hHow Generative AI Is Reshaping the Gaming Industry KellyOnTech

5 min readDec 4, 2023

I will use the gaming industry as an example to discuss the third stage of generative AI development — the original stage. According to Robin Li, the founder of Baidu, generative AI will be divided into three stages of development: assistant stage, collaboration stage and original stage. According to Sequoia Capital’s prediction, by 2030, AI will have original capabilities. I would like to give special thanks to Mr. Tao Jianrong, CTO of Netcase Gaming, for providing first-hand information on the gaming industry.

In what ways has the AIGC impacted the gaming industry?­

At present, the impact of AIGC in the gaming industry is reflected in the whole value chain of R&D and O&M (Operation and maintenance), see the table below.

Table: The impact of AIGC in the gaming industry

I specifically introduce Yuyan language large model, Danqing multimodal large model and AI generated game policies.

Yuyan Language Large Model

The Yuyan language large model is a series of Chinese text pre-training large models independently developed by NetEase Fuxi. NetEase Fuxi is NetEase’s top organization specializing in game and AI research and applications. The large model is based on NetEase Fuxi’s self-developed GPT2 generative model, which uses hundreds of billions of parameters and supports expansion from text to multiple modalities such as graphic, music, and behavioral sequences.

With a large language model like ChatGPT, why do we need to develop the Yuyan language large model?

Generalized large models do not appear to be deep enough in understanding domain-specific needs, while Yuyan language large model can deeply understand the copy writing needs of the game domain while maintaining the basic capabilities of the large models, and can automatically generate copy. And after human game planning fine-tuning, the AI can continue to assist in the creation of generating more new copy. It can set up different copywriting styles and word game play.

Danqing Multimodal Large Model

NetEase’s self-developed text and image generation model,”Danqing”, is a semantically enhanced diffusion model for text and image generation, with training based on 800 million graphic data. Through text input, it can quickly generate high-quality images. At present, it has safely and confidentially hosted more than four thousand projects and opensource models.It is the largest AI drawing model library in China.

What are the results of the images generated by the Danqing Multimodal large model?

Song Huizong Zhao Ji created a painting called “ Plum blossoms in the waxing moon and Mountain Fowl Scroll”, and the Danqing large model created a painting based on the poem that Song Huizong mentioned to the painting.

Image source: Baidu Encyclopedia. Song Huizong Zhao Ji’s famous painting “Waxing Plum and Mountain Fowl Scroll” Taipei National Palace Museum
Image source: Infoq. Netease Danqing multimodal large model generated by the image “Danqing”

The advantages of the Danqing multimodal large model lies in its deeper understanding of Chinese culture compared to other models.In the corpus, there is a large amount of data on traditional Chinese culture, idioms, paintings and calligraphy, poems and food, .etc. The graphs generated by the Danqing large model are more oriental aesthetic. For example, the following picture is created by using a famous poet Li Bai’s famous line meaning water falling from a 3,000-foot-high waterfall).

Image: InfoQ. Ink landscape painting generated by the Danqing Multimodal large model

In addition, to improve the quality of AI-generated images, Danqing large model introduces manual feedback in both the training and generation phases. In the training phase, manual evaluation from multiple dimensions, screening out many high-quality graphic matching, high aesthetics data, to help the basic model to get better results; in the generation phase, manual scoring of the model’s semantic generation capabilities and image aesthetics, screening out a large number of high-quality generated results, introduced into the model as positive feedback.

AI Generated Game Policy

Through reinforcement learning, AI can set different game styles, different segments of the game strategy, and adjust the strategy in real time according to player’s mindstream and teammate feedback. In the 2022 Hangzhou Asian Games, eSports was first listed as an official competition event. I take one of the events, “Dream of Three Kingdoms 2”, as an example.

“Dream of Three Kingdoms 2” was developed by Hangzhou Electric Soul Network Technology Co. in 2014. The founder, Yu Xiaoliang, was one of Tencent’s earliest game producers. This traditional style MOBA game, also known as multiplayer online battle arena (MOBA) uses AI to generate game strategies. With its self-developed E-soul engine, the game has lots of innovations in terms of image quality, game content, game strategy and so on. “Dream of Three Kingdoms 2” was released in 2015 and selected as one of the esports programs for the 2022 Hangzhou Asian Games in 2021. Although this game is not as popular as the League of Legends, which is represented by Tencent, its local development team and references to Chinese culture elements are officially revered.

Image: Chinese team won the gold medal in the “Dream Three Kingdoms 2” program at the Hangzhou Asian Games

As technology continues to advance, generative AI will be able to perform all aspects of game development and reach a level comparable to professional developers and designers. This further improves the efficiency of development in the gaming industry while reducing costs. Will the gaming industry be the next field to be replaced by AI?

Video version

The original article comes from Mans International website blog




Tech Investor / Founder at Mans International / Author