In early October 2024, several investors who had invested in large language models gathered to discuss how to sell their shares in large language model startup companies. The investment market that once scrambled for shares is now filled with pessimism.
In less than 12 months, the positions of startups and large companies in the large language model race have reversed. ByteDance has emerged as a latecomer in large language models. Although it didn’t catch up with the first starting gun last year, its reputation has reversed since this summer. ByteDance has launched multiple AI applications and hardware. For example, the Doubao App has a daily active user count of 9.45 million, becoming the AI product with the largest daily active users in China. It has also launched the Ola Friend headset and is researching and developing AI glasses. During the same period, many startups entered a period of adjustment. Companies like Zhipu AI, 01 All Things, MiniMax, Dark Side of the Moon, and Baichuan Intelligence have seen personnel adjustments, and there have been no news of large language model financing in nearly half a year.
Last winter, Alibaba and ByteDance fired in opposite directions. Alibaba’s aggressive investment reversed the market sentiment. For example, investing in the Dark Side of the Moon greatly increased its valuation, bringing ammunition and momentum and stimulating investment institutions to continue financing. Since last winter, ByteDance has also had new actions. Founder Zhang Yiming has communicated with many people. This year, he has introduced many technical talents and integrated internal forces to establish the Flow department. Zhu Jun and Zhu Wenjia are respectively responsible persons, and there are multiple teams under it.
Startups face a dilemma. To C AI products are difficult to break out of the mobile internet ecosystem and are easily in direct competition with large companies. In terms of advertising and traffic acquisition, large companies have obvious advantages. For example, ByteDance controls the穿山甲 platform and super apps like Douyin. Doubao surpasses Kimi, which has a lot of advertising materials, in terms of daily active users and monthly active users. Large companies are also competing for large language model API customers. Alibaba’s open source large language model affects the orders of startups. In model research and development, ByteDance also shows a “saturated” offensive stance, such as in the field of video generation models.
The possible way out for startups is to find niche directions or be willing to do the dirty and tiring work. For example, in the field of companionship and virtual social products, the firepower of giants is currently relatively weak. Or they can wait for internal tensions to emerge in large companies. But in China, there are more than just ByteDance that possess extremely strong product capabilities and traffic resources. WeChat has not made a move yet.
The era of large language models is like the era of computer vision back then. The outcome of the competition between startups and large companies depends on the degree of disruption of new technologies, thresholds, and whether a new ecosystem can be found. Currently, large language model products are in a state where semi-mature technology meets a semi-new market. The immaturity of technology is an opportunity for entrepreneurs, while the semi-new market is a weight for large companies. Those who were full of ambition and crowded onto the large language model rocket last year need to adjust their postures to adapt to the new reality.
