Picture from Visual China

Blue Whale News, December 23 (Reporter Wu Jingjing) TuSimple, a former autonomous driving star company, has re-chosen the path of entrepreneurship after experiencing a series of storms such as delisting.

In August 2024, the company first revealed that it would target the AIGC direction. Four months later, on December 17, the company announced the launch of a new brand, CreateAI, and released its first big TuSen video model product “Ruyi”.

Currently, the Ruyi-mini-7B version is officially open source on HuggingFace, and users can download and use it. It is understood that “Ruyi” is designed to run on consumer video cards (such as RTX 4090).

Is it okay to switch from autonomous driving to AIGC?

Why is the company moving from autonomous driving to a vastly different visual model circuit?

In an exclusive interview with Blue Whale News, the person in charge of future technology at Tucson told Blue Whale News that this transformation mainly starts from the perspective of company transformation and business development. On the one hand, the company has already accumulated algorithm, computing power, and data experience in the AI field when doing autonomous driving; on the other hand, co-founder Chen Mo has accumulated resources in the gaming industry and has the opportunity to quickly implement technology.

Chen Mo also mentioned in an interview with the media earlier that in the future, Tucson is trying to find a faster way to “save” the company with existing resources. With the continuous advancement and disclosure of visual model technology, AIGC is currently the best choice in terms of commercial space and technical accessibility.

According to Blue Whale News, Tucson's future visual model team is the company's previous autonomous driving team, and some technical experience can be reused.

The person in charge of future technology at Tucson said that AI visual models and autonomous driving all rely on the “troika” of algorithms, computing power, and data to drive technological development, and both rely on large amounts of data for training and optimization. However, video generation technology is very similar to sensing modules in autonomous driving. Both are mainly data-driven, the R&D path is relatively short, and the technical foundation they rely on is also relatively clear.

In his view, autonomous driving technology involves various algorithm modules such as perception, positioning, planning, and control, as well as various fields such as software systems, hardware design, and vehicle structure design. In contrast, video model technology has a shorter development path and narrower technical aspects, mainly focusing on data processing and model training.

“The importance of data even surpasses algorithms,” said TuSimple technical director. He mentioned that Tucson has accumulated rich experience in data labeling in the field of autonomous driving, has a self-built labeling team and labeling platform, and has established a complete data processing process. “These experiences and tools can be applied directly to the data preparation phase of the video model, saving significant time and costs.”

Of course, in terms of the technology and effects of many visual models, Tucson still needs to explore from scratch. Currently, the company's video model mainly revolves around five key metrics: production quality, consistency, controllability, ease of use, and cost.

TuSimple Technology Director told Blue Whale News that the quality of model generation is the primary goal to ensure that the generated video content reaches a high level of image, motion, and detail. “The company adopts a spiraling R&D strategy to gradually improve the controllability, ease of use and cost efficiency of the model on the premise of ensuring the quality and consistency of production.”

Tucson chose the third path: don't rely on models to make money, end up doing the content yourself

Currently, new developments continue in the field of visual models. On December 9, local time, OpenAI of the United States officially released the latest version of the video generation large model Sora-Turbo, which can generate new video content based on text, image, or video input. Furthermore, in the domestic market, whether it is large technology giants such as Byte and Kuaishou, as well as startups Pika, Aishi Technology, and Shengshu Technology, etc., continue to actively promote technology and product iteration.

Did Tucson enter to get a share of the pie on the highly competitive visual supermodel circuit?

I learned from the company's current business developments and interviews that the answer is no. The person in charge of future technology at Tucson told Blue Whale News that the more accurate statement is that the TuSimple to become a content company rather than a big model technology company is not the same direction as platforms such as Kuaishou and startups such as Pika.

Currently, there are only two types of large-scale visual models on the market: one is Runway, Pika, and other companies, which provide paid video generation tools or services for C-end users, so that creators can pay to create their own content; the other is for B-side film, entertainment, and gaming companies to help reduce costs and increase efficiency in the industry.

TuSimple Technology Director told Blue Whale News that if positioned as a pure video model company, TOC and TOB both have obvious challenges:

On the one hand, in the TOC direction, the target user group of video generation tools is professional creators, not the general public. The fee model and profit prospects are unclear, and the video model requires a large amount of computing power support. The operating cost is high. The domestic market makes it difficult to rely on fees to attract users and achieve profits in the short term.

On the other hand, the implementation of technology empowering the B-side alone is a huge challenge, because it is difficult for technology companies to thoroughly understand the requirements of specific scenarios, and it is also difficult to effectively integrate technology into actual production processes to control the quality and style of content.

Compared to many video models that are focusing on the versatility of technology, Tucson chose a different third path: directly open source model technology, not rely on models to make money, buy classic IPs, and end up using big models as content themselves.

According to Blue Whale News, the company currently has a dedicated animation and game team already developing new projects.

“We want to build a video content creation company driven by AI technology to create an end-to-end video content generation chain. Ultimately, high-quality content is used to attract users and achieve commercial value.” Tucson's Head of Future Technology said. “Technology is just a tool; the ultimate goal is to provide users with content.”

Currently, Tucson has set up an animation and game section, and its new brand, CreateAI, has been authorized by the famous martial arts IP “The Legend of Jin Yong Xia” to develop a large-scale martial arts open world RPG game. In August 2024, the company also officially announced that it had reached a cooperation with Shanghai Three Body Animation Co., Ltd. to jointly develop the first animated feature film and video game in the “Three Body” series. It is reported that the company will also launch SLG game tools and the game itself in December.

“We now have two top IPs, “Legend of Jin Yong” and “Trisomy,” and our goal is to achieve $1 billion in revenue by 2027.” In a recent interview, Chen Mo talked about Tucson's future goals in the AIGC direction.