New stories cannot fill in old losses.
@New knowledge and originality of science and technology
Author/Wang Siyuan Editor/Yi Page
After all, Shangtang Technology failed to regain the confidence of investors by adding the big model of chasing the wind.
On April 10, Shangtang released the "Nirixin SenseNova" large model system, and displayed a number of products at one go, including the ChatGPT like product "SenseChat", the second painting platform consistent with Midjournal's painting style, the AI digital human video generation platform Ruying, and the 3D content generation platforms Qiongyu and Gewu.
A large number of product matrices have the potential of "AI family bucket", but this has not met the expectations of the secondary market. After the upward jump on April 11, the stock price fell to below HK $3 again.
In fact, Shangtang Technology, as the first company to write AI big model technology into the prospectus, does have enough strength and unique advantages. However, this does not mean that Shang Tang can take advantage of the East Wind to turn around and sing. Whether it can seize the opportunities given by the market, there is no positive conclusion at present.
"Shell Doubt"
Like other domestic companies that release products based on the GPT concept, the Shangtang model was questioned once it appeared.
Some netizens found that Shangtang showed the same picture as that on the AI model station Civitai when demonstrating the second painting platform, and even there were text notes showing Civitai at the bottom of the model. Therefore, it is doubted that the self researched student map is "a map directly from Civitai".
Shangtang responded: "SenseMirage includes Shangtang's self-developed AIGC model, and also provides third-party community open source models to support importing open source models from multiple platforms or uploading users' local models. Users can avoid localized deployment processes, and efficiently generate more diverse content based on the open source model self training model."
In plain terms, Shangtang has trained its own big model, but Seconds is a platform that provides third-party open source models, and models on other third-party websites can also be imported normally.
For example, although I am a supermarket owner, only a small amount of goods in the supermarket are produced by myself, and most of them are distributed by other manufacturers.
The so-called self research model of Shangtang has also been criticized by netizens as "cloud based stable diffusion". These guesses are not groundless. From an example displayed at the press conference, it can be seen that although it is a Chinese prompt, its style is the tone of translation from Chinese to English. For example, "amazing image and color grading should be professional".
It is commendable that many of the products released by Shangtang are directly demonstrated in Demos, rather than full concept PPT.
When demonstrating the medical scenario case of "negotiation", after the user proposed symptoms, "negotiation" did not give disease diagnosis, but made multiple inquiries. According to the length of illness, other symptoms, etc. of the user, the possible symptoms and recommendations of the medical department were given, more similar to the actual manual consultation.
"Seconds Painting" is a platform for creating cultural and biographical images. Its logic is similar to that of Midjournal. In addition to self selecting styles, users can also add batch images and train their own image styles.
"Ruying" is the AI digital human video generation platform launched by Shangtang. Users only need to upload a piece of live video material to generate a digital human avatar with natural voice and movement, accurate mouth shape and multilingual proficiency.
Although these platforms have not yet been widely opened to the public, and the actual level is unknown, it is not difficult to find from these landing demonstrations that the idea of Shangtang is slightly different from that of other manufacturers, and the layout of large models is more like to sell computing power and services. Shangtang also said frankly that "Nissin SenseNova" will provide a variety of flexible API interfaces and services for government and enterprise customers for subsequent access by partners.
Of course, this is also the advantage of Shangtang, but it is not clear how long it can lead and how much real revenue it can bring.
Arithmetic is not equal to strength
In the short term, computing power is undoubtedly the top priority in this round of AI big model business competition.
The AI big model is different from the traditional "recovery" AI, which is composed of many simple neurons. These neurons are interconnected to form a huge neural network. These neural networks need to learn from a large amount of data in order to better complete tasks, such as natural language processing, image recognition, etc.
The training involves a lot of complex calculation processes such as matrix operation, gradient calculation and parameter update. These calculations need to be carried out on large-scale data sets so that the model can learn enough information. The huge amount of computing naturally requires the use of powerful computing resources such as high-performance computers and GPU clusters.
This is also the reason why GPU hardware is snapped up by enterprises entering GPT all over the world.
Shangtang just has an extremely sufficient reserve of GPUs. "There were 10000 A100 chips before it was discontinued last year, which can completely cover the consumption of training a language model with 100 billion parameters."
Industry insiders said, "To train generative AI like ChatGPT, the computing power required needs at least 10000 Nvidia A100 chips." According to public information statistics, at present, only six companies in China have such hardware strength, namely Shangtang, Baidu, Tencent, Zijie, Ali and Magic Square.
However, there is a big gap between Shangtang and the computing power reserves of several Internet giants. Relevant information shows that there are at least tens of thousands of A100s on Alibaba Cloud, which can reach 100000 on the whole. Alibaba Group has five times the size of Alibaba Cloud; Baidu has self-developed and mass produced GPU chips, and there is no need to worry about computing power.
In addition to NVIDIA's dedicated graphics card, Shangtang also purchases domestic GPUs. It is reported that many Cambrian and Haiguang GPU cards have been adapted in its large device.
However, the adaptation of domestic GPUs is also a problem. An industry expert said, "Currently, only the A100 and A800 are competent for large-scale model training, and the ease of use and cost performance of domestic GPU cards cannot be compared". That is, the current domestic GPU is not able to support the training of super model, and more investment is needed to optimize it.
That is to say, in the future, high-end GPUs will be in short supply due to comprehensive factors, which can be covered up by rich training experience and greater cost investment; However, in the long run, how to continuously obtain high-end GPUs is the issue that needs to be considered by Shangtang. If the embargo still exists, and the domestic GPU cannot adapt to the training of the super model, the distance between the domestic GPU and the international cutting-edge AI enterprises and the most advanced large model will continue to be widened.
On the one hand, we cannot deny Shangtang's hard power in calculating power reserves, but on the other hand, we cannot ignore Shangtang's lack of high-quality data corpus.
"In China, most of the high-quality Chinese language materials are in the industry, not in the public domain Internet, and even large factories are difficult to obtain key data of the industry segments." An AI co founder said.
If the data is insufficient, the model cannot learn from more aspects. This may cause the accuracy of the model to decline, making it unable to complete the task well. Missing data may make the model more vulnerable to adversarial attacks and interference, leading to reduced robustness, making it more vulnerable to attack and deception by attackers.
Fortunately, Shangtang's business model prefers to provide computing power, so customers can upload data for training. But this will involve another problem. Big factories have their own big models, and there is no reason to cooperate with Shangtang. Some small and medium-sized factories, because they are relatively vertical, do not have much data. Under the general background that it is not difficult to build a big model, and in combination with the optimization of algorithm efficiency, they can completely set up their own big model teams for training. This can also ensure the security of their own data. After all, no one wants to give their core data to others.
In addition, domestic AI models are entering the "100 model war". In addition to Baidu, Huawei, Ali and other major manufacturers, Kunlun, Zhihu and other waist enterprises are also "entering" one after another. From the perspective of 'new knowledge of science and technology', an obvious trend in the future is that the leading enterprises in all industries and fields will almost launch more vertical large model products, and small factories in the track are also more inclined to stand on the shoulders of their own industry giants. At that time, it is still worth thinking about where Shangtang can put itself.
Of course, this is just a conclusion based on expectations. The development of the big model is bound to bring new growth points to Shangtang, but can this change its dilemma of long-term losses?
Can't solve old problems every day
Although Shangtang has laid out AI for many years and has a good accumulation in some subdivisions such as computer vision, its profitability has been worrying due to factors such as commercialization and R&D investment.
In 2022, Shangtang Technology will achieve an operating revenue of 3.809 billion yuan, a year-on-year decrease of 18.97%; The net profit attributable to the parent company was 6.045 billion yuan, up 64.73% year on year, and the adjusted net loss increased 233.9% year on year to 4.736 billion yuan.
This indicates that its performance in the last year has suffered a considerable decline. On the one hand, the decline in the income contributed by the two major businesses, Smart City and Smart Commerce, led to the expansion of the net loss; On the other hand, the increased losses were caused by R&D investment, impairment of financial assets and contractual assets, and net foreign exchange losses.
In 2022, Shangtang will spend 4 billion yuan on R&D, up 11% year on year; The R&D human efficiency has further improved by 90% compared with last year, with 9.35 R&D models per capita per year; The cumulative number of commercial models increased by 93% to 67000.
Despite the special "money burning" label of AI enterprises, investors of Shangtang have little patience in the face of increasing R&D investment year by year and long-term unrealized performance.
On June 30 last year, the equity of investors and cornerstone investors before the listing of Shangtang was basically lifted. That is, within a short day after the lifting of the ban, the market value of Shangtang Technology lost 91.5 billion Hong Kong dollars. In the following six months, Alibaba, Softbank, Global Capital and other investors have reduced their holdings and cashed out.
If combined with the experience of OpenAI, Microsoft has "burned" 13 billion dollars since it injected 1 billion dollars into it in 2019, but has not yet earned optimistic income. According to PitchBook, a research organization, OpenAI is expected to generate only $200 million in revenue this year.
In the face of poor performance and poor capital market, Shangtang has no small doubt about whether it can continue to invest capital in the future compared with the Internet giants that entered the market in the same period.
Optimistically, the Shangtang model can well empower the current four major businesses, and use the previous customer channels to promote the RRS model system. However, the amount of added value that can finally be brought remains to be tested by the market.
After the listing, Shangtang kept telling stories to the capital market, from smart cars to the meta universe to today's big models, covering almost every big mouth, but the result is that the more concepts are built, the less money is earned. In addition, once the liquidity of the capital market decreases, the storytelling mode becomes less and less interesting. In the era of the big model, if the business soup with inherent advantages can change the normal situation, it might be better to give more patience.
(Statement: This article only represents the author's view, not Sina.com's position.)
