Game-Time reports/games, as a product of the combination of culture and technology. Not only does it serve as a testing ground for various creative ideas, it is also one of the best implementation scenarios for cutting-edge technologies. In the past year, the breakthroughs and evolution of large models and AIGC technology have brought huge changes to the game industry. They have not only changed the game development process, but also brought greater room for imagination to the entire game industry.
Imagine being able to have anthropomorphic, impromptu, and dynamic conversations with NPCs in the game. These characters also have different personalities that evolve over time, as well as accurate facial animations and expressions – what really makes “Westworld” and “Runaway Player” The plot is brought into reality, and players can get a good experience and immersion in the virtual world.
Now this is no longer just a romantic fantasy. With AIGC innovating the traditional game content experience, the ideal intelligent NPC no longer seems out of reach. As more and more leading companies such as NVIDIA and NetEase fail, more and more NPCs in games will be given “souls” by AI.
The wave of AI is surging, and intelligent NPCs cannot be ignored
At COMPUTEX 2023 in May last year, NVIDIA released its new AI technology Omniverse Avatar Cloud Engine (ACE). This is a real-time AI solution that can provide AI models for speech, dialogue and character animation in games, that is, adding real-time interaction functions to NPCs. Nvidia said: “This service brings intelligence to NPCs through AI-driven natural language interaction, thereby changing the game.”
Recently, John Spitzer, vice president of performance technology at Nvidia, also publicly stated that all developers are interested in the way AI affects the game development process. He will provide hardware support for a key NPC test.
AI is a huge opportunity for the game industry and will bring unlimited possibilities to game developers and game operating companies. It has been recognized by more and more companies. Jeff Skelton, EA’s director of technical partnerships, has publicly stated that AI will become the standard for the gaming industry moving forward. Mature AI tools can directly be used in game design, modeling, adding expressions, voices and even actions. The intelligent NPC born based on AI role-playing capabilities can be said to be the cutting-edge application of AIGC in games.
As the carrier of the game’s worldview, NPC is an important tool to improve players’ sense of immersion. In the past ten years, against the background of gameplay innovation and diminishing margins of player screen experience, NPCs have become increasingly important as promoters of storylines and companions of player experience.
Similar to “The Witcher 3”, “Red Dead Redemption 2” and “Cyberpunk 2077”, the NPCs in the game show great freedom. They are not just background elements but an important part of the story, adding depth and realism to the narrative. However, the logic behind these products achieving this effect is still ancient, that is, through a large number of “pile of people” in the team and the use of massive scripts to achieve this effect.
There are more than 1,000 NPC characters in “Red Dead Redemption 2”, distributed in more than 100 missions in 6 chapters. Each NPC has its own singer, art writer and voice actor. The production of these NPCs took nearly 8 years and cost nearly US$500 million in research and development.
Despite this, research shows that 52% of players complain that the current NPCs “only repeat conversations”, 99% of players are looking forward to smarter NPCs, and 81% of players are even willing to pay more for them.
The emergence of large models is expected to change the current NPC production process and game experience. In fact, products represented by the mobile game “Ni Shui Han” have been at the forefront of integrating large models into games, and the use of AI character models has effectively improved Game immersion. The interaction of some in-game NPCs achieves the effect of thousands of people, bringing players a personalized gaming experience.
Unlike previous script-driven NPCs, these AI-driven intelligent NPCs can make decisions based on player dialogue, give real-time responses to player behavior, and even show some human-like emotions. This advancement makes the game world more realistic and interactive, with each NPC providing a unique and dynamic contribution to enhance the narrative depth and interactive dimensions of the game, thereby enhancing player immersion.
In the past two years, the industry has entered an era of high-quality products, competition in the global game market has intensified, and super-large productions such as cross-platform and open world have begun to dominate the market. As games become more and more complex, the creation of NPCs is also constantly upgraded, and the importance of intelligent NPCs is self-evident.
Are AI NPCs integrated into the gaming industry?Many problems need to be solved urgently
From a macro perspective, the existence of intelligent NPCs will completely change the traditional game content experience, especially plot-heavy content products such as RPGs.
By then, players will no longer just be in a world with preset navigation. Instead, they will interact with anthropomorphic characters that can learn, adapt and provide real-time feedback. This interaction will also bring unprecedented depth to the game narrative. Every decision and action of the player may lead to diverse and unexpected results.
Although the prospects of intelligent NPC are very bright, and the current large language model technology is also in a stage of rapid development, integrating it into mature game development is still full of challenges.
First of all, the biggest shortcoming of intelligent NPCs at the moment must be “not anthropomorphic enough”. Regardless of platforms such as GPT-4, Claude2.1, or Inworld, which focuses on intelligent NPCs, although the character generation capabilities are very powerful, players can clearly perceive the behaviors and emotions of these characters during the dialogue and interaction with the characters they generate. differences from humans.
This deficiency can ruin the player’s sense of interacting with “real” characters, reducing immersion.
Secondly, as of now, large models have been used for less than two years, and many AI platforms and tools based on large models still have obvious shortcomings in terms of cost and ease of use.
On the one hand, high-quality AI tools and platform API calls are expensive and difficult for small and medium-sized developers to use. The huge economic threshold limits the rapid popularity of cutting-edge AI technology in the industry. On the other hand, most AI platforms are not designed specifically for gaming. Lack of industry-specific use cases and optimized configurations make it difficult for developers to seamlessly integrate these tools into games.
In addition, API stability and the maximum number of conversation tokens (Token) restrictions, etc., all hinder the integration of AI technology into the game industry to varying degrees. Even with advanced tools like GPT-4, Token limitations still restrict the depth and length of conversations between users and NPCs, which causes the AI to often forget the original conversation with the user during the conversation. In commercial applications that pursue stable performance, its negative effects will only become more prominent. After all, long and complex dialogue is crucial to in-game storytelling and character development.
How to solve these challenges is critical to fully unleashing the potential of AI in creating truly immersive, dynamic and highly interactive gaming experiences.
The future of AI applications: meeting the needs of developers and industries
In this regard, many domestic and overseas companies, including large AI model companies, are exploring feasible paths, such as Claude2.1’s 200K maximum Tokens, NVIDIA’s ACE for Games, etc. On January 9, domestic AI startup Baichuan Intelligence also took an important step towards smart NPCs, releasing the Baichuan-NPC character model and Baichuan character creation platform, bringing new possibilities to the integration of AI and the game industry, arousing heated discussions in the industry. .
Baichuan Intelligent Character Creation Platform Official Website
The main reason why it has attracted widespread attention is that it solves the various shortcomings mentioned above. This also raises a question that needs to be re-explored: What qualities do AI large models and platforms need to have when they are commercially applied in the game industry?
First, for intelligent NPCs, the most important issue is whether the acting is “like or not”.
Just like talented human actors can shape characters well, whether an intelligent NPC can act resemblingly depends first on whether the basic model is smart enough, that is, whether the model’s general intelligence level is sufficient. The intelligence level of the model is usually determined by pre-training. However, compared with general models, game intelligent NPCs are vertical large model applications, which means that the large models behind them not only require complex general text data, but also require a large number of high-quality dialogues, rich Plot narration and other content, and targeted training around “role playing” to demonstrate a high sense of personification in the dialogue.
In order to better improve the model effect, Baichuan Intelligence optimized the base model from multiple dimensions as early as the pre-training stage. On the one hand, it collects massive industry websites, high-quality books, and high-quality script data to pre-train Baichuan-NPC with more than 3T Tokens in domain knowledge; on the other hand, Baichuan-NPC innovatively uses artificially synthesized data to enhance domain knowledge in the pre-training stage. , greatly improving Token utilization efficiency.
A very important point in the pre-training process is that due to differences in data proportions, the growth curves of the model’s basic capabilities and each special ability in the pre-training phase are actually not consistent. In other words, even if a certain amount of high-quality novel dialogue, script dialogue and other data are added to the pre-training, the “playing” effect of the role model will not necessarily improve simultaneously with the model’s general ability. In this regard, Baichuan Intelligent has proposed a course learning method that focuses on ability balance, dynamically regulating the data distribution at different stages in an adaptive manner, and maintaining a balance between basic model capabilities and professional abilities. The popular understanding is that this kind of course learning method can make the model’s academic performance and literary, artistic and sports performance improve together and develop in a balanced way like a human being.
It is of course important to improve the capabilities of the base model, but role models often face the obstacle of “identity consistency” during their implementation. That is to say, NPCs built based on universal language models often have character knowledge, background common sense, behavioral habits, and tone styles that deviate from the character design during the dialogue process. For example, when we asked the big model “Li Bai” “What should we pay attention to when flying an airplane?”, he would talk about the modern thing “airplane” instead of telling us that he didn’t know what an airplane was.
This phenomenon mostly stems from two aspects: first, the alignment goals are inconsistent, and large language models often aim to serve as qualified assistants instead of playing “roles”; second, the model’s thinking ability is insufficient and cannot truly take on the role setting.
Inspired by a number of technologies, Baichuan Intelligence’s first “Thinking Chain Alignment Technology” has largely alleviated this pain point. That is, during the model training process, the data construction method with thinking chain and the enhanced alignment method with thinking chain alignment are used to greatly improve the consistency of the characters. This significantly enhances the model’s basic dialogue capabilities and role interpretation capabilities, making it closer to humans.
Second, based on the data set, if AI wants to truly create a realistic game character, it not only needs to achieve “real performance” through technical means, but also uses engineering means to optimize the character construction and deployment process and reduce character customization. comprehensive cost.
According to Game-Time, intelligent NPCs are not only tools to create realistic interactions, but also promote the development of complex story lines in the game. Beyond dialogue, AI is more conducive to creating characters whose actions, reactions, and emotions are indistinguishable from real humans. The latter includes non-verbal cues and behaviors that add depth to character interactions.
Currently, character construction or reproduction in the industry is mainly achieved through API calls. In actual creation, multiple departments such as product, operations, and technology often need to collaborate and debug repeatedly. The development threshold is high, the cycle is long, and the efficiency is low. The final character effect It’s still difficult to guarantee.
So, how to obtain a high degree of freedom, low-cost, zero-code customized solution? Baichuan Intelligence further launched a development kit composed of “character creation platform + search enhanced knowledge base”. It can effectively improve the freedom of character customization through its two unique technologies of self-developed strong multi-round alignment and search-enhanced knowledge base.
Currently, the Baichuan Intelligent Character Creation Platform provides 30+ government-created characters and is still expanding.
For example, the strong multi-round alignment technology carefully designed the role setting field in System Prompt. To put it simply, whatever characteristics the user defines in System Prompt, the generated character will “perform” according to the user’s settings.
The competitive, nagging, timid Sun Wukong
In addition to “performance” ability, knowledge ability is another key factor that determines whether an AI character is “real”. Only by having knowledge consistent with his identity can the character “perform” without immediately “acting out”, but this often has a very high development threshold.
After deeply combining search capabilities with role knowledge capabilities, Baichuan Intelligence optimizes database construction and recall algorithms for role-playing scenarios, providing a flexible “many-to-many” association method for roles and knowledge databases. In other words, developers can enrich the “connotation” of the character by uploading the knowledge documents required for the character and customizing the response methods, ensuring the freedom of customization while greatly reducing the difficulty of development.
Third, while being targeted, AI tools also need to meet the customized needs of developers and different companies.
AI should allow developers to quickly define and create characters specific to their projects, giving developers and players more control over character creation and interactions. In terms of commercialization solutions, AI companies should also provide more simple, easy-to-use and low-cost solutions, so that AI can transform from a toy of large companies into a weapon that can truly benefit the entire industry.
Baichuan Intelligent’s character creation platform is an excellent example. After the character is created, the platform not only integrates the character parameter fine-tuning options and character dialogue effects into the same interface to achieve “what you adjust is what you see”, but also achieves it through “zero code” “What you adjust is ready to use.” Even after the role debugging is completed, the role code can be connected to the business scenario with one click to complete the deployment. Flexibly and efficiently integrate into various product development processes to create lifelike intelligent NPCs.
After role debugging is completed, you can deploy it to the business scenario with one click.
Conclusion
In movies such as “Westworld” and “Runaway Player”, NPCs with “consciousness” interact with the game world and are regarded as the ultimate form of games. With the rapid rise of generative AI and large models, this dream no longer seems far away. NPCs given “souls” by AI are profoundly changing the game industry.
But currently, due to limitations in technology, capital, and talent, this new trend only stays among leading companies. The emergence of Baichuan intelligent character large models is expected to change this situation. Its convenience, ease of use, and high degree of customization mean that smart NPC is no longer a competition ground for big manufacturers, but can also benefit the entire industry and small and medium-sized developers.
“Baichuan Character Model + Character Creation Platform” will promote game creation to achieve richer story expressions, more novel interactive methods, and promote the transformation of traditional narrative experience. The deep meaning conveyed behind it is not only conducive to the implementation of more creative ideas, but also accelerates the development of the game into the next cycle.
Baichuan character model official website:https://npc.baichuan-ai.com/index
