Although the scuffle between the world's mainstream large models is still ongoing, in some regions, the ecological niche has been initially established. In China, the war of 100 models has come to an end, but due to the slower than expected landing, it is still too early to talk about the market structure.
Wen Niu Hui. Edited by Zhao Yanqiu.
The competition between the world's mainstream models shows no signs of slowing down. At the beginning of 2024, the large model market will once again be stirred by major releases from companies such as OpenAI. Although the war is hot, in some regions, the ecological niche of the large model has been initially established, and it is rapidly transmitted to the downstream, including application development, mobile phones, and other enterprises, have also joined the war in the last mile of the large model.
In China, although the war of 100 models has ended, due to challenges such as the performance of the basic large model, the landing is slower than expected, and it is still too early to talk about the market structure.
In 2024, the State-owned Assets Supervision and Administration Commission (SASAC) will hold a special promotion meeting on enterprise artificial intelligence on the first working day after the Spring Festival. E Weinan, academician of the Chinese Academy of Sciences, leading artificial intelligence companies, Huawei, as well as State Grid, China Mobile and other leading enterprises, made exchange speeches, and also mentioned OpenAI's latest large model SORA. The meeting emphasizedAccelerate the construction of a number of intelligent computing centers, carry out AI+ special actions, and build a number of industrial multi-modal high-quality datasetsand other specific initiatives.
We have sorted out the wars between some mainstream large models in the world from 2023 to the present, and got a glimpse of the next trend and pattern of large models.
Mistral: "OpenAI" in Europe is gaining momentum
In 2023, some domestic industry insiders discussedAre there any opportunities for Europe to develop its own large-scale model industrial system? So far, although Europe does not have a complete industrial system like the United States, Europe's big model is catching up.
On February 26, 2024, Mistral AI, known as the "European version of OpenAI", released a new large model, Mistral LargeThe official introduction said that the performance is second only to GPT4
However, Mistral Large is no longer open sourceThis also means that mistral has a higher demand for commercialization, after all, open source money is not easy to earn.
The competition among cloud vendors for Mistral is also extremely fierce for advanced and regionally distinctive large models. In the news on the morning of February 26, Mistral Large was also only available through its own locationSelf-built infrastructure in Europe, as well as Microsoft Azure cloud to provide business servicesIn the evening, Amazon Web Services officially announced that the Mistral AI model was about to land on Amazon. Moreover, everyone noticed that this is Chinese ***It indicates that Mistral AI's services should also cover China.
Microsoft, the big winner of the big model, and the cooperation with Mistral AI have also attracted attention. Microsoft is a major investor and business partner in OpenAI, howeverThe relationship between Microsoft and OpenAI is also friend and foe, and Microsoft is also seeking benefits in cooperation with other large-scale model companies。Microsoft, for example, has a partnership with Mistral AI. These delicate relationships will also affect the market competition pattern.
Founded in May 2023, Mistral AI is less than a year old and is based in France and founded by former scientists from Deepmind and Meta. In September 2023, they launched oneOpen-source large language models, although the parameter is only 7b, it outperforms LLAMA2's 13b model in all benchmarks.
After the release of Mistral Large, some netizens ran to the social network of Meta Chief Scientist Yang Likun to ask when llama3 would be launched, and Yang Likun said that he hoped that netizens would be more patient.
OpenAI: Unbeatable, but increasingly opaque
10 days before Mistral AI launched a new large model, on February 16, OpenAI coldly launched a new Wensheng ** large model Sora, with amazing results, which once again sparked heated discussions in the industry. Some industry thinksIt also further widens the gap between OpenAI and its peers. So far, on a global scale, no company has said what time period and how to catch up with SORA's level.
In the past year, OpenAI has made new moves almost every month, from product technology, to business monetization, to cooperation and investment.
In 2023,OpenAI's revenue exceeded $1.6 billion, becoming the fastest-growing product company in history. It is currently the main oneThere are three monetization modelsFirst, software developers call APIs and pay them for traffic; the second is ChatGPT Plus membership, monthly subscription fee; The third is to share from Microsoft's revenue from selling OpenAI models, but it is lower than its own direct sales revenue.
OpenAI also launched the accelerator Converge at the end of 2022 to build its own industrial territory, investing in companies involving chip manufacturing, AI application assistants, robots, etc.
But as OpenAI grows bigger, it becomes more and more opaque, its product information is becoming more and more closed. And in November last year, CEO Altman Sam was abruptly removed by the board of directors and returned to the CEO role a few days later. This turmoil has presented OpenAI's internal ** to the world. Many analysts believe that this is a battle of ideas, and OpenAI has not given any official explanation, making people doubt the further development of AI.
Ultraman's words and deeds are also worth playing. Although he himself has repeatedly stated that he needs to establish a regulatory framework, in that internal corporate conflict, Altman was considered a "radical", and he disagreed with the "establishment" on the contradictions between development, commercialization and security, AI for good and the profitability of AI companies.
At the beginning of this year, it was rumored in the industry that Ultraman planned to raise funds7 trillion dollarsConstruction of energy and chip factories, when Intel CEO Gelsinger asked Altman himself for confirmation the other dayUltraman denied itBut then he said that AI chipsets alone require a lot of global investment, beyond imagination," he saidWe don't have a specific number yet”。In addition, there are rumors that OpenAI will launch GPT5 in the second half of this year.
Judging from a series of actions by OpenAI and Ultraman, it has the ambition to become a giant in the AI era, and its goals far exceed Microsoft, Apple and Google.
Microsoft: Old-fashioned strategy
In addition to OpenAI, in this wave of large-scale model melee, Microsoft's strategy has to be said to be forward-looking and sophisticatedThe action is similar to the "Thousand-armed Guanyin", and the influence goes straight in all directions.
When ChatGPT caused a sensation around the world, Microsoft quickly decided to invest an additional $10 billion in OpenAI and won 49% of the equity; On the other hand, they quickly applied GPT to their own applications.
The first product officially announced by Microsoft is the search product Bing, so that Bing, which accounts for only 3% of the world's total, has a surge of millions of new registered users in a weekHit Google's stock price**
In addition to Microsoft's close cooperation with OpenAI, its Azure cloud also serves large models such as Meta's Llama, France's Mistral, Canada's Cohere, and the Middle East's G42. The industry believes that Microsoft's move, in addition to establishing a complete AI ecosystem, also avoids the suspicion of monopoly.
Recently, according to the information of the official account of Xiaodiantable, Microsoft has reached a cooperation with the domestic office IT ** business Yidianyun to jointly launch Yizhihui, which allows Chinese enterprises to use Openal's GPT capabilities.
Microsoft's foresight and insight, which seems to have returned to its peak, has also caused the industry to be very anxious. When it acquired GitHub in 2021, it first introduced the concept of "Copilot", which means "AI assistant". It has to be said that this is as forward-looking as Microsoft's investment in OpenAI in 2019. At Microsoft's blockbuster global technology conference Ignite in November 2023, Microsoft CEO Nadella saidMicrosoft has become a Copilot (AI assistant) company
It is worth noting that at the same time as joining forces to launch a "bombing" war against Google,Microsoft's relationship with OpenAI is not monolithic. For example, in August 2023,OpenAI launches ChatGPT Enterprise subscription service, directly replacing some of Microsoft's cloud AI business services, to Microsoft's displeasure. In the OpenAI turmoil in November, although Microsoft and CEO Nadella quickly stabilized the situation, many users switched from directly calling OpenAI's services to using the Microsoft cloud platform. In the future, it is not excluded that the relationship between the two will be more competitive.
Google: Under red alert, continuous frustration
In December 2022, a month after the release of ChatGPT, Google CEO Sundar Pichai issued a "code red" within the company, expressing that Google was in an urgent crisis. But more than a year has passed, and Google, the vane of the artificial intelligence industry, is still being pressed by OpenAI.
Now, there is a lot of in-depth analysis in the market - why does Google "get up early in the morning and catch up late"? For example, the "History of the Disappearance of Google Romance" by Lao Wan Story Club.
Google has indeed been "stunned" by the cooperation of OpenAI and Microsoft, and has been rushing to respond in the past year, urgently launching the ChatGPT-like chatgpt-like conversation service Bard in February 2023, the Palm 2 benchmarking GPT4 in May, and the Wensheng diagram model benchmarking DALL·E 3 in December. Although they were launched one step later, the effect was not as amazing as OpenAI's products, and even like Bard in the release demo, the reply was wrong, and Google's market value evaporated by more than $100 billion overnight.
Interestingly, in the past, Google's new product releases were mostly concentrated in May and June every yearBut in December 2023, Google went against the norm and began to release large models. Unfortunately, these models didn't make much of a splash. On February 15, 2024, Google just released the multimodal large model Gemini 15 Pro, the hind foot was silently drowned out by the Sora thrown by Openal.
It is worth noting that Google, which has used an open source strategy to defeat Microsoft's mobile ecosystem, has taken a closed-source road on a large model.
In 2023, Zuckerberg's Meta will launch the LLAMA model and open source it, which will have a profound impact on the global model ecology and artificial intelligence entrepreneurship.
At present, in the overseas market, OpenAI, Microsoft and Meta are in a leading position in terms of technology, business services, and the impact on enterprises and developers.
: The domestic leader, in the midst of a difficult catch-up
Objectively speaking, in the domestic market in 2023, the system and rules of punching on large models are ahead of other domestic Internet companies, which can be seen in its past accumulation in artificial intelligence.
From the aspects of large models, large model tool platforms, large models to their own applications, and large model ecology, including developer incubation, a systematic approach has been carried out. At present, the improvement of the performance of basic large models and the mining of landing application scenarios are the focus and difficulties of domestic large model enterprises. For OpenAI's new SORA launch in February, no relevant plans have been released.
Huawei: It has taken the controversial but feasible path of monetization
In the AI industry, there is controversy over whether the HUAWEI CLOUD Pangu model is a large model. However, some industry players believe that, in essence, HUAWEI CLOUD's large-scale model path is a business path that can be implemented, realized, and does not burn money.
In July 2023, HUAWEI CLOUD Pangu Model 30 was released, proposing "not to write poems, but to do things", which sparked heated discussions; In September, at the cloud summit, they responded to "compose poetry and do things", but the two sides have very different ways of playing on the battlefield of large models, and large models do not have clear industry attributes.
Since then, Huawei has launched strategic cooperation in various industries, explored the implementation of applications, and won large orders in industries such as meteorology, medicine, mining, and government affairs. Similarly, the HUAWEI CLOUD Pangu model faces the challenge of improving the performance of the basic model.
Unlike other companies in the industry, Huawei also has computing and network technologies in the field of large models. Ren Zhengfei talked about "selling shovels in the new gold rush era" in his internal speech, and proposed a special area for 100 modes and 100 states.
In addition to Huawei, there are also major manufacturers in the Chinese market, such as Zhipu, Alibaba, iFLYTEK, Tencent, ByteDance, and Inspur.
Although the domestic 100-model war has come to an end, the ecological niche has not yet been fully established. Due to challenges such as the performance of the basic large model, the application landing in 2023 is not as expected, and it is still too early to talk about the market structure.
In the spring of 2024, there will be quite a few progress updates. With the improvement of the performance of the basic large model, as well as the promotion of the first and central state-owned enterprises, the progress of the implementation is expected to accelerate this year. 2024 is still a year of fierce competition among domestic large-scale model enterprises.