The big model is undoubtedly the hottest word of the year. Looking at 2023, from the "Hundred Model Battle" competition to the "Hundred Flowers Blooming" of the application layer ecosystem, the AIGC industry has ushered in an explosive period. Data shows that as of October, China has released 238 large models, a threefold increase from 79 in June. Since the emergence of the Big Model technology boom, from venture capital to capital markets, from the B-end to the C-end, AI has been mentioned everywhere.
After the popularity of big models for half a year, there has been a debate in the industry: who is the future between big models being universal and vertical? At the Geek Park Innovation Conference 2024 on December 16, Robin Lee said: "Only the native application of Volume AI is valuable, and the progress of the big model is not an opportunity for most people".
Standing at position C, the highlight moment of the big model of the text heart
The emergence of ChatGPT has made "big models" a hot topic in the global technology arena, and has also made it the absolute "C position" of this year's World Artificial Intelligence Conference.
Among numerous products, the dynamics of Baidu's Wenxin Big Model are undoubtedly one of the most closely watched. On March 16 this year, Robin Lee, the founder, chairman and CEO of Baidu, officially released ERNIE Bot, a pre training generative big language model. Robin Lee, founder, chairman and CEO of Baidu, revealed that the first version of Wenxin model in 2019 had been released, and since then, new versions have been iterated every year.
When people's evaluation of the application of AI big model mostly stayed in the impression of "an easy-to-use intelligent software", Robin Lee took the lead in proposing it in China. This AI wave can be compared to the "fourth scientific and technological revolution".
On June 26th, during the Nishan Dialogue on Digital Civilization at the Internet Conference, he reiterated that "the big model is a major technological change that affects human development, an engine that drives global economic growth, and a significant strategic opportunity that cannot be missed."
Another business tycoon who shares the same opinion is Bill Gates from across the ocean. "Whether you want it or not, this will lead to a new industrial revolution, with significance beyond the invention of the internet and iPhone."
Baidu ERNIE Bot's iterative optimization has also been accelerating. The Wenxin Big Model 3.5 released in June has greatly improved training speed and inference speed. Only four months later, it was replaced by the fully upgraded Wenxin Model 4.0. Robin Lee said that its comprehensive ability was "no inferior to GPT-4".
At the end of August, ERNIE Bot's big model became the first batch of AIGC products that were open to the public, an experience and a touch. On the day when it was open to the public on August 31, Wenxin answered a total of 33.42 million questions from netizens in one word, ranking first in each app store.
"This marks a significant breakthrough for Chinese industry leaders and a milestone in their rapidly developing artificial intelligence industry," Bloomberg wrote in a report.
Many foreign media regard ERNIE Bot as the leader of China's big model. "ERNIE Bot has aroused the greatest excitement", The Economist wrote in its report on September 3.
In the list of top 100 global AI figures released by Time magazine, Robin Lee was rated as the global AI leader, and was listed alongside Mask, Huang Renxun, OpenAI CEO Sam Altman and others.
Baidu's big model has also been recognized by society, and the key technology of the Wenxin big model has also won the Leading Technology Award at the 2023 World Internet Conference.
The large model is biased, and the future is the era of native applications
Accustomed to the food delivery and transportation battles of the Hundred Regiments, and returning to the Hundred Model Wars at the beginning of this year, it seems like another miraculous Internet battle of financial and endurance.
Some scholars believe that the essence of AI in the era of big models is not much different from that of the previous era of deep learning, both of which are patching up the gaps in the Internet. But on December 16, at the Geek Park Innovation Conference 2024, Robin Lee, Baidu's founder, said frankly: "Only the native application of AI is valuable, and the progress of the big model is not an opportunity for most people".
As early as March 2023, Robin Lee reminded the industry to "not build wheels repeatedly", and said that application is the real opportunity of the big model. But countless amounts of money are pouring into the AI big model race. For many companies, instead of thinking it out and doing it slowly, it's better to simply and roughly say it again, at least to boost stock prices or valuations.
In Robin Lee's opinion, the big model track in China is actually full of "traps":
Firstly, the big model is a universal model. A large model can serve hundreds of industries, so it conforms to the winner take all rule, and its development dividends are difficult to spread to too many people and too many enterprises. The fundamental difference between big models and previous AI waves lies in their universality, namely the emergence of intelligence. This means that large models naturally possess the Matthew effect. In terms of final efficiency, those with larger scale, earlier launch, and stronger company strength undoubtedly have a first mover advantage, and only a few large models will eventually be able to compete. Just like mobile operating systems, having 1-3 top companies in a market is sufficient. And the number of large models in China has already reached over a hundred levels, most of which are destined to be cannon fodder.
Secondly, like operating systems, large models themselves do not possess any ability to create value and must have applications to monetize their potential. As early as at the Baidu World Conference in October this year, Robin Lee said: "Without the rich AI native application ecology built on the basic model, the big model is worthless."
Since the 1970s, AI has experienced multiple challenges of becoming a hot topic and cooling off. The most recent one came in 2016, when Alpha Dog defeated the world Go champion, refocusing the world on artificial intelligence. But just a few years later, the financing events and scale in the AI field plummeted.
Robin Lee attributed AI's fever reduction to its application: "At last, we found that this thing is useless". For example, in the last AI trend, a large number of AI companies emerged in China, among which the AI Four Dragons were among the leaders. However, their income is somewhat meager compared to their valuation. The scale of applications often focuses only on a few areas such as monitoring and facial recognition.
Thirdly, historically, the total potential of the application layer has been far greater than that of the underlying layer, and it is also more friendly to most companies. On the application layer, there is currently no killer level application at home and abroad. The most downloaded application may be ChatGPT itself, which has a daily activity of over 100 million. It is generally believed at home and abroad that the big model will restructure all industries, but up to now, generative AI has no "WeChat" or "Tiktok" of its own. For large models, if they do not have a good landing scenario, the models that cost billions or even billions can only be "applied" in score running software, becoming an image project that satisfies vanity.
Under the constant appeal of Robin Lee, in recent months, more AI start-ups have also begun to focus on the application level. The transformation of these enterprises will make the ecosystem of the big model more prosperous, bring more users, form a positive cycle, and ultimately benefit the entire industry.
The imagination space for AI native applications is huge, and the market is looking forward to phenomenon level applications
How profitable is the combination of AI and native applications? Robin Lee gave an example in the conversation at the Geek Park Innovation Conference 2024: Microsoft combined Office with AI and launched the product Microsoft 365 Copilot, whose annual revenue reached $5 billion. According to Sam Altman, CEO of Open AI, disclosed in October this year, the company's annual revenue is currently about $1.3 billion. As the world's best big model company, OpenAI's revenue cannot keep up with a subsidiary product of Microsoft, as evidenced by the monetization ability of AI native applications.
Robin Lee also believes that small factories also have great opportunities in AI. Even a company with a large scale cannot cover all industries, and small and medium-sized enterprises can leverage their flexible advantages to create large-scale applications. As Robin Lee said, "(in the era of big model) start-ups may be able to produce three or five Super Apps, or hundreds or thousands of very valuable vertical applications, which is also very likely."
At present, there are also many AI native applications with huge monetization potential in China, and their imagination space is not smaller than that of Wenxin Big Model or GPT itself. At the recent Baidu World Conference, Baidu released over ten native AI applications, a considerable portion of which have already demonstrated commercial potential. Taking "Baidu New Library" as an example. With the support of the big model, "New Library" has features such as AI writing and AI PPT making. These are all essential needs for domestic users, and their commercial payment models have been validated by Kingsoft, Microsoft, and Baidu themselves, with the potential to create huge revenue.
Having applications alone is not enough, there must also be a complete ecosystem to serve them. Baidu launched its own AI native application store at this year's World Congress, solving the distribution problem of AI native applications. In addition to the previous set of infrastructure such as Feijiang, Baidu AI Cloud Qianfan Model, ERNIE Bot, etc., domestic AI native applications already have an ecology similar to mobile Internet. Developers can develop AI native applications at a lower cost and find their own distribution channels.
More than half a year later, the development of the industry is verifying Robin Lee's view. Standing in the present, only when more and more people realize the importance of native applications for big models and reproduce the combat effectiveness of the Chinese army - just like the prosperity of apps in the mobile internet era, can the AI big model industry enter the true and potentially changing trend of the entire era.
Silently climbing for 10 years, Baidu's big model has climbed to the top like this
Behind the Baidu Wenxin Big Model lies the determination of Baidu All In AI and an endless pool of AI talents.
The seeds of the big model had already been planted 10 years ago. In 2013, a small team of three people who had only published a 9-page paper was auctioned by a Sino US technology company, and the final bid exceeded 40 million US dollars.
In the winter of 2012, Professor Jeff Hinton of the University of Toronto, along with his two students, published a 9-page paper titled "ImageNet Classification with Deep Convolutional Neural Networks". They established the neural network algorithm AlexNet and won the championship in the ImageNet image recognition competition with an accuracy of 84%. The training cost was very low, using only four Nvidia GPUs.
Robin Lee keenly found the value of this research, and immediately arranged Tang Hesong, the vice president in charge of Baidu's strategic investment at that time, to contact Sinton and send an olive branch to invite him to join Baidu. This move has driven two other tech giants, followed closely by invitations from Google and Microsoft. Due to various reasons such as not wanting to cross the ocean at an advanced age, Xinton ultimately decided to join Google and pass by Baidu. In the end, Google acquired a company that had been established for less than a month, had no products, no revenue, only three employees, and a few papers for a price of $44 million.
In this auction, Robin Lee strengthened his determination to invest in in-depth learning and accelerate the deployment of AI technology. In January 2013, Robin Lee led Baidu to establish the world's first in-depth learning research institute and personally served as the president. He began to recruit well-known talents such as Wu Enda, Xu Wei, and Zhang Tong from around the world, many of whom personally visited and received them.
In the following years, despite constant hot topics, Baidu has always adhered to the ALL in AI strategy. Robin Lee made a speech at the 2020 World Conference on Artificial Intelligence: "My belief in AI has not changed at all". In that speech, Robin Lee gave his own insistence. A series of AI underlying technologies and products, such as PaddlePaddle, Kunlun Core and Smart Cloud, have been born in recent years, gradually forming a four tier architecture of Baidu AI "chip framework model application".
The Wenxin model, which stood in the spotlight this year, was actually first released in 2019, 16 months before ChatGPT's basic model GPT-3 was released.
In July 2019, the Wenxin model was upgraded to 2.0; In December 2021, Baidu launched a 260 billion parameter knowledge enhancement model based on the Wenxin model; In May 2022, 11 industry models based on Wenxin were released. By March 16, 2023, with the expectation of the whole industry, Robin Lee announced the release of ERNIE Bot.
"In a sense, we have prepared for this for many years. We began to invest in AI research more than a decade ago, and launched Wenxin big language model in 2019. Today's release is a continuation of the efforts of the past years." On the stage, Robin Lee is as warm and firm as ever.
Robin Lee once wrote in his letter to shareholders in 2020, "We are determined and patient, and we can endure extreme loneliness and hide under the stars and sea.". Now, the ERNIE Bot big model is carrying Baidu's dream and heading for the sea of stars.