According to Caishang News on September 21st (Reporter Cui Ming), the gaming industry's focus on Generative AI and large models shifted from cost reduction and efficiency enhancement last year to promoting gameplay innovation and experience upgrade this year. At the recent 2024 Yunqi Conference, the 'AI Endgame Challenge' gameplay of the mobile game 'Space Killing' under Giant Network made its appearance at the exhibition and attracted attention.

Caishang News journalists learned that the 'AI Endgame Challenge', based on Giant Network's self-developed Multi-Agent gaming technology framework, is the industry's first gameplay of a multi-agent AI native game. In the game, AI players can exhibit high-intelligence behaviors, strategize, disguise, and even 'group hug,' which sets it apart fundamentally from the current market's AI companion type and AI NPC games.

"We believe this is a relatively successful experiment," said Ding Chaofan, director of Giant Network's AI laboratory, in an interview with Caishang News reporters. He mentioned that the 'AI Endgame Challenge' is a practical product of the team's AI Agent technology, and judging from the heat and player feedback since its launch in August this year, the technology has achieved a relatively ideal combination with gameplay.

Ding Chaofan believes that the future form of games will break traditional rules. The game world can be updated based on player feedback data, not only enabling dynamic expansion in design but also including random events and trigger Easter eggs, creating a non-linear open world.

"From a global perspective, the industry's exploration of AI Agents is still in the early stages, which is also a research hotspot in the AI field this year," Ding Chaofan disclosed. Internally, the team has different directions in AI technology applications, but overall, they will focus on breaking through multimodal capabilities and AI Agents.

Self-developing large models focusing on the application in the gaming field.

As an important technology in the field of artificial intelligence, Multi-Agent large models have been under great scrutiny in the past year. They combine the powerful language processing capabilities of large language models and the collaborative characteristics of Multi-Agent systems to achieve intelligent independent decision-making and interaction in complex environments.

Ding Chaofan told the Caixin journalist that the core of the Agent is still driven by large models. In the past, traditional AI technology solutions had a common problem of insufficient generalization ability in new environments. With the emergence of large language models, there has been a very strong generalization ability, which in turn has brought the ability to create new content and new forms, making the "AI Endgame Challenge" gameplay possible.

In order to successfully implement multiple intelligent agents with large models in the popular game "Space Kill" with a user base of 0.2 billion, the Giant Network AI Lab team balanced the illusion problem, ensured content security, and reduced reasoning costs to make this AI gameplay both fun and controllable.

"Our exploration of AI focuses on the gaming industry." Ding Chaofan said that the company had already established a complete set of AI industrialized production systems last year, including large models, visual, and audio aspects. AI as a productivity tool has achieved significant results internally in the company. "For example, the all-in-one AI art production platform we developed internally has become the daily tool used by colleagues in the art and design departments."

Ding Chaofan stated that, relatively speaking, AI is currently more mature in 2D art production and voice generation, but there are still certain technological bottlenecks in 3D model generation. Due to the limited number of 3D assets, it is difficult to form a universal and effective model. This is a problem faced by the entire industry.

Bullish on multi-modal and AI Agent directions.

Considering the current reasoning costs and return on investment, it is difficult for AI products to achieve large-scale commercial implementation. Apart from a few major companies investing heavily, most AI products developed by enterprises are only targeted at specific users, not open to the general public.

Ding Chaofan mentioned that in terms of reasoning costs, current computational pressure can be alleviated through some technical means. "The reasoning costs are decreasing at a rate of tens of times per year. In the future, reasoning costs will definitely not be a problem. We need to overcome these so-called bottlenecks to think about the future benefits to the business."

Although the gaming industry's discussion on generative AI and large models has shifted from cost reduction and efficiency improvement to gameplay innovation, Ding Chaofan believes that the two are not contradictory. "From my personal perspective, this year I am more focused on how the capabilities of large models and multimodal AI can be combined with the gameplay itself to optimize or even revolutionize players' gaming experiences. Our exploration in this area is quite rapid."

Cailian Press learned that at the 2024 Yunqi Conference held recently, Giant Network has launched two self-developed 'game+AI' large-model applications, one being the character large model GiantGPT, whose application effect has been verified in the Giant Network's online games business scenarios. The other is the voice model BaiLing-TTS, which is currently applied in practical scenarios such as game NPC voice-overs and video creation.

Ding Chaofan revealed that this year the team will focus on two main directions: multimodal and AI Agent. From a gameplay perspective, without being bound by rules, maintaining a high degree of freedom while allowing players to feel involved, we believe this will be a future form of gaming and was our original intention to explore the combination of AI and games.