In the year of the explosion of large-scale model applications, who can take the lead in breaking through?
Author |Zhang Kaijing.
Edit |Fun about business.
At the end of the year, large model manufacturers showed another wave of "muscles".
On February 1, ByteDance, which had been in a "silent" state in the field of large models, finally made a new move, and launched the "Coze Button" AI bot development platform；Different from the form of chatbots in the past, "buttons" are more like GPTS released by Open AI in November 2023, which allows users to create a personalized version of the bot through chatting, calling plug-ins, etc., to achieve "0**" development.
Screenshot of Weibo.
In addition to the byte field, Orion Star, founded by Cheetah Mobile CEO Fu Sheng, also released its own large model Orion-14B on January 21；Fu Sheng emphasized that in enterprise application scenarios, the Orion Star model can achieve a model effect of 100 billion parameters when combined with enterprise private data and applications.
And such as iFLYTEK, 360, etcManufacturers that grabbed large models for the first time in 2023 are now iterating rapidly；The former recently released the Spark large model v35, the latter launched the large model search app "360 AI Search".
Not so long ago,Mobile phone manufacturers have also poured into the large-scale model track。On January 10, Honor released its self-developed 7 billion parameter end-side AI large model "Magic Big Model", and since then, Huawei, Xiaomi, OPPO, vivo, and Honor five domestic mainstream mobile phone manufacturers have gathered in the field of large models. At the same time, Apple across the ocean is also rumored to be testing the introduction of generative AI features in iOS 18.
Screenshot of Weibo.
The "100 Model War" does not seem to have died down with the passage of time, but has intensified. What kind of calculations are manufacturers making now? Will the large-scale model track in 2024 usher in new major changes?
ChatGPT didn't come out without warning.
OpenAI, which created it, was founded as early as 2015, and in 2018 it launched its first generative pre-trained model, GPT-1, which was also the original prototype of ChatGPT.
Prior to this, NLP models (natural language processing) in the industry were mainly trained on a large amount of annotated data for specific tasks, and their capabilities were limited.
GPT has made breakthroughs in natural language reasoning, question-answering tasks, and common-sense reasoning, for which it has also received a $1 billion investment from Microsoft. Subsequently, from GPT-1 to GPT-2 to GPT-3, GPT's capabilities continued to improve in just 2 years.
Screenshot of OpenAI's official website.
On November 27, 2022, with the release of GPT-35. The launch of ChatGPT, an artificial intelligence conversational chatbot, which has been popular in a small circle, has only been recognized by the public for the first time. Being able to chat, draw, write copywriting, and edit **, powerful functions have made ChatGPT quickly popular once it was released, with more than one million registered users in 5 days and 100 million monthly active users in two months, becoming the fastest-growing consumer application in history.
When the news reached China, one stone stirred up a thousand waves. How far has artificial intelligence come to this? Shocked, selling ChatGPT accounts on ** has even become a business. At the same time, bigwigs from all walks of life who saw huge business opportunities also came down one after another, stating that they would launch their own large models as soon as possible.
From March to September 2023, various major Internet companies will compete for the first place in this track. From the large language model "Wenxin Yiyan's intelligent brain model", to Alibaba's "Tongyi Qianwen" model, iFLYTEK's Spark model, and then to Tencent's hybrid yuan model, the three traditional domestic Internet giants "BAT" have all come to an end.
In addition, Huawei, JD.com, SenseTime, NetEase Youdao, Kunlun Wanwei (300418SZ) and so on have successively launched large-scale model products, and even the three major operators of China Mobile, Unicom, Telecom, Changhong and other home appliance manufacturers, Tsinghua, Fudan, Chinese Academy of Sciences and other scientific research institutes and universities, have released their own large models.
HUAWEI CLOUD AI***
At the Zhongguancun Forum in May 2023, experts said that the number of large models of China's "above 1 billion parameters" was still 79; According to GitHub's statistics, by the end of 2023, nearly 300 large language models have been released in China。The "100-model war" is no longer a lie.
However, compared with the original "Hundred Group War" in the takeaway industry, the "Hundred Model War" has a more demanding demand for funds.
According to NVIDIA's official information, in the training stage of the underlying model, it takes 34 days to train GPT-3 with 175 billion parameters and uses 1024 A100 GPU chips, and in order to maintain daily reasoning, OpenAI needs at least 3240,000 A100; Based on this calculation, ChatGPT's hardware cost alone exceeds $800 million
Fang Han, CEO of Kunlun Wanwei Group, said publicly"Without 2,000 A100 cards, the experiment would not have been possible." To this end, "AI concept stocks" such as Haitian AAC and Insai Group have even successively released fixed increase plans to raise funds for training large models.
Canned Gallery.
At this time, how to find the direction of application landing as soon as possible while narrowing the gap with ChatGPT and realize self-hematopoiesis has become a problem that every participant has to face in the "100 model war".
From a business perspective, the opportunities brought by large models can be summarized into three categories: cost reduction, efficiency improvement, expansion of original market demand, and creation of new market demand.
The large model's super human-computer dialogue and audio generation capabilities have not only made it widely used in traditional customer service scenarios, but also had a profound impact on games, film and television production, etc.
Alibaba, Meituan's first-class intelligent customer service algorithms, as well as China Mobile's "nine-day model", China Telecom's telechat model, etc., all belong to this kind of products, and the application scenarios are directly locked in intelligent customer service, smart government and other aspects. At the 2023 Asian Games, iFLYTEK and China Mobile jointly launched a 5G new call based on the Xinghuo model.
In terms of games, film and television production, director Lu Chuan once said in an interview, "Using AI to draw movie posters, the effect of 15 seconds is better than that of a professional poster company for a month." ”
Screenshot of Weibo.
The expansion of the original market demand is reflected in the upgrading of traditional services, which is also the most widely used field of large models.
Taking a traditional search engine as an example, after accessing Wenxin Yiyan, enter a question in the search box, and what will be given can no longer be a link, but a more certain answer. Based on this, applications such as maps, network disks, and libraries can be reconstructed by accessing large models.
Tencent, which has a large number of businesses, has also completed the test of accessing Tencent's hybrid model for a number of businesses and products such as Tencent Cloud, Tencent Advertising, Tencent Games, and Tencent Meeting, and has achieved initial results.
Screenshot of Tencent's mixed yuan official website.
In addition, in the traditional education, medical care, automotive and other fields, large models have also been widely used.
After being connected to the iFLYTEK Xinghuo model, iFLYTEK's learning machine has realized functions such as AI one-to-one assisted teaching, Chinese and English composition correction, and oral sparring. Launched the industrial-level medical industry model "Lingyi"; The empowerment of HUAWEI CLOUD's Pangu model has enabled the new M7 to be "far ahead" in the field of intelligent driving, with more than 100,000 units set to exceed two months.
In terms of creating new market demand, demand for AI super assistants and AI robots is also constantly being created. In the former, various general large models, including Wenxin Yiyan, Xunfei Xinghuo, Tongyi Qianwen, etc., have corresponding products, most of which can understand the user's language semantics, and have image understanding capabilities, and can help users complete tasks by calling software APIs and using a variety of tools; The latter has no less than 10 robot companies, including UBTECH, Dreame and Unitree, which have exhibited related products.
Screenshot of Tongyi's official website.
However, behind the prosperity, there are also hidden worries. It is not difficult to find that whether it is to improve efficiency or expand demand, the vast majority of applications on the market have similar functions.
Taking the AI learning machine as an example, in addition to iFLYTEK's related products, NetEase Youdao with access to the Ziyue large model, Good Future with access to MathGPT, homework help with access to the Galaxy large model, and 360 with access to Wenxin Yiyan and 360 Intelligent Brain have similar products on sale. In terms of functions, what they advertise is also similar, AI one-on-one tutoring, general AI homework assistant, virtual oral language coaching, etc., from the perspective of consumers, it is almost difficult to appreciate the difference.
Canned Gallery.
Although each company can come up with a bunch of ranking lists to argue, its own large model scores higher and is more capable; But when it comes to practical application, the difference of a few percent or even a few thousandths still makes people wonder: do we really need so many large models?
Although the large models in the market are dazzling, the industry has formed a certain consensus on the development trend of large models.
The founder, Robin Li, once said: ".The sign that mankind has entered the AI era is not the production of a lot of large models, but the production of a lot of AI native applications。Zhou Hongyi, the founder of 360, also talked about the development trend of large models at the beginning of this year2024 will be the year of large-scale model application scenarios, and there will be "killer applications".
This also means that the distance between the large model and the C-end user will get closer and closer.
Byte's "clasp" platform is an example of this. According to "Fun Solution Business", it has an infinitely expanding set of capabilities, and users can continuously strengthen the ability to customize bots by adding plug-ins; In addition, users can upload local files to the bot's knowledge base for learning; The created bots can also be deployed on different social platforms and applications.
Screenshot of the official website of the button.
This is tantamount to providing users with the opportunity to develop their own chatbots, so that more people can participate in the construction of the AI ecosystem.
At the same time, the large model is also undergoing the process of software and hardware integration and collaboration.
In this regard, smartphone manufacturers are undoubtedly the representatives of the industry. According to incomplete statistics from "Fun Solution Business", among domestic smartphones, Huawei Mate60 Pro, Xiaomi 14 Pro, Vivo X100 series, OPPO Find X7 series, Honor Magic6 series and other mobile phones have been equipped with large models.
In addition to making AI assistants more intelligent, these large models also have a key application area of mobile phone photo albums. It turns out that if you want to eliminate other tourists in **, you can only use ps, and test your skills and techniques; Now you can directly apply the AI erasure function, and you can do it in one step and there are almost no flaws.
Similarly, Meitu (1357HK) self-developed AI visual model "Fantasy Intelligence". It has partnered with Samsung to allow users to experience AI image editing on the Galaxy S24 series phones. Not only can it be "smart P map", but it can also further generate "AI painting style" through the ** given by the user.
Screenshot of Weibo.
The scenario that is benchmarked against a mobile phone is a PC (computer).。In January, Lenovo has released AIPC products, which have stronger computing power support capabilities, more intelligent human-computer interaction, and a more open application ecology after embedding large models.
And Kingsoft Office (688111SH) is a software product that can be applied on the PC platform. It focuses on intelligent documents, which can help users easily create texts and PPTs required for office, and can also independently digest documents such as PDFs and answer questions about these materials.
In addition, education is also an important application scenario. Manufacturers such as iFLYTEK, Good Future, Homework Help, and NetEase Youdao have integrated large models into AI learning machines, and the sales of educational learning tablets will skyrocket in 2023. And based on the iteration of the product, the ** of the learning tablet is still rising.
Canned Gallery.
Zhang Xiaorong, president of the Deepin Science and Technology Research Institute, believesIn the future, large models may develop in the direction of specialization, personalization, and low threshold. The functions of the model will be more granular and optimized for specific domains or specific needs; At the same time, by providing a more user-friendly interface and a more convenient interface, the difficulty of use is reduced, and more people can participate in the development and research based on large models.
And due to the limitation of computing power, large models may be more deployed in the cloud and at the edge. This reduces the consumption of computing and storage resources, and improves the responsiveness and availability of the model.
But whatever the trend, it has to be combined with the actual cost; Otherwise, it is obviously unsustainable to just blindly invest.
360 (601360.)SH) in the first half of 2023, although the emerging business "360 Intelligent Brain" generated nearly 20 million yuan in revenue, it was in 9Of the total revenue of 100 million yuan, it accounts for only 21%。
Canned Gallery.
iFLYTEK (002230.)SZ) is expected to deduct non-net profit in 2023 by 71%-81% year-on-year, mainly due to the company's increased investment in the research and development of cognitive large models on an independent and controllable platform.
In addition, how to enable users to better understand the decision-making process and results of large models and improve their trust is also a key issue.
The domestic large-scale model industry is in full swing, and the United States is even more so. According to the "Chinese Artificial Intelligence Large Model Map Research Report".Among the large models released in the world, China and the United States account for nearly 80% of the large models。As early as May 2023, the number of basic large models with more than 1 billion parameters in the United States has exceeded 100.
In addition to the well-known ChatGPT, representative general model companies in the United States include Anthropic, Cohere, and Google.
Among them, Anthropic is known as the "fierce rival of OpenAI". Its developed chatbot Claude can summarize about 7At 50,000 words, it's better than ChatGPT for long conversations and content, in-depth analysis of large documents, and faster average response times.
Screenshot of Weibo.
Cohere is characterized by its differentiated positioning. Unlike OpenAI, it has firmly chosen the TOB route, providing flexible storage and data privacy protection paths, emphasizing security, privacy and customized services.
As for Google, the latest development is the launch of the AI model Gemini, which is characterized by multimodal processing and the ability to understand complex logic. In the industry-standard MMLU (Multitasking Language Understanding) benchmark, Gemini is the only AI model that outperforms the results of human expert tests.
Screenshot of Weibo.
Wang Peng, a researcher at the Beijing Academy of Social Sciences, believesThe differences between China and the United States are mainly reflected in three aspects: the level of financing, the level of development of the basic model and the level of development of the application layer.
According to incomplete statistics, in the first half of 2023,In the AIGC primary market in the United States, Silicon Valley has raised a total of about $14 billion in the field of artificial intelligence, accounting for 55% of the world's total financing, the average round of financing amount is 3$300 million. In the same period, the domestic artificial intelligence field was much more cautious, with the number of investment events falling by 49% year-on-year, involving a total amount of 617.4 billion yuan, down 62% year-on-year.
In terms of the development level of basic large models, there are still problems such as lack of total data, lack of computing resources, and limited scenario penetration rate of domestic large models. After all, in terms of the amount of public data, English data itself dominates, and the United States is still taking various ways to restrict China's access to the core resources of computing power.
As for the application layer, China is also in a state of following; Among them, it lags behind the United States in the office, finance and medical fields.
Fun about business.
In response to the fact that many domestic manufacturers have claimed that their large models have surpassed GPT-4, Zhang Xiaorong believes: ".Theoretically, some manufacturers may be locally ahead of GPT4, but considering the investment of both parties in algorithms, computing power and data resources, it is relatively unlikely that the domestic model will comprehensively surpass GPT4
In his view, it is necessary to face up to the gap between domestic and foreign large models, which involves technology, talent, capital and other factors.
The good news is that China has a large market scale and rich application scenarios, which provides a broad space and conditions for the landing and application of large models. And the more data and scenes, the more practical the large model can be. This gives China a chance to catch up with and surpass the United States even though it is slightly inferior to the United States in terms of underlying R&D technology.
However, whether it is China or the United States, there are still many problems to be solved on the track of large models. The most typical problems are the lack of credibility, stability, and security of the output results.
It's going to be a long-term learning process for both humans and AI.
ByteDance