Image source: Generated by Unbounded AI

In the history of artificial intelligence development, 2024 is destined to be a year worth remembering. This year, Chinese AI companies staged a thrilling battle to catch up and surpass. From following the pace of GPT-4 at the beginning of the year, to facing the impact of GPT-4o in the middle of the year, to going head-to-head with OpenAI's o1 series at the end of the year, Chinese companies have shown amazing execution and innovation potential.

This is an era where heroes are rising together. The "Six Little Dragons of AI" such as Zhipu AI, Dark Side of the Moon, MiniMax, Baichuan Intelligence, Zero One Thing, and Step Star are competing fiercely with Internet giants such as BAT. The favor of capital, the flow of talents, and technological breakthroughs are intertwined to form the most exciting chapter in this industry. When the bell rang at the end of the year, Chinese AI companies had already achieved true surpassing of international giants on some tracks.

In this silent race, every OpenAI innovation will trigger a wave of catching up. However, Chinese companies do not simply imitate, but gradually develop their own characteristics. From Zhipu's full-stack layout to the ultimate experience of Dark Side of the Moon, from MiniMax's pursuit of efficiency to DeepSeek's vertical breakthrough, every company is looking for a unique development path.

This is a dramatic story and a legend about innovation and persistence. Let us enter 2024 and relive this exciting competition that changed the pattern of China's AI industry...

January's competition among the heroes: the competition to start the year

The first snow of 2024 has not yet fallen in Beijing, and competition in the AI ​​industry has become fierce. In this industry, people have long been accustomed to the continuous pace of work - after all, OpenAI never gives its opponents a chance to breathe.

It has been nine months since the release of GPT-4, and this monument that stands in front of all companies is still daunting. Although Baidu claimed in October 2023 that Wenxinyiyan 4.0 was "not inferior" to GPT-4, the market and users are obviously still waiting for more challengers to appear.

January 16, this seemingly ordinary working day, is destined to be recorded in the history of China’s AI development. Zhipu AI, founded by Professor Tang Jie of the Department of Computer Science at Tsinghua University, launched a new generation of large base model GLM-4 on this day. When the test data came out, the entire industry was excited: in multiple authoritative evaluations such as MMLU and GSM8K, GLM-4 reached more than 90% of the level of GPT-4. What is even more remarkable is that in terms of Chinese alignment capabilities, GLM-4 has surpassed GPT-4 for the first time.

Coincidentally, MiniMax, known as one of the "Six Little Dragons of AI" in the industry, chose to show off its sword on the same day. they publishThe first MoE (hybrid expert model) large language model ABAB6 in China has been developed. This innovative hybrid expert architecture shows unique advantages in processing complex tasks and improving computing efficiency.

Just two weeks later, MiniMax took action again, launching abab-speech-01, a large speech model that can complete sound reproduction within 6 seconds. It used millions of hours of audio data training results to demonstrate Their strength in multimodality.

Since then, voice reproduction has also become a standard feature of domestic AI chatbots.

On January 22, a heavyweight joined the battle with his new work. Lingyiwanwu, led by Kai-fu Lee, released the Yi-VL multi-modal language model. This product achieved great results on the MMMU and CMMMU data sets as soon as it was launched, showing strong interdisciplinary knowledge understanding capabilities, making this low-key company In one fell swoop, it became the focus of the global open source community.

Just four days later, Dark Side of the Moon, the team known as the "most knowledgeable about AI", demonstrated their ultimate pursuit of product polishing. The major update of Kimi Chat not only comprehensively improves the basic model capabilities, but also achieves breakthroughs in user experience: the response speed is doubled, and network search, contextual learning, and multi-scenario capabilities are all available. In particular, the new voice input function in the mini program version makes this product far ahead in terms of user-friendliness.

The perfect presentation of these innovative functions has made this always low-key company quickly become a rising star in the industry. In the first half of 2024, Kimi became one of the most popular AI applications, attracting a large number of users who originally focused on using ChatGPT to switch to it. Its influence is so great that it has even set off an investment boom in "Kimi concept stocks" in the capital market.

Just as this magnificent January was coming to an end, on January 29, a piece of bombshell news exploded. Baichuan Intelligence, known in the industry as "China's most unique AI unicorn", has launched Baichuan 3 with over 100 billion parameters. This behemoth lived up to expectations, showing the best performance in medical evaluations, surpassing GPT-4 on Chinese tasks, and also performing well in the fields of natural language processing and code generation.

At this point, the wrestling in the first month of the new year is coming to an end. The five "AI Six Little Dragons" have all made their debut, each showing unique technical routes and product features, while the other Six Little Dragons are still gathering momentum and will make their debut in a few months.

However, amidst the cheers, a thought-provoking voice came from another technology giant.

On January 30, ByteDance CEO Liang Rubo rarely showed a strong sense of crisis at the annual all-staff meeting. This leader in the mobile Internet era seems to have lost its way in the AI ​​era. "GPT won't be discussed until 2023, and the industryThe large model startups that have done well were all founded from 2018 to 2021. "

These words were like a thunderbolt, echoing in the technology circle for a long time. When AI startups were chasing after each other, this Internet giant, which was once known for its acumen, finally realized that it was here. ByteDance’s self-reflection also reflects the entire Chinese science and technology industry. The embarrassment of the technology industry in the field of AI: The former innovators now have to find the courage to catch up again.

No one knows that this magnificent competition has just begun, and the Spring Festival is coming. On this occasion, a storm named Sora is brewing on the other side of the ocean...

Spring Festival Thunder: Shock caused by Sora

New Year's Eve in 2024 has not yet arrived, and Silicon Valley in the United States has sent a "special New Year gift" to the Chinese AI circle.

February 10, the first day of the Lunar New Year, O penAI quietly released the Vincent video product Sora on social media. This model, named after the Icelandic word "story", can convert text descriptions into high-quality videos of up to 60 seconds, supports multi-angle lens conversion, and can even accurately restore complex scenes. physical movement Regularity.

For a time, Chinese engineers who were supposed to be enjoying the reunion were forced to interrupt their vacations and stay in front of the computer to study Sora's demonstration video.

This sudden "Spring Festival gift." ", causing the entire Chinese AI circle to sink into Thinking. The emergence of a new track in video generation means another shift in technology paradigm. An anonymous CTO of an AI startup lamented: "While we are still studying how to catch up with GPT-4, OpenAI is already pioneering. Next track. "

March: Undercurrent

While companies were still digesting the impact of Sora, the capital market took the lead in reacting, with Dark Side of the Moon exceeding 10. The US$100 million in financing is the first shot in the spring financing war. Its luxurious lineup of investors-Alibaba, Sequoia China, Xiaohongshu, Meituan and Lisi Capital, etc. have joined in, demonstrating the capital market’s support for Chinese AI companies. firm Confidence.

On March 6, MiniMax also received US$600 million in financing led by Alibaba, and its valuation exceeded US$2.5 billion. The favor of capital seems to be preparing for the upcoming technology competition.

p>

Almost at the same time, Zero Yiwu released and open sourced the Yi-9B model. This model has excellent performance in code and mathematical capabilities, with its excellent Chinese and English processing capabilities and low-cost deployment.Advantages, quickly caused a sensation in the open source community.

On the product battlefield, Alibaba Cloud takes the lead. On March 14, Tongyi Qianwen demonstrated its eye-catching scene innovation capabilities. Its one-click analysis function for quickly reading tens of thousands of pages of documents, as well as the subsequent launch of the 6-hour ultra-long audio and video transcription capability, demonstrated its leadership in the vertical field. The results of deep cultivation.

A few days later, Dark Side of the Moon reported another good news. Kimi intelligent assistant increased the lossless context length to 2 million words. This achievement far exceeded OpenAI and set a new benchmark for the industry.

In this month of March, which is full of uncertainties, a piece of heavy news has excited the entire industry.

On March 19, NVIDIA released the new generation AI chip GB200 at the GTC 2024 conference. This "super chip" equipped with 208 billion transistors has redefined the limits of AI computing with its training computing power four times that of the H100 and lower energy consumption.

This news has given the Chinese AI circle both hope and pressure. The difficulty and cost of obtaining these top-level computing power may intensify industry differentiation.

On March 23, the most low-key "AI Six Little Dragons" step stars were finally unveiled. Their Step series of universal large model matrices - including the 100-billion-parameter language large model Step-1, the multi-modal large model Step-1V and the trillion-parameter MoE language large model Step-2 preview version were also unveiled.

At the end of the month, Alibaba once again demonstrated its technological innovation strength. The first open source MoE model Qwen1.5-MoE-A2.7B launched on March 29th can match the traditional model with 7 billion parameters with 2.7 billion parameters. It not only significantly reduces the training cost, but also explores a way to improve efficiency for the industry. new path. This is another important breakthrough for Alibaba in model architecture innovation after the launch of the Qwen 1.5 series before the Spring Festival.

From the Sora shock during the Spring Festival, to the release of Nvidia GB200 in March, to the hurried technical layout of various companies, this spring is destined to be unrest. The temporary "aphasia" of Chinese AI companies in the field of video generation and the urgent need for top-level computing power indicate that the industry structure may undergo drastic changes. With the arrival of April, a multi-dimensional competition involving technology, capital, and talents is quietly unfolding...

April: Hundreds of rivals compete

In April, industry competition has further intensified. On April 3, Alibaba Cloud launched the AI ​​programming tool "Tongyi Lingma", which supports more than 200 programming languages, demonstrating its ambitions in vertical fields. Three months later, this product was named one of the treasures of the museum by WAIC (World Artificial Intelligence Conference).

A week later, they launched 3The 2 billion parameter Tongyi Qianwen (Qwen1.5-32B), this model that balances performance, efficiency and memory usage, provides the industry with a more efficient and economical solution.

On April 17, MiniMax released the abab 6.5 series of models, including the abab 6.5 with trillions of parameters and the more efficient abab 6.5s. These two models support 200k tokens context length and can process nearly 30,000 words of text in 1 second, showing amazing processing efficiency. Of even more strategic significance is that on April 29, MiniMax announced a comprehensive upgrade of its open platform API services, with prices reduced by more than 50% and processing speed doubled, taking an important step towards inclusive AI.

On April 18, Kimi Intelligent Assistant received a major update, including improved model capabilities, new common phrase functions, voice input/reporting, and search and reference tracing, aiming to improve user experience and efficiency. The new version of Kimi performs better in logical reasoning, mathematical programming, Chinese-English translation, etc. It also supports personalized common phrase settings and voice interaction, making information acquisition more convenient. A new citation tracing function is added to the search results to ensure the rigor of answers.

On April 30, an important evaluation result attracted the attention of the industry: Baichuan Intelligent’s Baichuan 3 ranked first in the country with a score of 73.32 points in the SuperCLUE Chinese large model evaluation, surpassing GPT-4-Turbo and other 32 A large model. In particular, its outstanding performance in the two key dimensions of knowledge encyclopedia (82 points) and logical reasoning (68.60 points) allowed the industry to see the true strength of the Chinese model.

Fierce battle in May: Awakening of giants and price storm

May , competition on the artificial intelligence track has suddenly accelerated. On May 9, Alibaba Cloud released version 2.5 of Tongyi Qianwen and open sourced the 110 billion parameter model Qwen1.5-110B. This version has significantly improved understanding, logical reasoning, etc., and its Chinese ability leads the industry. In multiple benchmark evaluations, Qwen1.5-110B successfully surpassed Meta's Llama-3-70B model, marking the first time that Tongyi's large model has caught up with GPT-4 level.

On May 13, OpenAI released GPT-4o ("Omni"). This all-round model can not only analyze and generate text, images and sounds, but is twice as fast as GPT-4 Turbo and costs Only half of the latter. For a time, the technological gap that Chinese companies had finally narrowed was widened again.

Zero One Wanwu, led by Kai-Fu Lee, is also not to be outdone. On May 13, they released the 100-billion-parameter AI model Yi-Large and announced a dual-track strategy of open source and closed source, showing a clear and prudent business route.

May 15, MiniMax launched the AI ​​chat APP "Little Conch AI". This application based on multi-modal large models can quickly process large amounts of text, understand emotions, and supports multiple file formats and voice interaction.

In terms of industry governance, Zhipu AI has chosen a unique entry point. On May 21, they signed the Frontier Artificial Intelligence Security Commitment with 15 top AI companies including OpenAI, Google, and Microsoft, joining the global AI governance dialogue as equals. On the same day, Baichuan Intelligent's Baichuan 4 set a new domestic record in the SuperCLUE comprehensive benchmark evaluation with a score of 80.64, surpassing GPT-4 Turbo, demonstrating the firm determination of Chinese companies to catch up in technology.

Following this, Baichuan Intelligence launched a new generation of large model Baichuan 4 and the first AI assistant "Baichuan" on May 22. This assistant has significantly improved its general abilities, mathematics and coding abilities, ranking first in domestic evaluations.

Two long-dormant technology giants also woke up this month.

On May 15, ByteDance released a large model of Beanbao and announced usage prices that were enough to stir up the industry. For the pro-32k version of the Doubao universal model, the model inference input price is only 0.0008 yuan/thousand Tokens, while the pricing of models with the same specifications on the market is generally 0.12 yuan/thousand Tokens, which is 150 times the price of the Doubao model. The pro-128k version model inference input price is 0.005 yuan/thousand Tokens, which is 95.8% lower than the industry price.

At the end of the month, Tencent also joined the fray. Tencent Yuanbao AI products are developed based on Tencent's Hunyuan large model and have multi-modal capabilities. They are designed to provide instant answers, creative inspiration and fresh information, covering multiple scenarios such as knowledge learning, life encyclopedia, workplace office and fun creation.

However, all is not well in this fierce competition. Regrettably, after Baichuan Intelligent’s blockbuster release in April 2024, there was no similar follow-up for more than half a year. At the end of the year, news of the resignation of co-founder and commercialization leader Hong Tao set the stage for this quarter. Cast a shadow.

As the smoke of May gradually dissipates, the sunshine of June will shine into this track full of expectations and uncertainties. Who can stand out in this war without gunpowder is still an unsolved mystery.

Who fired the first shot in the price war?

In the fierce race of artificial intelligence, ByteDance, the former sleeping giant, finally woke up, only to find that the world had changed beyond recognition. If DeepSeek fired the first shot of the price war with that price bullet, then Byte was like a heavy gunner, instantly detonating a comprehensive price war.

In early May, the private equity giant Huan Fang’s in-depth search was like a dark horse., taking the lead in launching a price reduction raid with the cost-effectiveness of the DeepSeek-V2 model. This model is close to GPT-4 in terms of mathematics, programming, Chinese and English capabilities, and its price is only 1/35 of GPT-4o, which instantly ignited the sensitive nerves of the industry.

ByteDance immediately joined the battle and, in its usual aggressive style, slashed the input price of the Doubao universal model Pro-32k to 0.8 yuan/million tokens. Tan Dai, president of Volcano Engine, even said boldly: "Large models will be priced in cents from now on." This move is tantamount to issuing a general mobilization order for price wars to the entire industry.

Alibaba Cloud and Baidu followed suit, slashing prices and even launching free models. Not to be outdone, iFlytek and Tencent announced that Spark Lite API and Hunyuan large model lite 256k are free of charge. In just a few days, the domestic AI large model market has fallen into a comprehensive price war.

However, this seemingly fierce price war actually exposed the deep anxiety of domestic AI companies. A price war is like drinking poison to quench thirst. It seems pleasant, but in fact it is full of dangers. Companies are trying to gain market share at low prices, but behind them are confusion about business models and uncertainty about the future.

ByteDance’s move may seem powerful on the surface, but in fact it reveals its helplessness on the AI ​​track. As an Internet giant accustomed to winning through scale and traffic, they seem to have yet to find a real way to win in the field of AI. Price war is just one of the few weapons in their hands.

This price war without gunpowder is dragging all participants into an increasingly narrow channel. Technological innovation is submerged in the vortex of price, and commercial value is diluted by disorderly competition. Whoever can maintain rationality and long-term vision in this seemingly fierce competition is actually internal friction, and whoever can finally stand on the commanding heights of this emerging track.

When the smoke of the price war dissipates, what may be left is only a piece of chicken feathers and confusion about the future. In AI, a track destined to change the world, price should never be the ultimate winning weapon.

Undercurrents in June and July: The game of talent, capital and computing power

The smoke of the price war has just dissipated, and the AI ​​industry has not ushered in peace. Instead, it has entered a more cruel competition - a fierce battlefield of talent, capital and computing power. If May is the stage for the awakening of giants and price storms, then June heralds that deeper competition is about to begin.

In June, the sense of smell of capital was extremely sensitive. On June 3, Zhipu AI was the first to gain favor from international capital, receiving US$400 million in financing from Prosperity7, a fund under Saudi Aramco, with a valuation exceeding US$3 billion. This is undoubtedly a strong endorsement of its technical strength and development prospects.

On June 17th, Hangzhou DeepSeek announced that DeepSeek Coder V2 is open source. This model is close to GPT-4-Turbo in terms of code and mathematical capabilities, with a total parameter of 236 billion, ranking among the best in the world. The comprehensive open source of models, codes and papers has set a benchmark for open sharing for the industry and accelerated the prosperity of China's AI ecosystem. DeepSeek's open source move not only demonstrates its technical confidence, but also attracts the attention of more developers and researchers, paving the way for the subsequent battle for talent.

If the flow of capital is an undercurrent, then the competition for talents is an open game. In July, the battle for talent became intense. ByteDance showed shocking offensiveness, and hired "the person who knows the most about Alibaba Big Model" with high-standard treatment - Zhou Chang, the former head of Tongyi Big Model, and his team, and even recruited zero-one things. Many top talents in the AI ​​field include former vice president of algorithms Huang Wenhao and former core members of Wall-Facing Intelligence.

This series of heavy signings has instantly rewritten Byte's position on the AI ​​talent map, and also indicates that this Internet giant will launch a more intense offensive in the AI ​​field. Although this kind of "poaching" behavior is common in business competition, it also reflects the extreme scarcity of talents in the AI ​​field and the desire of enterprises for top talents.

At the same time, the situation of Lingyiwanwu seems to be extremely difficult. After Technology Lianchuang Huang Wenhao left, co-founder Li Xiangang also chose to return to real estate trading platform Beike, and product leader Cao Dapeng left soon after.

This AI upstart, which was founded just over a year ago, is facing the continued disintegration of its core team when its valuation once reached 20 billion yuan. There were almost no major product launches in the second half of the year, casting a shadow over the development of this star company and raising industry concerns about the stability of talent at AI startups.

The craze of capital has not diminished due to the flow of talents. At the end of July, Baichuan Intelligent completed a 5 billion yuan Series A financing, with top investors such as Alibaba, Xiaomi, Tencent, and CICC investing in the company, once again proving that the capital market continues to be optimistic about the AI ​​​​track. In August, Dark Side of the Moon received another $300 million in financing, with its valuation rising to $3.3 billion, consolidating its leading position in the AI ​​field.

In August, Step Star welcomed a heavyweight—Zhang Xiangyu, one of the four authors of ResNet. The joining of this post-90s AI guru later attracted the follow-up of Tencent Research Director Yu Gang and Microsoft Asia Research Institute Duan Nan. This series of talent acquisitions not only brings technological boost to the company, but also sends a key signal to the industry: on the road to catch up, original technological innovation may be the real magic weapon to retain talents. Instead of poaching employees with high salaries, it is better to build a more attractive technology platform and research atmosphere.

At the end of the year, Xiaomi, a company that has always been relatively "conservative" in the field of AI, also began to make efforts, poaching Luo Fuli, one of the core developers of DeepSeek V2 and a genius born in 1995, with an annual salary of tens of millions. Breaking the external investment in the field of AIInadequate stereotypes.

At the same time, another shocking story is unfolding in the global capital market. On June 5, Nvidia's stock price surged, and its market value exceeded US$3 trillion for the first time, surpassing Apple to become the second-highest market value company in the world.

As of June 19, it surpassed Microsoft and became the most valuable company in the world with a market value of US$3.34 trillion. In the past five years, its stock price has soared 3477.31%, far surpassing Microsoft and Apple, becoming the undisputed computing power leader in the AI ​​era, which also highlights the key role of computing power in the development of AI.

The flow of talents, the influx of capital, and the rise of computing power dominance together constitute the main theme of the AI ​​industry in June. The upcoming July will be an important stage for technology demonstrations and industry exchanges.

[Please continue reading the next part]