Burning Dimension (ID: chaintruth) original
Author | Mashu leaf
Edit | Xie Zhongxiu
Overnight, big factories have rolled up AI models.
Following Baidu Wenxin's words, the 360GPT large model and Huawei's "Pangu large model" quickly followed, and only on April 10, 2023, three large models competed for the "official announcement" in one day, including Sogou founder Wang Xiaochuan's "Baichuan Intelligence", SenseTime's Riday-new SenseNova large model, and Kunlun Wanwei's large language model "Tiangong" 3.5.
On April 11, Alibaba announced that all products will be connected to the "Tongyi Qianwen" model. In the upcoming May, Tencent's hybrid model will also be released.
In addition to large factories, there are also many "star entrepreneurs" who have entered the game.
In February 2023, Wang Huiwen, the former co-founder of Meituan, released the "AI Hero List" and quickly launched a round of financing; Subsequently, Kai-Fu Lee, CEO of Innovation Factory, Wang Xiaochuan, former CEO of Sogou, Zhou Bowen, former head of JD AI, and Li Zhifei, founder of Mobvoi, set off a wave of AI large-model entrepreneurship. Ali AI bull Jia Yangqing was also exposed to devote himself to AI entrepreneurship.
Large factories "open the roll", and capital quickly smells the smell. "The AI big model brought by ChatGPT is undoubtedly the direction of important attention of capital in the past six months." William Iam, a tech track investment follower, said.
"In layman's terms, the AI big model is an infrastructure, more similar to the role of the public cloud in the Internet, covering more knowledge areas, more intelligent output content, for example, ChatGPT through machine learning to achieve complex speech, image, speech processing, can not only like a doctor, but also like a writer to write articles, but also as an intimate customer service." William explains.
Entrepreneurs and segments have also benefited. "Now that big factories roll up AI models, we entrepreneurs are looking forward to more and more life." AI freelancer Ah Bin said excitedly.
"Large manufacturers can invest more resources in the research and development of large models, bear higher costs, and continue to enrich AI application scenarios." Ah Bin said to Ran Dimension, "With ChatGPT, a liberal arts student of mine actually built his own language learning website, and even the illustrations of the website are generated by Mid Journey, as long as the current work and AI tool process is run, the AI big model allows entrepreneurs to complete the original 1-2 days of workload in 3 hours." ”
In addition to using AI large models to work for themselves, AI entrepreneurs can also "redevelop" based on large models, "For example, using the Wen Xin One Word Big Model, based on its voice and text processing functions, developers can add online accounting small models, and they can use voice input to achieve real-time bookkeeping, and the convenience is second to all accounting apps on the market." AI entrepreneur Obayashi added.
Especially in some vertical application scenarios based on the development of AI large models, it has also become a new opportunity.
For example, Huawei's recently released Pangu Meteorological Model, a high-resolution global AI weather forecasting system, can predict global weather conditions in the next seven days in seconds, which is 10,000 times faster and 20% more accurate than traditional forecasting algorithms. At the same time, 650 enterprises have been the first to access the "Wen Xin Yiyan", covering various scenarios such as the Internet, media, finance, insurance, automotive, and enterprise software.
However, in the eyes of many industry insiders and observers, we should not be too optimistic about AI large models.
After all, the current domestic computing power scarcity, insufficient data model training, and even copyright disputes and other problems are difficult to ignore, "the current big language model still has the problem of 'nonsense' output error information, ChatGPT pearl jade in front, leaving little time for big manufacturers." Ah Bin added.
On April 11, the Cyberspace Administration of China issued the "About
It's just that maybe the road is bright, but the process towards the goal must be careful and steady, in order to go further and better.
01
Big factory unwinding
"Facing the AI era, all products are worth upgrading with large models." On April 11, at the Alibaba Cloud Summit 2023, Alibaba Group Chairman and CEO Daniel Zhang described the vast future of AI big models.
But behind this sentence, AI large models have long become the focus of recent development of domestic Internet manufacturers, and they are poised to "open the book".
In 2023, ChatGPT will explode, and while bringing the fire of AI models, it will also drive domestic manufacturers to enter the game and compete for the next "domestic ChatGPT".
On March 16, Baidu took the lead in releasing "Wen Xin Yiyan", at the press conference, Baidu founder, chairman and CEO Robin Li said, "Wen Xin Yiyan's expectation is to benchmark against ChatGPT, and even GPT-4. ”
Then, on March 29, 360 founder, chairman and CEO Zhou Hongyi showed off its first artificial intelligence chat platform. On April 9, 360 officially announced that the artificial intelligence product matrix "360 Wisdom Brain" developed based on the 360GPT large model landing search scene will be open for internal testing for enterprise users on the 16th.
On April 8, Tian Qi, chief scientist of HUAWEI CLOUD's artificial intelligence field, introduced the progress and application of Huawei's "Pangu model" for the first time.
Immediately after April 10, three large models competed for the "official announcement" in one day, including Sogou founder Wang Xiaochuan announced that the large-model product that was about to be put into entrepreneurship was called "Baichuan Intelligence", and announced that he had received $50 million in start-up capital; There is also SenseTime, known as one of the "Four AI Tigers", released the new SenseNova large model; Boutique game developer Kunlun Wanwei announced the upcoming release of the large-language model "Tiangong" 3.5, and launched an invitation test on April 17.
At that time, Kunlun Wanwei also said that the "Tiangong" large model was very close to the intelligence level of ChatGPT.
The "volumes" are endless. On April 11, Alibaba also said that all Alibaba products will be connected to the "Tongyi Qianwen" model in the future. At the summit, Kunlun Wanwei also announced cooperation with Alibaba Cloud in the construction of intelligent computing centers and large model training.
Source/Visual China
On April 11, ByteDance's office software Feishu released a video announcing the exclusive intelligent assistant "My AI", different from the AI intelligent tools of other large manufacturers, "My AI" seems to have no intention of copying a large language model similar to ChatGPT, more similar to Microsoft's Office Copilot.
Now, many domestic technology giants, including Baidu, Tencent, Alibaba, Huawei, JD.com, 360, etc., have announced that they have corresponding layouts in the generative AI industry, and in May, Tencent's hybrid model will also be released.
Behind the big factory rushing into the AI big model battle, perhaps to create another future.
Jiang Han, a senior researcher at Pangu Think Tank, said that there are also "unspeakable hidden" behind the rush of large manufacturers to launch AI large models: "At present, if enterprises do not have AI large models, it is likely to trigger market distrust of enterprise artificial intelligence technology." Therefore, "as long as it is an Internet artificial intelligence company, almost all of them are rolling up the big model." ”
For large manufacturers, in addition to using AI models to bring new user traffic and user attention, senior Internet observer Ding Daoshi pointed out, "Using AI models to transform and upgrade business and improve revenue may be the ultimate goal of large factories." ”
Just as GPT-4 is fully integrated into Microsoft Office and Bing, according to statistics Data.ai mobile Internet research company, the number of global downloads of Microsoft Bing Search has increased by 8 times. Today, Baidu Wenku also provides users with functions such as dialogue interaction and content creation by accessing Wenxin Yiyan.
The layout of AI large model tracks and large factories is frequent, and there is also an "economic account" behind it. "Big models are the development trend of artificial intelligence, and large models have high financial and technical thresholds, and the development and training costs about 12 million US dollars at a time, which requires deep optimization of AI frameworks and super parallel computing capabilities, and large models may incorporate highly customized small models in the future, resulting in market concentration to large companies." Tian Qi pointed out.
It's just that whether the market will pay for the new story of the AI big model of the Internet giant, it seems to be observing.
On March 16, Baidu, the "first to eat crabs", fell by more than 9% on the day of its release, and the next day its stock price soared higher, closing up more than 13%. On March 30, 360 released the large model and the next day, the stock price opened 5.53% higher, and then rose again on April 3. On April 11, Kunlun Wanwei rebounded strongly the next day after plunging more than 17% on the day of the release, rising 9.5%, and the highest intraday rise was nearly 16%.
To a certain extent, whether it is Baidu, 360, or Huawei, Alibaba, Kunlun Wanwei and SenseTime, the market also votes with money, indicating the thirst for ChatGPT-like AI models.
This has also become a major reason for attracting Internet manufacturers to open AI models.
02
Entrepreneurs drink soup
Big manufacturers have become the main force of open-book AI models. After all, as Tian Qi said, although large models are the trend of artificial intelligence development, large models have high financial and technical thresholds, and they are not a cake that entrepreneurs and small companies can easily eat.
But when the big factory is opening the volume, small factories and entrepreneurs have also found opportunities in the wind.
"Although small start-up companies do not have as strong manpower, material and financial resources as large factories, if they can solve the pain points in a certain field through small models based on the AI models of large factories, they will also receive the attention of capital." William said.
In addition to the large manufacturers of "domestic ChatGPT" under the unified volume, the figure of small entrepreneurial projects is also very active.
At present, william's investment institution is paying close attention to the small entrepreneurial team represented by digitextend, taking digitextend as an example, "The founder team is from iFLYTEK and Microsoft, and based on ChatGPT self-developed algorithm made a small model, for enterprises to provide similar assistant summary and induction services, for the answers generated by large models, through small models to further improve the accuracy, can become the advantage of small factories." ”
In addition, Tsinghua-affiliated AI startup DeepLang AI has recently completed a new round of high-valuation financing, with a latest market valuation of about $100 million. Facewall Intelligence, which focuses on the acceleration of large models and application landing, also previously reported that it was acquired by Wang Huiwen, and high-level financing has also advanced rapidly.
In addition, AI big models are also providing possibilities for more AI entrepreneurs.
"AI big models are the strongest assistant for small and medium-sized entrepreneurs." Obayashi said that as one of the first ChatGPT users, he noticed that ChatGPT has begun to support entrepreneurs to develop plugins for their user portals, "for example, users can search for travel guides, book travel agencies and hotel tickets through ChatGPT Q&A." ”
For Daelim, "this allows entrepreneurs who originally had to reach users through the app store to see a new traffic portal, and entrepreneurs can save themselves from being charged 30% of the advertising promotion fee by the app store." ”
At present, Daelim is developing an online accurate translation APP based on ChatGPT's powerful text translation capabilities, "As long as a certain amount of manual intervention and correction of ChatGPT's translation results is carried out, a more accurate translation effect can be achieved." At present, Daelim has contacted 2-3 investment institutions.
Photo: AI entrepreneur discussion group (left) A Bin's website made with a large model (right) Source/Screenshot of Burning Dimension
Like Dalin, Ah Bin also sighed, "AI big model is undoubtedly the hottest cyber partner, in the current AI entrepreneurship group, discussing how to use large models to start a business has become a new trend, an entrepreneur based on the language model developed an AI reading APP, a week of users exceeded 10,000." And "selling prompts for sorting, using various AI tools to assist writing, drawing and making product drawings to receive business orders, etc. have also become new wealth codes."
Previously, as a film and television producer, Ah Bin never imagined being able to enter the artificial intelligence industry, but under the initiative of HustleGPT initiators, Ah Bin participated in the challenge of "giving AI $100 to make money as a business partner", and now, "With the help of the big model, I took 5 days to build a website page, which was unimaginable in the past." ”
AI models may have the power to subvert and reinvent and reshape. Daniel Zhang said at the summit, "All industries, software and services deserve to be redone based on new artificial intelligence technologies, which will not only bring innovative customer experience, but also change the paradigm of our production, work and life." ”
Robin Li, founder, chairman and CEO of Baidu, also pointed out, "In the era of big models, the biggest entrepreneurial opportunity is on the application side. In the era of mobile Internet, there are actually few operating systems, and the most successful are WeChat, Douyin, Taobao and other applications. In the next 10 years, opportunities worth 10 times more to them may be born. ”
From this point of view, AI big models have great prospects.
03
ChatGPT replication is not easy
However, whether it is the opening of large factories, or small factories and entrepreneurs drinking soup, there are many difficulties on the road to domestic AI large models.
First of all, the big manufacturers of AI models are difficult to avoid the shadow of "ChatGPT", and because of insufficient data model training, it has attracted many doubts.
"Training large models requires a lot of high-quality data. Although domestic large manufacturers, such as Huawei, Ali, Baidu, etc. have already had a large number of data resources in the field of artificial intelligence, such as Wen Xin Yiyan's official model parameters are 260 billion, higher than ChatGPT3's 175 billion, but the same NLP large model, ChatGPT3.5 has accumulated more than trillions of human language datasets, and Wenxin Yiyan's initial text data is only 330 billion. Ah Bin pointed out.
In addition to data accumulation, data collation, model training, etc. require a lot of time, "The AI models of domestic manufacturers want to catch up with GPT4, and the time cannot be crossed." Ah Bin also said.
"Nowadays, the biggest challenge for large manufacturers to make AI models is actually fierce competition and the development of user systems. At present, for AI large models, not only is the competition fierce, but the user's taste has obviously been fed by ChatGPT, and the big manufacturers will also face higher user requirements. Zhang Shule added.
Taking Baidu's Wenxin Yiyan model as an example, not only was it ridiculed by the group on March 16, the stock price fell by more than 9%, but in late March, the controversy about "Wen Xin Yiyan" drawing is "shell", "painting skin" and "fake", some netizens complained that Wen Xin drew "mouse and bus", and the result was "mouse and bus", because its English translation is "mouse" and "bus", and questioned its reference to a foreign open source AI tool for drawing.
Source/Visual China
In this regard, Baidu said in a statement on March 23 that "Wenxin's words are completely Baidu's self-research, and Wenshengtu's ability comes from Wenxin's cross-modal large model ERNIE-ViLG", and said that at present, Wenxin's words still need to be "further tuned and iterated through training".
Dabayashi also pointed out that behind the ridicule and controversy about Wen Xin's statement of "artificial intellectual disability" is the disappointment of users compared with ChatGPT, "It can be said that although the domestic AI big model now has a large amount of Chinese data, there is still a lot of room for improvement in the understanding and use of Chinese data, and even the generation."
However, "even ChatGPT has problems such as the timing and insufficient coverage of 'learning materials', deep learning needs to digest a large amount of data, and the insufficient training of domestic AI large model data is not a temporary solution." Zhang Shule pointed out that the current domestic and foreign AI big models are just a "editor" who performs well when playing normally, and does not have real thinking and innovation capabilities, "At present, it is still the learning time of AI large models, and domestic AI large models still have opportunities." ”
Secondly, in addition to data, the AI models of large factories are also constrained by insufficient computing power.
"The computing power of domestic companies is too scarce." Dabayashi said that the AI large model has extremely high requirements for computing power, and the more application scenarios, the demand for computing power can increase by 10-100 times, taking ChatGPT as an example, the data center for training computing power includes servers, etc., and the investment scale is at least 10 billion yuan.
"If you want to make China's ChatGPT, the lack of industrial supply and policy support for computing power chips is also a piece of paper." Ding added.
In addition, with the influx of large factories and star entrepreneurs, AI models have obviously become the new "traffic password" of enterprises, but it is inevitable that the fish will be mixed and the mud and sand will fall.
Jiang Han said that the current head players of AI large models, among which the advantages of Baidu and Ali are clearer, and because Baidu Wenxin released earlier, the first-mover advantage is more obvious.
"Including iFLYTEK and SenseTime, they have accumulated a lot of artificial intelligence before ChatGPT became popular, and they are more likely to make a large model of ChatGPT-like natural language generation in the true sense, but some companies still use the concept of hype and cutting leeks, so we should pay attention to distinction." Ding Daoshi said bluntly.
What's more, large manufacturers have accelerated the development of AI large models, and the potential risks of artificial intelligence technology have also triggered many discussions.
Recently, thousands of top people, including Musk and Apple co-founder Steve Wozniak, have publicly called on AI labs to suspend training more powerful models for 6 months, after OpenAI data leaks also caused privacy disputes, "user data security and privacy protection, as well as copyright and intellectual property disputes caused by AI creation, are all problems that cannot be bypassed by the development of AI large models." Ah Bin added.
According to the Ministry of Industry and Information Technology, the scale of the mainland computing power industry has grown rapidly in recent years, with an annual growth rate of nearly 30%, ranking second in the world in terms of computing power. By the end of last year, the total scale of computing power in mainland China reached 180 billion billion billion floating point operations per second, and the scale of the core industry of computing power reached 1.8 trillion yuan.
The Measures for the Administration of Generative Artificial Intelligence Services (Draft for Comments) also clarifies the attitude of support and encouragement for the generative AI industry.
From the general direction, the future is bright, but back to the road under the feet, there are many difficulties, whether the AI big model can roll up a new world and a new future, we have to be cautious.
Bibliography:
"Artificial Intelligence Industry Special Report: Big Model Applications Blossom, AI Development Enters a New Era", source: Future Think Tank.
*The title image and some of the accompanying pictures are from Visual China.
*In the text, Ah Bin, Daelim and William are pseudonyms.
*Disclaimer: Under no circumstances does the information in this article or the opinions expressed constitute investment advice to anyone.