GameBot's CEO Kakar Liu recently spoke to GeekPark, a Chinese tech media, on how the company is evolving game AI from reinforcement learning to LLM-powered agents. He highlighted efforts to create human-like, emotionally engaging AI characters and the Game NPC Ecosystem behind demos like Living Chang’an City. Kakar also reflected on the challenges of applying AI in game development and the importance of team focus and resilience.

Read the translated full interview below.

W. Brian Arthur, a founding figure in complexity science, once said, "New technologies arise from the combination of existing technologies."

GameBot was founded in early 2019, at the tail end of the peak of what is now considered the previous generation of AI, focused on pattern recognition. A wave of hype followed. At that time, OpenAI and DeepMind were still focused on teaching AI to play games using reinforcement learning (RL). RLHF, the core method now used to fine-tune LLM outputs, also emerged during this period.

Today, as the marginal returns of the Scaling Law diminish, the tech world is once again turning its attention to reinforcement learning. As a result, the combination of Scaling Law and RLHF has become one of the most closely watched technical paradigms.

Kakar Liu, founder and CEO of GameBot, previously led Tencent’s Chinese Go AI "Jueyi" and the Honor of Kings AI “Juewu." It was through training AI to play games using RL that he gained confidence in AI’s potential.

“It was like teaching a child—watching it go from knowing nothing to gradually mastering skills...and ultimately doing things humans couldn’t. It was a truly mind-blowing experience,” he said.

It was that early confidence in AI that led Liu and his team to leave Tencent and found GameBot. Since then, public confidence in AI has fluctuated—falling from a peak into a trough, only to surge again with the rise of AI 2.0.

With the emergence of LLMs, discussions around game AI have reignited. Some new-generation entrepreneurs have labeled GameBot as part of the "previous generation." Indeed, the company has remained relatively low-profile in recent years.

Recently, Kakar Liu spoke with GeekPark in an interview, reflecting on GameBot's journey—from focusing on RL to LLM+ RL. He discussed the company's ongoing efforts to develop AI-native games and shared his views on the future of game AI.

Kakar Liu, Founder & CEO of  GameBot

01

Game Agents: From Outperforming Humans to Human-Like Interaction, and Finally Emotional Resonance

GeekPark: Since the advent of LLMs, there has been a growing discussion about how AI is transforming the gaming field. What changes have you observed in this area?

Kakar Liu: I’ve observed two main paths.

One is whether we can use LLM and AIGC to create more differentiated content at a lower cost. Currently, we are already seeing efforts to produce various types of content. The other direction is focused on GameAgents—whether we can make bots and NPCs in games feel more alive.

In the past, we used reinforcement learning to make bots extremely skilled. However, this was only one aspect of the gaming experience. Players still desire more differentiation and a more human-like experience.

Many people believe that adding AI capabilities to a game automatically makes it an AI game. We see this as diluted — or an overly generous interpretation of what constitutes an AI-native game. There should be clearer definitions. In AI games, AI should be integrated into the core architecture of the game, not just added as a peripheral element and then called an AI game.

GeekPark: How do you interpret the “human-like” experience in games?

Liu: In the past, we used reinforcement learning to make in-game bots extremely powerful, sometimes even stronger than human players. However, for players, they don’t want bots that completely crush them. They want the bots to be kind and empathetic. For example, when I encounter a teammate or character in-game, I’d like them to follow my lead and support me in meaningful ways.

We previously combined Supervised Learning (SL) and Reinforcement Learning (RL) to solve the problem of making bots more "human-like" in their actions.

For example, there are physical limitations in human actions—such as the inability to instantly turn 180 degrees or the need for 0.1–0.2 seconds to make a decision. These limitations don't apply to AI.

By modeling, we can incorporate the constraints and behaviors found in human actions into the game world. This was one of the challenges we frequently tackled.

GeekPark: After the emergence of LLMs, how do you think this technology will enhance the gaming experience?

Liu: Players also hope that agents behave more like humans in the game world, with richer interactions. This means that agents should not only respond to player commands but also proactively make requests to players.

Before LLM technology came out, it was difficult to meet the need for interactivity with the previous technology stack.

Our goal has always been to improve the autonomy and interactivity of agents. Before 2021, we focused on improving autonomy, enabling AI to make decisions in complex scenarios. In 2021, we released "Orion α," which broke through GameAgents in 3D environments, allowing AI to play shooting games. This was a global first.

GeekPark: After the emergence of LLMs and research like Stanford's Small Town, people have started imagining more specific forms of agents in games.

Liu: Nowadays, it's not enough for agents to simply make autonomous decisions. They also need to have interactivity, which means they must generate enough content through interactions with each other. To demonstrate this, we created a demo called Living Chang'an City, featuring characters from different backgrounds and professions, with complex relationships forming a self-operating small society. Alongside this demo, we also released our Game NPC Ecosystem Technology.

Back in 2019 and 2020, we discussed the concept of agents. At that time, the word "agent" was translated as "proxy”, and it was only in the past couple of years that we began using the term "intelligent agent."

In that context, introducing it as "proxy" would have left people confused. Therefore, we opted for the term "GameBot" as a temporary solution.

GeekPark: How do you envision the future development of agents?

Liu: In the future, we will see more and more agents in our daily lives, such as robotic dogs, drones, and other devices. In a sense, robots are a type of agent, and the future of hardware and embodied intelligence will continue to advance.

Specifically in gaming, agents will step out of the virtual world. Through extensive interactions with agents in games, players will deepen their understanding of them, and strong emotional bonds will form. These bonds can extend into real life.

When entities from games, such as characters, appear in the real world, they could become advanced companions in your life. They will create "Both Reality" with you, exploring the game world together and interacting with you in real life. We believe the future will be a world where one billion people live alongside ten billion AI agents.

02

Game Development is a Complex System Engineering Process

GeekPark: How should we understand the "Game NPC Ecosystem Technology" behind the Living Chang'an City virtual city demo?

Liu: In breaking down the technology for agents, our approach is similar to most teams — we divide it into different modules, such as Control, Plan, Memory, and Reflection. These modules form the underlying architecture of each agent.

Additionally, for the ecosystem to operate more effectively, agents need to interact with each other more smoothly. This requires higher-level design. You can think of it as an "event trigger," where top-down control is applied. Without this top-down control, relying solely on emergent behaviors from the lower layers, the generated content would quickly become monotonous.

The most challenging part is determining the level of control at the top — in other words, event triggering. Once an event is triggered, it can alter many states or values in the game world, which in turn affects the agents at the bottom.

It's similar to the relationship between individuals and countries, or the geopolitical dynamics between nations. A top-level event can change many things, which are then transmitted down to the individual agents, affecting their actions. Before large models emerged, or when they were not as powerful, implementing event triggers was quite difficult.

GeekPark: Does this require writing a lot of rules?

Liu: It relies heavily on planning and requires many adjustments. However, after the release of models like GPT-4, combining top-level event triggers with LLMs has proven to be quite effective. Firstly, significant events don't occur very frequently and don't change constantly. Secondly, when changes occur, it takes time for these changes to be reflected in the behavior of the lower-level agents.

GeekPark: The interaction between the lower-level agents and the top-level trigger is crucial. How does that work?

Liu: I think the most exciting part is this entirely new system. The interactions at the bottom level can also influence the logic of top-level event triggers. If it were purely bottom-up, it would be like giving a starting point without knowing where it will go — it would be completely uncontrollable. However, if I provide several key nodes, like a, b, c, d, and e, these nodes are fixed. How they interact with each other is left for the AI to decide, while ensuring the overall storyline makes sense.

GeekPark: Has this new technology framework been applied to any specific game development projects yet?

Liu: Recently, we've been collaborating with a Steam game team to integrate this technology stack into a commercial game.

The game involves space merchants and management gameplay. The basic premise is that there are many planets in space, each with several agents. The relationships between countries, like Country A and Country B, will influence the business dealings between the merchants. For example, if the relationship between the two countries becomes tense, Country B might suddenly stop trading with Country A or increase prices.

GeekPark: What is the biggest challenge when applying this new technology framework to game development?

Liu: It has caused a significant disruption to the existing game development pipeline.

Today, game development tends to focus on certainty. However, game development itself is a very complex system with many different roles and professionals involved, and it requires tight project management.

The biggest feature of AI is uncertainty. Making a management system that demands high certainty compatible with uncertainty is a very difficult task.

For example, during our collaboration, there were times when both the AI and game design teams were frustrated. Figuring out how to reconcile both sides was a huge challenge.

This also explains why AI-native commercial games haven't been released yet. It's not just about having an idea — when it comes to implementing it with the team, it often drives people crazy. Either the game team goes crazy, or the AI team does.

GeekPark: Based on GameBot's experience, what insights can you share?

Liu: The team is critical. The requirements for team members have become more demanding. We now need people who truly believe in the project, those who, after going through a period of madness, can then calm down and begin discussing how to solve problems. For the entire project, patience is essential.

In this process, besides the uncertainty of AI, the overall game experience is also uncertain. This makes the process different from traditional games, where milestones are clearly defined and expectations for the next version can be set.

When it's difficult to establish expectations, maintaining team cohesion becomes crucial. It requires inner strength.

03

Resilience is a Matter of Chance

It's something you inherently possess.

GeekPark: The outside world often views GameBot as a game AI company, but it seems that this doesn't align with how GameBot sees itself. Can you elaborate?

Liu: We're somewhat helpless about this. The "game" label may be heavy on us, but we only see games as a scenario where AI technology can rapidly land and be implemented.

Previously, games were a good fit for reinforcement learning and unsupervised learning iterations. Now, the LLM+RL tech stack is still suitable for iteration in GameAgent scenarios.

GeekPark: From a commercialization perspective, GameBot is essentially the largest third-party AI NPC provider.

Liu: As a technology company, we don't have the luxury of being supported by a big corporation like DeepMind. So, we need to think about commercialization, and we've done quite well in the gaming field.

We are also trying to apply our technology in other fields, such as using the agent technology in low-altitude traffic, including autonomous driving. Another important direction is agent-based simulation. For example, we're working with a university to simulate ancient cities and model human behavior within economic systems.

Many complex systems cannot be reasoned through inductively. Using AI technology to simulate certain scenarios and then translating these back to real-life situations can help solve some problems.

GeekPark: What do you think of Google's AI game engine released this summer?

Liu: If they intend to create a new game engine to replace the existing ones, I think it will be very difficult. Today's game projects are already incredibly complex, with very intricate pipeline management. There are many different elements involved, such as content, operations, and commercialization. If you want AI to try an end-to-end approach, it would make collaboration among people much more difficult. Even today, when generating videos, people still need to handle post-production to make it usable.

If it's an enhanced version of video generation, adding interactivity on top of just presentation, I think it's a good research project. It would help AI better understand the physical world. Content generation requires higher consistency, and if we can really solve that consistency problem, it will certainly deepen AI's understanding of the physical world.

GeekPark: What do you think about the real-time interactive world model, Oasis, recently developed by a startup?

Liu: Many people are now trying to use natural language chat to create a game, and I'm following some of these projects myself. My biggest takeaway is that writing code isn't necessarily more complex than using language.

GeekPark: But it lowers the threshold.

Liu: I tend to believe that in the future, content produced through language interaction will struggle to create very complex systems. Describing something complex with language is difficult, and the efficiency is low.

GeekPark: This perspective may relate to GameBot's experience as a mature technology provider. Apart from technical breakthroughs, you also emphasize engineering stability, efficiency, and other factors.

Liu: I'd like to elaborate on this. Many times, when people talk about technology, they focus more on algorithms, such as the interest in PPO and DQN algorithms in reinforcement learning.

However, with the emergence of large models, people have started paying more attention to engineering. For example, achieving parallel processing with 100,000 A100 or H200 GPUs presents significant engineering challenges.

From my perspective, this is just the offline aspect of engineering. Another very important aspect is the online side — how to deliver services with models.

For instance, we need to provide services globally. In many places, especially in South America and Southeast Asia, the IDC lacks the capability to deploy the latest GPUs. In such cases, providing real-time services itself becomes a massive challenge. We have been working on solving issues like how to make compressed small models run on CPUs from ten years ago, performing extensive compression and predictions, and achieving low latency even in regions with poor network quality.

It's great to see that people are now focusing more on engineering. I also believe that with the capabilities of large models accumulated over the past two years, as they are scaled for real-world services, high concurrency and low latency in online scenarios will present many significant challenges.

GeekPark: As an entrepreneur, you've certainly faced many challenges over the years, with ups and downs. Are there any insights you'd like to particularly share?

Liu: One thing that really surprised me about myself is that the mistakes I thought I would never make, I ended up making twice. From a rational perspective, you think you would never make these mistakes, but you still do — it's a lack of focus.

Everyone knows that in entrepreneurship, focus is essential, and you must use limited resources to work on the most important things. That's something we all understand, but I still made the mistake. I think the reason is that, in my previous experience at a large company, there was plenty of support and resources, and we were riding the wave of trends. Coupled with a strong team, for a long time, we had the mindset of "there's nothing we can't do, only things we haven't thought of yet." This led to an unconscious over-optimism.

But after April of this year, we deeply realized the importance of humility. After going through those lessons, I now remind myself that controlling our actions and staying focused without spreading ourselves too thin is a very difficult, but incredibly important task.

GeekPark: I've heard that you often write internal letters to the team. What was the focus of your most recent letter?

Liu: The last internal letter I wrote had the title, "Resilience Like a Song, Dancing with the Wind." Resilience is something that's very hard to learn by effort alone — it's something you're naturally equipped with by chance.

In that letter, I also shared a line from an English article I recently read: "Whispers of Resilience: Dancing with the Winds of Change."

It's hard to capture the full nuance of that phrase in Chinese. What touched me most was “Whispers of Resilience” — it captures the idea of enduring hardship with a kind of quiet strength: soft yet confident, a lasting force from within. It reminded me of Homer, the ancient Greek bard. His resilience mirrored that same quiet power. Though blind and alone, he created the timeless epics The Iliad and The Odyssey. He wasn’t defeated by fate — instead, he used poetry and music to depict war, exile, hope, and courage. His verses, like whispers in the wind, carried grace through adversity. Through his art, he brought light to the world — showing us that even in the face of storms, we can still dance with destiny.