Talent remains the most valuable asset in the 21st century.

Zhang Yiming has retired, but not completely.

According to the Financial Times, despite having announced his retirement and stepped down as CEO, Zhang Yiming remains attentive and actively involved in ByteDance's AI strategy, personally overseeing the recruitment of AI talent from competitor companies. Companies such as Alibaba and Zhipu Qingyan have already seen a significant outflow of talent to ByteDance.

The former CEO's transformation into a 'headhunter' to recruit talent for the company underscores ByteDance's emphasis on AI business. In fact, among domestic internet companies, ByteDance's layout in the AI field may be deeper and more extensive than that of internet giants like Alibaba and Tencent.

Simultaneously Developing Multiple Large Models, ByteDance Becomes the Leading Player in AI

Currently, ByteDance's AI products can be classified into three categories: intelligent agent development platforms, various AI large models, and AI applications for apps or web pages. The development platforms mainly include Kouzi and Coze, targeting domestic and international markets, respectively. There are over ten products each for AI large models and application platforms, covering areas such as chat, social media, images, text, video, music, education, office work, coding, e-commerce, and more.

Among ByteDance's products, the most widely recognized is Doubao, a large language model comparable to ChatGPT. This model supports input in various formats, including text, images, and PDFs, and can output text and images based on user requests. Lei Technology has tested Doubao multiple times, and among freely available AI large models in China, Doubao's experience and functionality richness are undoubtedly top-notch.

According to DianDian data, Doubao ranks second on the App Store's free app chart. In October this year, Doubao had 51.39 million monthly active users, ranking second globally among AI applications, trailing only ChatGPT with 258 million monthly active users.

(Source: DianDian screenshot)

Other AI applications, such as audio generation app Sponge Music, video generation app Jimeng, code generation app CodeGen, image generation apps Xinghui and Xingtu, among others, are comparable to AI products from other internet companies. Limited by factors such as cost and technology, many AI companies only focus on specific areas; for example, Kimi, the top AI app for buying traffic in China, still does not offer drawing functions and focuses solely on text content.

With strong financial and technological foundations, ByteDance has chosen a 'big and comprehensive' approach, launching relevant large models and apps in various fields such as audio, video, drawing, and graphics, to compete with other AI company products. Additionally, while most domestic AI companies' operations are limited to China, with function development centered on domestic user needs, ByteDance attaches great importance to overseas markets, ensuring that its AI applications have corresponding apps in international markets, such as Doubao and Cici, Xingtu and Hypic, hinting at greater future development potential.

Simultaneously developing a large number of AI large models has the advantage of covering multiple fields and reducing the risk of missing out on opportunities but also consumes more resources, requiring a large number of talents and computing power. Zhang Yiming's personal involvement in poaching talent from other internet companies is to prepare for the simultaneous development of multiple large models.

In terms of computing power, ByteDance has purchased a large number of NVIDIA GPUs over the past two years. Its data centers deployed abroad are not subject to bans and are also purchasing H100 and Blackwell chips. The Financial Times also reported that after poaching talent from other chip companies, ByteDance has formed an AI chip R&D team that is developing AI chips for machine learning. News of ByteDance developing its chips has been exposed multiple times before, and it is rumored to adopt TSMC's 5nm process technology, with mass production expected in 2026.

ByteDance has perfectly timed two key trends in the graphic and video era, launching Today's Headlines and Douyin, both of which have become industry leaders. Facing the AI era, ByteDance is clearly well-prepared, purchasing computing power chips while aggressively poaching talent. However, for ByteDance, the AI business is still in the layout and loss stage, with profits remaining elusive. Achieving profitability in the AI business as soon as possible is essential to avoid long-term investment losses.

Creating Specialized Sub-scenarios: Will AI Ultimately Become ByteDance's Cash Cow?

OpenAI CEO Sam Altman once planned to raise $7 trillion to reshape the global chip industry and lay the infrastructure for AI technology development. In response, NVIDIA CEO Jensen Huang bluntly stated that this amount of money could buy all the GPUs in the world, given that the current total value of global data centers is only around $1 trillion.

Although Sam Altman's idea is somewhat fanciful, his point is correct: AI technology development is extremely costly. More critically, many current large AI models have yet to find a suitable profit point. Services like Kimi, Doubao, Zhipu Qingyan, iFLYTEK Spark, and ERNIE Bot are all free to users, with ERNIE Bot even retaining the free-to-use ERNIE 3.5 large model.

(Source: Doubao screenshot)

I once subscribed to ERNIE Bot's large model membership, which cost up to 50 yuan per month. While it seemed to offer many functions, the necessity was limited. Coupled with other large language models available, I did not renew my subscription. Based on my experience, there is still significant room for improvement in the intelligence of large language models. At this stage, when generating images, they cannot recognize proportional requirements. Moreover, with too many similar products, competitiveness inevitably declines significantly after charging, leading AI companies to hesitate to adopt charging schemes for their large language models.

Insufficient intelligence leading to limited practicality and too many similar products are the two main factors contributing to the difficulty of profitability for large language models. Currently, revenue from AI large models relies heavily on customized needs from B-end customers.

Despite having the second-highest monthly active users after ChatGPT, Doubao's average daily usage time from June to August this year was only 5.3 minutes, leaving significant room for improvement compared to mainstream internet apps like short video, live streaming, and social media.

The key to eliminating the plight of large language models and achieving profitability lies in creating specialized usage scenarios to increase user stickiness. In fact, in today's world where large language models are generally free, specialized large models with fewer functions tend to charge fees. For example, ByteDance's image and video generation large model Jimeng charges 659 yuan per year for basic membership and 5,199 yuan per year for premium membership. Similarly, KuaiShou's Keling and Alibaba's Tongyi Lingma, both specialized large models, have also adopted charging schemes.

(Source: Jimeng screenshot)

ByteDance, with its AI large models covering multiple fields, already has inherent advantages. However, Doubao, which has the largest user base and unlimited potential, urgently needs to find a breakthrough.

At the recent re:Invent conference, Amazon Web Services announced that it will launch a versatile large model capable of inputting and outputting content in any modality next year. Various specialized models are emerging, but only large language models have the potential to encompass everything and complete any task according to user requirements.

Amazon Web Services revealed that more than half of its customers use different large models in different scenarios. Completing different tasks with different large models not only reduces work efficiency but may also require subscribing to multiple memberships, increasing cost expenditures. Enhancing product intelligence by integrating the capabilities of other large models into Doubao may enhance its competitiveness and help ByteDance create a versatile model before Amazon Web Services.

Leveraging the performance of versatile models in specialized scenarios to improve Doubao's competitiveness and drive traffic to ByteDance's other less well-known large models for specialized scenarios is crucial for ByteDance to rapidly improve the profitability of its AI large models.

AI Large Models Clashing: Will ByteDance's Aggressive Hiring Weaken Competitors?

The vast consumer market of 1.4 billion people, coupled with the large number of talents graduating from institutions each year, has laid a solid foundation for domestic enterprises to lay out emerging industries. Coupled with the country's extensive support, domestic competition in industries such as new energy vehicles, XR, and AI large models is extremely fierce.

I strongly support this national support scheme because competition drives progress. The recent bankruptcy of Northvolt, Europe's largest locally-built power battery enterprise, after burning through $14 billion (approximately 100 billion yuan) is a case in point.

However, for enterprises, excessive market competition reduces the chances of survival. Following the AI wave sparked by ChatGPT, domestic AI companies have sprung up like mushrooms after rain, with 237,000 new AI companies established in the first half of this year alone. Over the past few years, more than 80,000 domestic AI companies have gone bankrupt.

(Source: Doubao AI generation)

With strong financial resources, ByteDance has sufficient strength to simultaneously undertake the R&D of over ten AI large models. However, facing giants like Tencent, Alibaba, and Baidu, ByteDance cannot guarantee victory. Poaching talent from competitor companies not only enhances its strength but also weakens the opposition.

Once the talent war begins, domestic AI companies will increase the salaries of technical personnel and the R&D costs of AI large models, quickly eliminating smaller vendors. Following the development trajectory of new energy vehicles, 80% of domestic AI large model companies may go bankrupt in the next five years. Only companies that persevere will have the ability to share a piece of the AI industry pie.

The AI subsidiaries of internet giants such as ByteDance, Alibaba, Tencent, and Baidu have sufficient financial support and already possess the foundation to grow into leading enterprises. The ultimate outcome will depend on the performance of these companies' AI large models in specialized scenarios over the next two to three years, as well as their exploration of versatile large models.

Source: Lei Technology