giant network: big models drive innovation in gaming paradigms, and “game ai” 2.0 moves from concept to reality
2024-09-23
한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina
on september 23, ding chaofan, head of giant network's ai lab, attended the yunqi conference and gave a keynote speech. he revealed for the first time the technical details of the self-developed big models giantgpt and bailing-tts, and said that the big models have promoted the innovation of the game paradigm, and "game + ai" 2.0 has gone from concept to reality.
giantgpt is good at role-playing and "playing games"
at this year's yunqi conference, giant network launched two self-developed large models, giantgpt and bailing-tts, and their applications. among them, giantgpt is one of the first large models in the gaming industry to complete filing. since last year, giant network ai lab has continuously iterated and optimized it.
ding chaofan introduced that giantgpt has excellent role-playing capabilities, vivid situational reasoning capabilities, customized long- and short-term memory, and the ability to deeply support game scenarios. it can be called a large role-playing model that "can play games."
data is the core of big model capabilities. giant network has built a huge data set based on public data on the internet and its own data. it has the advantages of large scale, diversity, and high quality, and forms a complete and efficient data production chain.
currently, giantgpt has been implemented in many giant network products such as "zhengtu". the companion-type intelligent npc built based on giantgpt has long-term memory of personality, emotions and adaptation, providing players with an intimate companionship experience and improving user retention.
in addition to the large language model, giant network ai laboratory also launched the industry's first large tts model that supports mixed speaking of mandarin and multiple dialects - bailing-tts this year.
at present, the speech synthesis big model technology has made significant progress in the field of mandarin, but the development in the field of dialects is very slow and cannot meet the diverse speech synthesis needs. china has dozens of major dialects, each with unique voice features and grammatical structures, which makes it extremely challenging to train a tts big model covering various dialects. in addition, the scarcity of dialect corpora and the lack of high-quality annotated data have further increased the technical difficulty.
in order to solve this problem, giant network ai laboratory built a mandarin and dialect dataset covering 20 dialects and more than 200,000 hours based on the chinese dialect system, and proposed a number of model-based technical innovations, enabling bailing-tts to achieve zero-sample cloning of mandarin and high-quality dialect speech and peking opera singing synthesis effects.
big models reshape gaming experience and productivity innovation
in his speech, ding chaofan demonstrated giant network's series of explorations in the application of large models, including ai painting platforms, ugc script creation tools, anthropomorphic intelligent question-and-answer systems, ai native game play, etc., reflecting the innovation of large models in gaming experience and productivity.
the one-stop ai painting production platform "giant mojing" focuses on supporting team collaboration. at the same time, it integrates a number of self-developed ai visual algorithm capabilities into the form of workflows to build a collaborative standard ai art production pipeline. without the need for frequent import and export or software switching, complex tasks can be completed on the same platform, thereby improving creative efficiency. at the same time, it integrates a one-click workflow to simplify a large number of complex operations, making it suitable for large-scale art production work.
based on the capabilities of multiple self-developed large models, the ai laboratory has also created an intelligent editing and creation platform for promotional videos, providing automatic analysis and style matching of hot videos. it combines the script large model with tts voice cloning capabilities to achieve the effect of one-click film production, greatly improving the production and creation efficiency of promotional videos.
the ultimate goal of implementing big model technology is to reshape the game experience and promote innovation in gameplay, and the "space killer" project has actively explored this. the game's ugc script creation tool introduced ai big model writing and tts functions, thereby lowering the threshold for content creation and stimulating players' enthusiasm for content creation; the ai native game play "ai endgame challenge" made players very "addicted", driving the game's related index on short video platforms to double, and a large number of players spontaneously shared various interesting gameplay and strategy skills.
the core of the "ai endgame challenge" gameplay lies in giant network's self-developed multi-agent framework design, which includes two major features: "collaboration" and "competition". how to build a balanced strategy based on a control system is the key to forming a high-quality collaboration and competition paradigm. in addition, because players need to be deeply involved in it, we must focus on the flexibility and freedom of players in performing tasks, as well as a good operating mechanism to ensure the reasonable evolution of the game process.
“game + ai” 2.0: from concept to reality
if productivity improvement is the "game + ai" 1.0 era, then the gameplay innovation based on aigc technology has pushed "game + ai" into the 2.0 era.
ding chaofan believes that "game + ai" 2.0 will create a future form of game: a non-linear world that can break the constraints of traditional rules, where the environment is updated based on player data feedback, the plot design is dynamically extended, and there are random events triggering it, giving players extremely high freedom and even providing players with customized game content.
“what’s exciting is that we see the possibility of this kind of game form in the ‘ai endgame challenge’ gameplay, where players can influence the environment and change the direction of the story through their own actions, and the whole process has a high-quality interactive experience. it is not only a successful technological breakthrough and attempt, but also realizes the advancement of a new game form from concept to reality.”
at present, giant network has built a full range of basic capabilities with self-developed large models as the core, covering large language models, visual content generation, voice generation, and ai agents. large models such as giantgpt and bailing-tts have achieved large-scale application in core game business scenarios, and have been deeply applied to various links such as game development, operation, distribution, and testing, forming an efficient production chain closed loop. at the same time, combined with the large model capabilities to go deep into the core gameplay level of the game, a companion ai smart assistant, emotion-driven decision-making ai, and a new game paradigm based on multi-agent large models have been created.
looking to the future, ding chaofan emphasized that giant network ai lab will more aggressively explore game play innovations driven by large models. "i think it is not far away for a native game world to emerge that interacts deeply with players and naturally produces lasting and highly attractive content."
