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

In the last two months, the second-dimensional people have found the "correct use" of AI_play.

Take NovelAI, which has the most stalks, as an example. No matter what pictures you feed, AI can be transformed into two dimensions and output beautiful pictures without losing their absurdity.

What do you think their original pictures are? This is just one of a handful of cases where AI paints the front. In the past two months, AI painting with a large number of stem fire out of the circle, driving the whole AIGC field investment and financing, AI industry also accompanied by "milk" a mouthful.

But AIGC can do much more than paint.

AIGC achieves creative freedom? AI-generated content (AIGC) means content produced by AI. Therefore, as long as the subject of creation is AI, the content belongs to AIGC regardless of whether the content is in the form of text, images, audio, video, and multimodal content arranged in combination with them.

In this category, AI-generated NPCs and game strategies in the game domain, as well as a large number of virtual digital people on the market, may be classified as AIGC.

In December 2020, the official website of investment institution a16z published an article proposing "four stages of content generation," namely PGC, UGC, AI-assisted generation and complete AIGC[1].

Photo credit: a16z [1] When this article was published, a16z was still very conservative and thought,"We are still in the first to second stages." In fact, the hard science and technology history article mentioned that Tencent launched the AI writing robot Dreamwriter as early as 2015 to cover events for the Rio Olympic Games, indicating that the content industry had touched the third stage_AI-assisted generation at that time.

At Baidu World Conference in 2022, Li Yanhong said that AIGC has three development stages: first,"assistant stage," AIGC is used to assist human content production; second,"collaboration stage," AIGC appears in virtual human form, forming man-machine symbiosis situation; third,"original stage," AIGC will independently complete content creation.

If this is the standard, we have already realized the first stage of AIGC, and there are many products in terms of AI generating text, images, videos, etc.

Although there are many AIGC projects, they can be divided into two categories according to different positioning: product type and R & D type.

Product-based projects are dominated by start-ups, whose products are generally more commercial and profitable.

They either charge users--prompt base, an AI painting company, directly charges $1.99 per piece;Discord's Midjourney allows users to try AI painting for free, but the free version has a limit on how many times they can use it, and they pay $30 a month to commercialize their paintings. In addition, if users earn more than $20,000 a month from Midjourney, they also need to share with Discord, which is about 20%.

Either charge the company_Li Di, CEO of Xiaobing, once told Shell Hard Technology that Xiaobing makes money from other companies by providing an artificial intelligence framework. Daily Economic News, Sequoia China and other companies have used Xiao Bing's solution.

R & D projects are represented by large-scale deep learning models of technology giants, which are generally dedicated to certain purposes or serve the underlying technologies of giants, and generally do not launch products for ordinary users.

Pre-trained language models from companies such as Open AI, Tencent and Ali. Take Ali as an example, its language model, as the bottom capability of natural language technology platform, supports product search, recommendation, advertising, customer service and other functions.

Source: Machine Heart [2] R & D projects even launch products for general users, but also small-scale testing, the purpose is to obtain more learning data in use, self-learning. For example, Baidu launched AI painting products based on Wenxin model, but the use of this function requires application, and there are restrictions on the number of times of use.

In addition, in order to demonstrate R & D strength, reduce duplication of labor, promote technical exchanges and other considerations, large companies generally open these projects for academic research, but this also gives a few AIGC projects a "shortcut to entrepreneurship."

Some AIGC projects on the market do not have their own models, but are purchased or copied from other projects.

Industry media EasyAI mentioned in its 2021 article that the model of artificial intelligence writing tool Copysmith AI and its similar product Copy AI is not self-developed, but uses GPT-3 language model. Starting in October 2020, GPT-3 began to open up its natural language processing capabilities to the outside world for a fee, and it is understandable that these companies will use GPT-3 for commercialization.[3]

But a few companies still plagiarize. In August, the media revealed that an AIGC startup called WriteSonic had plagiarized Stable Diffusion and launched Photosonic AI, an AI-mapping product. Stable Diffusion is a model that is completely open and free on GitHub, but WriteSonic does not label the technical source in the product, and the company founder has not responded, which is equivalent to default.[4]

Mature capabilities and open models have brought new opportunities to the artificial intelligence industry, as well as a large number of mixed AIGC projects, but capital does not mind this and even begins to grab projects.

On September 19, Songya Huang and Pat Grady, two partners at Sequoia Capital, wrote an article called "Generative AI, a Creative New World" in GPT-3.[5]

The article mentions that AIGC is becoming faster, cheaper and, in some cases, of higher quality than human creation, and says that "AIGC has the potential to generate trillions of dollars in economic value."

But the investment boom in the primary market had begun before that.

Some investors say that AIGC investment has been hot since July and August this year, and has been overwhelming by September [6]. The investment manager of a new first-tier city in Southwest China also told Shell Hard Technology that he had seen several AIGC cases in the second half of this year, but because the project was general, he did not take action.

On October 17, Stability AI announced that it had received a $101 million investment led by Coatue and Lightspeed Ventures, with a post-investment valuation of $1 billion, and was "respected" by the media as the first unicorn in the AIGC field.[7] Two days later, Jasper, a Texas-based AIGC company, raised $125 million in Series A funding led by Insight Partners, valuing it at $1.5 billion.

Domestically, since the beginning of this year, a number of companies involved in AIGC business have obtained new financing, almost all of their financing time occurred after July, angel round financing accounted for nearly half.

AIGC-related companies raised in 2022| For AIGC, financing is about expansion.

Developing algorithms and building and training models requires a lot of manpower and material resources, such as GPT-3 with 175 billion parameters, and the training cost is 12 million US dollars. In addition, in order to train GPT-3, Microsoft also invested 500 million US dollars to build a supercomputing center specially for it, loaded with 10,000 NVIDIA GPUs.[8]

Even if you apply an off-the-shelf model, you have to rent a lot of computing power maintenance services, and the more people you use, the more computing power you use. This is also why most AI painting products do not provide unlimited free services to users_each work has a calculation cost.

For investment institutions, AIGC could create significant business value.

We can look at trends in computer vision technology. Wu Wenchao, vice president of Chenshan Capital, believes that computer vision technology has been developing for more than 40 years, but there is no good way to commercialize it. Until killer applications such as Face Recognition and Industrial Vision Inspection emerged, computer data technology greatly improved productivity and created a lot of social value and investment opportunities. AIGC technology matures three to five years later than computer vision, and is now in the early stages of a major outbreak [9].

Source: Chenshan Capital [9] Some AI industry investors told Shell Hard Technology that the AIGC industry has reached an important point of commercialization, which is already a consensus in the industry and cannot miss this opportunity.

AIGC can vote, there are VC vacillation.

A first-line VC investment manager in Beijing told Shell Hard Technology that she was more optimistic about big companies than entrepreneurial teams on AIGC products. On the one hand, most start-ups could not bear the cost of running AI big models, and on the other hand, big companies had first-mover advantages, but only if the direction and mode were correct.

In addition, she also believes that AIGC cannot replace creative workers. No matter what content is generated, the ultimate creative source is people, and service creators are the most intuitive way to land AIGC products.

AIGC, don't make money by playing tricks. Whether we invest in AIGC or not, we should still pay attention to those problems with universal significance--

What should a good AIGC product look like? In the last two months, the AIGC that has been used to play with the stem has been out of the circle one by one, but the stem map is separated from the productivity positioning of AI and cannot bring commercial value.

In the view of the aforementioned Beijing VC investment manager, investing in AIGC is "choosing business model." AIGC cannot replace creative workers. No matter what content is generated, the ultimate creative source is people, and service creators are the most intuitive way to land AIGC products. She told Shell Hard Technology,"AIGC, which can better replace the creative worker's operation process, is a good product. "

However, current AIGC products are still far from "productivity tools."

How far can AIGC go? On the whole, the role of AIGC products in assisting production is limited at present.

The hard shell technology history article also mentioned the problem of AI writing being highly stylized and products retreating behind the scenes. As for the audio and video field, AI also has hard defects such as set melody and effect rigidity, picture hunting and so on [10].

AI's understanding and ease of use are not strong enough in the most promising areas of painting to replace basic painters_there are unreasonable problems such as hand and foot incoordination, mouth and eye skew, including famous rollover cases such as "can't use chopsticks" and "don't distinguish salmon from sashimi."

For example, Baidu launched AI painting product "Wenxin Yige," gave me such a "chopsticks in both hands" work, and its keyword is "a girl eating boxed rice with chopsticks."

In short, AIGC's ability is still not satisfactory.

How far are we from the AIGC? We expect AIGC to play a role in highly commercial areas such as the film industry.

One of the reasons restricting the development of CG film is the difficulty and high cost of production. Avatar, for example, cost $261 million to produce, including special effects costs of up to $200 million. Transformers 5 cost $217 million to make; Transformers 5 cost $217 million to make, and special effects accounted for $130 million.

However, the White Paper on Artificial Intelligence Generated Content (AIGC)(2022) released by the Chinese Academy of Information Technology mentioned that AIGC can synthesize virtual scenes and combine with matting technology to generate scenes that cannot be photographed or are too costly [11]. A large number of scenes in "Hot Blood Chang 'an" broadcast in 2017 are virtually generated through artificial intelligence technology.

Although there are still mixed reviews of AI-generated video scenes, at least we have seen the shadow of AIGC applications.

References：

[1] Meet Me in the Metaverse 2020.12.7 https://a16z.com/2020/12/07/social-strikes-back-metaverse/

[2] WAIC 2021 |Ali Vice President Si Luo: How large-scale language models empower industries and create value 2021.7.18 https://mp.weixin.qq.com/ s/-SVH32XRzW5Y_FWAIr79MA

[3]Will AI change content marketing forever? 2021.4.23 https://easyai.tech/ blog / will-ai-change-content-marketing-forever/

[4]Model open source is still powerful? Mystery company became the AI "third pole" of hanging giant and learning valve 2022.8.31 https://mp.weixin.qq.com/ s / QLAwtrVeCx5vLOELr6cXUQ

[5] Generative AI: A Creative New World 2022.9.9 https://www.sequoiacap.com/article/generative-ai-a-creative-new-world/

[6]AIGC is on fire, VC is urging FA to push the case 2022.10.21 https://mp.weixin.qq.com/ s / RSSHIRlLDFI7zc6dGB47yQ

[7]With $100 million in investment, AI drawing startles unicorns, and the end of human painters is really coming? 2022.10.18 https://mp.weixin.qq.com/ s / cqiC11gAXbw4Xt3ZPWEr1g

[8] OpenAI Presents GPT-3, a 175 Billion Parameters Language Model 2020.7.7 https://developer.nvidia.com/blog/openai-presents-gpt-3-a-175-billion-parameters-language-model/

[9]AIGC: Automated content generation, AI's next tipping point?| Chensi 2022.8.24 https://mp.weixin.qq.com/ s / nJOR20fYgGVB147ZbO1JLQ

[10]Video creators lose their jobs! Meta and Google launched AI black technology to grab the job 2022.10.18 https://36kr.com/ p/1963050276766720

[11]Artificial Intelligence Generated Content (AIGC) White Paper (2022) 2022.9 http://www.caict.ac.cn/ kxyj/qwfb/bps/202209/P020220902534520798735.pdf

This article comes from Weixin Official Accounts: Fruit Shell Hard Technology (ID: guokr233), Author: Yang Jingyi, Editor: Li Tuo

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.