The curtain opens, where are the commercial protagonists of AIGC?

Source: Alpha Commune

Author: Xu Siqing

Last Thursday, a number of large language models such as Baidu Wenxinyiyan and iFlytek Spark Model were officially approved to provide services. AI has triggered a change in the paradigm of artificial intelligence. At the same time, people naturally ask - where is the money in the paradigm change? What is the protagonist?

In addition to the legend of MidJourney (11 people’s income reached US$100 million in 2022), entrepreneurs and investors have always been hopeful and looking forward to exploring the business models brought by AIGC (generative artificial intelligence). This article attempts to analyze the opportunities for entrepreneurship and investment from the perspective of the composition of large models, and by the way, at the end, it tells an entrepreneurial story of "falling into the eyes of money".

Figure 1: Example of a Vincentian image—a cyberpunk-style female killer in the rain (This image was generated by HiDream’s Pixeling V1.0.)

LLM (Large Language Model) - Large companies burn money to build an ecosystem, a trap for entrepreneurs

Let's take a look at the architecture of large language model applications (see the figure below). It is not difficult to see that the bottom layer is built on a solid hardware foundation. Currently, there are GPU manufacturers represented by NVIDIA and computing power providers represented by CoreWeave. Like mobile phones, hardware is always the last word in commercialization. This is one of the business opportunities, but there are not many opportunities for entrepreneurs and investors, except for optimizing computing power platforms for large models.

Figure 2: The structure of the large language model and some representative companies

On the other side of the ocean, there are currently only a few mainstream large models. OpenAI, Meta, Google and other major companies have begun to lay out their plans based on large models. There are also LLM (Adept, Cohere, Character.ai) founded by the authors of transformer, which are more trending. Differentiation rather than homogeneity.

From this, we see more of the ecological battles of big manufacturers - if you don’t have your own big models, there will be no future. People are optimistic about Microsoft over Amazon because it has the upper hand and has taken advantage of the two major models of OpenAI and Meta Llama, and all its products embrace AI. Those who win the model will win the world. There is a basic platform built with large models, and applications are built on the platform. They are trying to steal customers from Amazon by pulling out all the stops.

The basic large-scale model is a huge gold-swallowing beast. Large factories are trained to compete for land and settle big accounts. Small factories have almost no chance except for vertical models. Entrepreneurs should be cautious.

Figure 3: Evolution of AI models (sourced from

OpenAI and big factory models belong to the first wave, and the large models of transformer authors belong to the second wave. The various models that continue to emerge later, whether practiced by oneself or using open source models, should belong to the third wave. The development trend of large language models is from basic large models to vertical models, and then to long-tail models focused on enterprise applications.

Figure 4: Overview of China’s LLMs (pictures from the Internet)

It should be noted that domestic LLM is overwhelming. Except for a few major manufacturers, we have reason to guess that most of them are models trained with open source models (without sufficient GPUs, how can the porcelain work come from? ), belongs to model re-engineering, which saves the pre-training link that consumes almost 99% of the computing power (refer to the Alpha Commune article: Meta large language model Llama 2—turning the table and starting a new chapter | Investors said), how fast It is economical and worthy of recognition, but most of them are vertical models or Party A's large models, and are not true basic large models.

Training basic large-scale models is labor-intensive and can easily cost tens of millions of dollars. It is a game for the rich, so entrepreneurs should be cautious.

The reason is very simple. When we calculate the amount of money raised by startup companies, compared with foreign entrepreneurs of the same type, the amount of financing is about 1/10, which is not enough to train the model. If you really do pre-training from scratch, the GPU will heat up and the money raised will be used up soon after it is turned on.

AI Tool- Selling pickaxes in the gold rush, high technical content

While generative artificial intelligence creates a new paradigm, it also creates a large number of new tool opportunities. These tools range from evaluation models, operation and maintenance models, enhanced systems to operational tools, all of which are new opportunities brought about by paradigm changes.

Particularly worthy of attention are the tools for managing and training various types of data, as well as the security protection system. Security tools include both traditional security tools and model-oriented security and compliance. Just think about it, losing data is already terrible. If the trained model is invaded, wouldn’t it be an efficient tool to teach people to invade their own privacy?

Security should actually be more than just a tool, but an essential infrastructure.

There are so many old tools that can no longer be used. Entrepreneurs have more opportunities to show their talents.

New AI Application-A Paradise in Bloom

The application layer is a rich mine with many entrepreneurial opportunities.

Whether it is toB applications or toC applications, the new paradigm created by the new generation of artificial intelligence provides endless opportunities.

The characteristic of toC is its revolutionary entry into new scenes, while toB is the coexistence of old and new.

We can describe to C as a shallow application. Its characteristic is that product managers have many opportunities for innovation and can submit applications that users like without having to dig deep into the model.

It is not difficult to predict that a group of outstanding Chinese product managers will become major international players in the tide of paradigm change. Perhaps a new generation of Zhang Yiming is emerging rapidly.

The to B application tests practitioners' understanding of the pain points of enterprise needs and the use of new means to solve old problems. Except for a few disruptive innovations, most of them develop along the enterprise value path of cost reduction and efficiency increase. Technological changes provide an opportunity for one trick to be widely used, but revolution is not easy to succeed. In the final analysis, traditional players occupy all the application scenarios, and new methods and tools play an innovative role. It should be a pattern where the old and the new coexist.

Party A's large-scale model deployed privately should have rapid growth in demand for a long time.

The only thing worth noting is that there are huge differences between China's toB market and the US toB market due to different customer compositions and purchasing behaviors. Entrepreneurs must not copy them. I am going to Sanya to attend the China CIO Summit in the past two days. I look forward to exploring it with industry experts. This topic will be reserved for future special discussions.

Wensheng Vision - a proven bonanza; the Chinese version of MidJourney - a product of both pleasure and pain

MidJourney directly appeals to designers and greatly improves the efficiency of art design. It can produce a picture in seconds. It can be said that it has completely changed the production method. Users have a strong willingness to pay. 11 people earn 100 million US dollars a year, becoming the AIGC so far. The highlight of business.

At the beginning of this year, the author and the designer worked together for a long time, studying MidJourney, and experienced on-site the C-side payment and B-side payment - the hard truth of the business model.

At the same time, I also strongly feel that the threshold for use is high - a great artist constantly corrects complex and professional English descriptions, and through "repair" (prompts), in order to produce satisfactory creative drawings, while at the same time enduring the defects of the Chinese style of the model - —Since at that time there was only a model like MidJourney that only knew English and had only studied Western cultural materials, designers were forced to use various means to improve the accuracy of English, and had to endure the dilemma of expressing "Chinese" as Japanese or Korean style .

Subversive means bring about a revolution in design efficiency. Designers simultaneously endure the mismatch between language and painting style. The combination of intense pleasure and pain makes us firmly believe that China needs its own MidJourney. Not only that, we also believe that MJ should be surpassed for two reasons: first, the market calls for tools that understand human language better than MidJourney; second, Chinese elements or global local elements should be well expressed. In addition, e-commerce and advertising have a huge market for high-efficiency production tools with a large number of pictures and short videos as the core, and for docking production systems.

God-given opportunity, I hit it off with Academician Mei Tao, a disciple of HKUST who has been trying to "do things" together for several years. A veteran who understands algorithms, knows models, and has practiced in the industry for many years, so we have seed round support from Alpha Commune and HKUST alumni. HiDream started by Academician Mei Tao (see article: What other entrepreneurial opportunities does AIGC have in China? Academician entrepreneurs and angel investors say this | Alpha Founders Club). In just a few months, HiDream has demonstrated unimaginable iteration speed and stunning results, completely overturning the long-standing "scholar entrepreneurship bias" in the industry.

The following is a comparison chart of examples generated by MidJourney and Hidream under the same conditions.

：Dungeons and dragons character portrait, dark short - haired woman, Wild West tracker holding a brass spyglass, Emma stone, clever, upbeat and witty, and strong

：A young French Bulldog appears confused after being ambushed, wide-eyed and stunned, pixar style

：Alien looking strange cute happy little bunny

Figure 5-7: MidJourney and Hidream graphs generated by the same

MidJourney has done valuable pioneering work. If latecomers can better apply transformers, build stronger model capabilities, and no longer simply rely on the Diffusion Model (MJ continues to optimize the current tool based on this model), they should There is much to be done. We also firmly believe that MJ will take advantage of the early start, get rid of dependence on the diffusion model, and continue to bring surprises to the market.

Whether it is tools, superficial applications, or deep enterprise applications, they are all in the early stages of scale generation and undercurrents. AIGC has just begun, and we look forward to the innovation and iteration of algorithms to continue to show more brilliant chapters for the industry.

**In the early morning of September 6, 2023 in Hefei. **

**The author of this article is Xu Siqing, founding partner and CEO of Alpha Commune. **