Wanwei?

Take Kunlun Tiangong AI image capability as an example. It is understood that Tiangong Clever SkyPaint is a multi-language Stable Diffusion branch model.

In other words, while retaining Stable Diffusion's ability to "understand" English text, it increases its ability to input Chinese prompt words.

Not only that, the parallel corpus has reached 150 million levels, including a collection of Chinese and English corpus, ancient poems, Chinese and English corpus, subtitle corpus, encyclopedia corpus, picture and text description corpus that can be used frequently. In terms of training,

uses a model distillation scheme and a bilingual alignment scheme. While distilling the student model using the teacher model, it is supplemented with the decoder language alignment task-assisted model training. In terms of experimental performance results,

is also better than the AI ​​image model that supports multilinguals that has been released:

AI music, Tiangong Yuefu SkyMusic is the first commercial-grade composition AI model in China. This model introduces global information constraints and implicit chord representations, and the chord accuracy reaches 92.5%, far exceeding the current industry's best 64.7%; it supports 31 languages ​​including Chinese, English, Japanese , Korean, French, etc., supports generation of lyrics from melody and copywriting materials, and supports global multi-music styles.

It is worth mentioning that the 20 AI music generated by Tiangong Yuefu has been released in more than 180 audio and video platforms such as the United States, South Korea, Indonesia and more than 180 audio and video platforms around the world.

Kunlun Tiangong is a data set training based on a 14 billion parameter GPT-3 generation model Tiangong Miaobi SkyText in the field of AI text generation. A major feature of the

model is that it covers many subdivided tasks, including sequels, dialogue, Chinese-English translation, content style generation, reasoning and poetry couplets, etc. Not only that, but from the experimental performance results, it is also better than the currently open source GPT Chinese pre-trained model:

In the field of AI programming , Kunlun Tiangong provides a multi-language open source programming model.

It supports many mainstream programming languages ​​at present, including Java, JavaScript, C, C++, Python, Go and Shell, etc. After

is "embedded" into VS Code, it can continue writing the code, or just give it a Chinese comment.

In AI programming, the generation pass rate of the model can be said to be one of the very important performance measurement indicators, and Kunlun Tiangong still performs well in experimental performance results:

It can be seen that Kunlun Wanwei’s layout on AIGC is not a matter of overnight success, but a quantitative change after technology accumulation in the long-term development process.

The second point that "worthy of attention" is also attributed to Kunlun Wanwei itself - financial strength.

After all, the AIGC thing requires a large amount of money to invest in the early R&D investment, and it is not something that small players can "play". This point of

can be seen from the recent financial data disclosed.

According to Kunlun Wanwei's semi-annual performance in 2022, in the six months of the first half of the year alone, the company has achieved operating income of 2.25 billion yuan, and the overall gross profit margin has been further improved to 79%. If

pulls the time node longer, as Kunlun Wanwei expands its diversified businesses and continuously injects artificial intelligence technology into it, it will also undergo major changes in its segmented businesses.

, for example, Opera browser, has successfully transformed from traditional browsers to technology-driven information distribution and metacosmic platform, and the revenue changes brought about by this are also visible to the naked eye.

single is the first half of this year. Opera achieved operating income of US$150 million, increased by 34% year-on-year, and achieved adjusted EBITDA of US$23.92 million, a year-on-year increase of 572%.

As of the third quarter of 2022, the commercialization capabilities of emerging businesses such as Opera News and Metauniverse have been continuously improved. Among them, the cloud game community GX.Games has exceeded 2,000 games online, with Opera GX exceeding 18 million monthly active users, and Opera's average global monthly active users reaching 321 million.

, like its overseas social entertainment platform StarX, has more than 280 million registered users worldwide, and its revenue has maintained strong growth. Ark Games focuses on the self-developed new MMORPG product "Tower of the Holy Land", and its highest monthly revenue has exceeded US$3 million.

In addition, Kunlun Wanwei has also made a lot of achievements in the investment field. For example, Grindr, the world's largest LGBTQ social platform, was recently listed. The major shareholder behind it was Kunlun Wanwei. This investment alone made Kunlun Wanwei's profit exceed 3.3 billion yuan.

This time, the open source AIGC model is also the inevitable result of its continuous increase in investment in scientific and technological research and development.

For example, in AI images, Kunlun Wanwei uses the A100 (80G version) , the most computing power graphics card cluster on the market, which can provide ultra-fast video memory bandwidth , exceeding 2 trillion bytes per second, processing super-large models and data sets. The AI ​​painting team trained the model for more than 90,000 graphics card hours, which shows the cost of its R&D training; not only that, in terms of AI text generation, it is also the final model obtained through A100 GPU cluster training.

According to the team, Kunlun Wanwei has invested tens of millions of yuan in funds since it began to deploy AIGC in 2020.

has technical accumulation, financial resources to research and development, and is also an Internet overseas giant. This is why Kunlun Wanwei's move is worth paying attention to.

From another perspective, Kunlun Wanwei's entry is also a recognition of the AIGC track.

Because players with strength are indeed stepping forward.

The popular AIGC track

ChatGPT is equivalent to the 3.5 version of GPT proposed by OpenAI, and the number of users reached as many as 1 million in 5 days of its launch.

, and earlier Stable Diffusion, also within just one month of its release, the company behind it was valued at 6.9 billion yuan. The popularity of

has already been speculated about the future market size and prospects:

By 2030, the AIGC market size may exceed one trillion yuan.

In addition to the two, there are countless foreign giant players entering the AIGC field, such as Google , Meta, Nvidia, etc., which have regarded this track as the focus.

Looking at the characteristics of these players, in addition to having sufficient R&D funds, it also includes technology, ecology, and landing scenarios.

In contrast, domestic development can be said to have not yet entered the real AIGC stage, which is mentioned in the "AIGC/AI Content Generation Industry Outlook Report" of the Quantum Bit Think Tank. Although

has appeared in the sub-track, some excellent pioneering companies and research institutions, the industry has not yet entered the stage of large-scale verification and systematic development.

For example, in terms of scenarios, the most important thing at present is that the demand in the content field is relatively saturated, and cost reduction and efficiency improvement are not urgent enough. In terms of segmentation, the virtual human industry will become a key driving force for the industry's medium- and short-term growth. Players need to look for scenarios with high fault tolerance, limited professionalism or creative requirements.

But in the long run, AIGC still needs to have clearer and clear consumption implementation scenarios, such as real-time scenarios (such as real-time interaction, real-time feedback scenarios) and highly personalized generation scenarios.

For example, in terms of players, there are currently fewer start-up players in China, with most of the start-up players in sub-tracks below 5, and most of them are internal exploration projects with major related business manufacturers/research institutions.

Because the market size is relatively limited in the short term, and the accumulation of channels and early resources takes a long time, we believe that high industry concentration will be more likely to occur in the later period. In terms of
technology, in addition to some companies with earlier layouts, the domestic AIGC field's ability to large-scale stable production needs to be further enhanced to complete the transformation from a taste-style approach to formal and business-oriented entry into various scenarios.

According to industry insiders, it is expected that within 2-3 years, various AIGC capabilities will be expected to achieve industrial production under the premise of controllable costs.

is finally in terms of monetization. Although the domestic fields such as Game AI, virtual people, TTS, structured text writing have formed stable business cooperation and revenue relationships, the overall monetization ability of the industry still needs to be further proved. And it is basically a project system/long-term project system, and standardized/platform products lack development.

Therefore, after various analyses, we can think that the current development status of domestic AIGC is still in an attempted stage. In most cases, AIGC has not yet entered the business scenario and its monetization capacity needs to be increased.

Then the next question needs to be discussed is: What changes will Kunlun Wanwei enter the game bring?

From the perspective of technology and R&D funds, as we mentioned just now, Kunlun Wanwei is qualified to enter the AIGC threshold.

. Judging from the open source and release of four major models, Kunlun Wanwei’s focus is to promote the development and construction of the ecosystem, not just to make pure applications, but to gain more initiative from the model level.

is first of all, from the perspective of Kunlun Wanwei itself, with an average global monthly active users of nearly 400 million, and its diversified businesses have covered many important AIGC scenarios.

This is exactly in line with the characteristics of foreign giant players who have "ecology" and "landing scenarios".

Not only that, Kunlun Wanwei also expressed its view on the most difficult monetization problem of domestic AIGC at this stage: the

model will not be charged after it is open source, but it can charge for commercial support and cloud services through methods such as Mango DB and Databricks to form a source of income.
Our benchmark companies are Databricks, MongoDB and the successful Stability AI. These companies have also had a low market value in in the capital market.
We firmly believe that the iteration speed of open source is far greater than that of closed source. The ultimate goal is to make the open source cloud project complement the large manufacturer Saash, and give small and medium-sized enterprises, small and medium-sized developers, and ordinary developers a better choice.

So what do you think about this set of "playing styles" of new domestic AIGC giant players?

—End —

Quantum bit QbitAI · Toutiao ID signing

Follow us and learn about cutting-edge technology dynamics