When OpenAI launched its conversational AI chatbot, ChatGPT, the rise of generative AI and large models swiftly permeated various industries. The gaming sector, standing at the forefront of technology, has inevitably embraced the wave of “Gaming + AI.” How should emerging technologies like large models be perceived, and what adjustments should the gaming industry make to align with this technological trend?
Recently, a representative from The Paper engaged in a dialogue with Ding Chaofan, head of AI Lab at Giant Network Group. Established in late 2022, Giant Network has been emphasizing the significant impact of AI on future developments, as highlighted multiple times by its chairman, Shi Yuzhu. Shi has stated that “Gaming + AI” is an inevitable trend and a primary area for the company's strategic focus. He envisions that in the future, the gaming industry will primarily benefit from AI’s ability to categorize players meticulously, tailoring unique gameplay experiences to match individual preferences.
Giant Network was among the first companies in China's gaming sector to complete the formal registration required for development. Ding Chaofan believes that gaming companies have an innate advantage in developing role-playing large models due to the existing scenarios and data. “Games are an excellent application scenario for large models; we provide the nails before crafting the hammer,” he noted.
Here are key insights from the conversation:
On the impetus for developing large models
The Paper: What prompted Giant Network to initiate the development of gaming large models?
Ding Chaofan: Primarily, it was driven by product and user needs. We believe in having the foundation before building a complex structure; we already had gaming products, which serve as a strong application scenario for our large model.
On the advantages of developing large models
The Paper: What specific advantages do you have in developing large models?
Ding Chaofan: Our core advantage lies in our extensive data. With 20 years in the industry, we have accumulated vast amounts of data that expedite our large model development. From deciding to develop a large model to team formation and initial project setups, we utilized about six months, aiming to complete the registration of our large model, GiantGPT, in February 2024.
On changes in gaming products using large model technology
The Paper: How have gaming products changed since implementing large model technology?
Ding Chaofan: The main highlights include character portrayal, context-driven reasoning, and long-term memory capabilities for AI agents. For character development, large models introduce varied speech patterns, allowing for more lifelike, personalized non-playable characters (NPCs). For instance, if a character like Zhang Fei is depicted as robust, his speech style would correspondingly reflect that.
Moreover, the AI can exhibit emotional intelligence—for example, if players repeatedly ask the same question, the NPC will show frustration, adding to the realism. We are not limited to intelligent NPC dialogues; we have integrated AI agents that can remember player preferences and personalities for a richer gaming experience.
On the influence of “hallucination” in gaming
The Paper: Does the phenomenon of “hallucination” in large models impact gaming products?
Ding Chaofan: Interestingly, this is somewhat beneficial in gaming. Hallucinations can introduce randomness and variety, enhancing player experiences with each session. We are exploring the integration of AI with social reasoning in gaming. The ambiguity introduced by these hallucinations may increase gameplay enjoyment while avoiding the costs associated with intense reasoning demands, which is a natural advantage of applying AI in gaming.
On the commercialization challenges of large models
The Paper: What challenges do you foresee in the commercialization of large models?
Ding Chaofan: There are significant challenges, particularly in AI product commercialization. For instance, a small percentage of heavy users consume a disproportionate amount of resources, making it difficult to cover development costs through minimal subscription fees. Currently, the high cost of reasoning limits broad accessibility, and any scaling requires careful consideration of return on investment (ROI).
Addressing these challenges
The Paper: How do you plan to tackle these challenges?
Ding Chaofan: We can mitigate reasoning costs through various technical means to reduce hardware needs and ease computational burdens by optimizing specific scenarios. We anticipate that reasoning costs could decrease dramatically within the next few years. Additionally, we are already seeing some promising small parameter models that are well-suited for our needs. As marginal costs drop, more business models will become viable.
We believe that traditional models will not impede the commercialization of AI; instead, it’s crucial to think beyond conventional frameworks. Many industries cite the difficulty of commercialization due to technological limitations, but that primarily stems from inadequate model reasoning capabilities. Compared to general large models, specialized models are likely to achieve commercialization more effectively. We predict that every industry will eventually have its own specialized large model—it's just a matter of time.
Further investments required
The Paper: Beyond reasoning costs, what other investments are needed?
Ding Chaofan: The primary investments are still in computational power and manpower. This year, we achieved breakthroughs in two areas: AI Agents and multimodal interaction. While the industry lacks a killer application, we are confident that the future applications will be grounded in our AI Agents.
These insights underscore the vital intersection of gaming and AI, illustrating how emerging technologies can reshape player experiences and the operational landscape of the gaming industry.
Ranking
