The biggest problem with AI games is on the table: in the end, they're just not fun enough
Three guests circle back repeatedly to an awkward consensus: Agent games are a hot concept, but the industry hasn't cleared the most basic hurdle—whether they're actually fun—let alone tackled retention or monetization.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
AI games are no longer a concept—they're about who executes deepest
Shao Yun argues that ‘AI-native games’ is no longer a zero-to-one proposition—everyone building games is using AI now, so that label has lost its currency. The real game is linear: who integrates AI capability into their content pipeline more deeply, more effectively, more quickly. That still hinges on the designer's taste and judgment about content value, the same things traditional game production has always tested. AI doesn't bypass that.
— Shao YunDeepSeek's open-source release plus Vibe Coding unlocked new game mechanics
Shao traces the tech timeline: Midjourney-class models solved art assets. O3-class models solved conversational understanding. But those were expensive and slow, too costly for consumer games. The real pivot came when DeepSeek open-sourced its language model—character-card PDFs flooded platforms like Xiaohongshu. Then Vibe Coding emerged earlier this year, and for the first time, ‘game-action primitives became AI-generatable and executable.’ The core mechanic broadened from pure chat to AI-driven, AI-executable behavior itself.
— Shao YunTraditional behavior trees can generate emergence that AI may not replicate
Shao's core claim: Agent systems are not a prerequisite for games. A cat in Dwarf Fortress dies of alcohol poisoning from drinking vomit. A Tears of the Kingdom player builds machines no one has ever imagined. Neither needed generative AI. He uses gravity as a metaphor: generative AI's value is like Newton observing a falling apple and deriving universal gravitation. But once the law is known, ‘the efficient path is just applying the formula, not re-deriving it from first principles each time.’ Use AI where it helps. Forcing it where it doesn't damages emergence, not enhances it.
— Shao YunToo few bugs in AI-generated content exposes a deeper problem
Shao raises a counterintuitive claim: ‘AI tends to produce too few bugs. Emergence is too weak. Regression too strong. What emerges is banal.’ AI is fundamentally a probability-regression model, so its outputs cluster around the average. But content creation demands the opposite: ‘I need something that scores 120 out of 100—fresh, creative, emergent.’ A hundred mediocre moments waste players' time. The ‘bugs’ people complain about—unintended sequences, unexpected consequences—are exactly what players share and discuss. AI-generated content regresses toward the bland.
— Shao YunThe designer's subject choice is Zhongzhen Simulator's real competitive moat
All three agree: Zhongzhen Simulator's strength isn't the AI tech. It's the choice to center on Zhongzhen—a tragic emperor whose individual decisions seem sound, but compound into catastrophe, naturally loaded with argument and imagination. That's the designer's judgment call, and ‘AI can't replace that. Only the designer chooses what to make. AI just executes.’ But Shao flags a durability problem: the mechanic of issuing decrees hasn't been decomposed into a progression loop with scaling difficulty and player growth.
— Shao YunToken spending isn't subscription; what you're selling isn't tokens
Zhongzhen's token model is often compared to legacy online-game point cards, but Shao highlights a crucial difference: point cards capped spending at a 24-hour ceiling; token consumption has no ceiling. A heavy player can spend 5,000 times what a casual player spends in a day. He reframes: business design should first ask ‘What am I selling, and to whom?’ then ‘How do I price it?’ Games never priced by cost—not networking fees, not server load. ‘We price by your experience.’ Zhongzhen hasn't yet designed a retention or identity loop for high-spenders to justify that 5,000-fold spend gap.
— Shao YunThe phone-scam game proves AI makes real moral hesitation a playable mechanic
Xiaoning describes a Gamescom demo where players roleplay as scam-call newcomers, using voice to talk targets into sending money. Over four days, dozens to hundreds tried it. The designer revealed only six actually closed. Many said they couldn't bring themselves to continue; they didn't know what to say next. Xiaoning sees this as a proof-of-concept for good game design: ‘Whether it's 'I can't bear to' or 'I don't know what to say,' you've already cast the other party as a person with emotional stakes. That's why it triggers real feeling.’ That real-time, voice-driven human response—the doubt, the hesitation—only AI's reactive generation makes playable.
— Xiao NingGPT-6 is a game engine—but that doesn't mean monopoly
Shao challenges the framing: ‘A game engine isn't an engine. It's not an aircraft engine either. A game engine is actually an assembly line for building cars.’ Engines are singular tools. Game engines are technical systems plus organizational structure and capability combined. History offers no precedent for a single company monopolizing that bundle. Large models are already game engines—and they're actually better paired with Godot than Unity or Unreal, since graphical interfaces designed for humans are token-expensive and awkward for AI operation. ‘There's no scenario where one company corners this.’ The future points toward open ecosystems, not convergence to a single foundation.
— Shao YunIn their own words · checked verbatim
What we call ‘AI-native games’—AI games—are no longer a zero-to-one question. It's become linear: who deploys it deeper, better, faster. The real test is whether you can twist AI capability into your own content-production pipeline.
所谓AI原生游戏 AI游戏 它已经不是一个0到1的命题了 它是一个线性化的命题 就是谁用的更深 谁用的更好 谁用的更快 你自己能否把AI的能力 拧入到你自己的 内容生产管线那边去
Shao Yun9:05
Generative AI's value is like Newton seeing an apple fall and deriving gravity. But gravity, for us, is already a known physical law. When I need to use it and think about efficiency—execution efficiency—I don't need to re-derive gravity from a falling apple each time. Gravity is just a formula I apply directly.
就我们可以想象 就生成式AI的作用在于 我像牛顿一样 看到苹果落地 让他推出了万有引力 但是万有引力 对于我们来讲 是一个已知的物理定律 我现在要用到 我要考虑到效率 其实就运行效率 其实我不需要重新 从苹果落地去推断一次万有引力 万有引力 我直接拿公式过来用就行了
Shao Yun31:22
What AI generates now isn't buggy enough, isn't emergent enough—regression dominates. What emerges is pedestrian. A hundred pedestrian outputs waste everyone's time. What's needed is something that scores 120 out of 100. That's what players want: fresh, creative, emergent content.
现在AI 往往做出来的东西 在我看来叫bug不够多 涌现性不够强 回归性过于强了 就是出来的东西 是平庸的东西 100个平庸的东西 在浪费大家的时间 我需要的是一个 120分的东西 这是玩家要的内容 就是新鲜的 有创意的涌现的内容
Shao Yun36:25
Characters can lie; they're open. But world state can't. So the model's job is to understand and propose actions. The rule system underneath—the designed behavior logic, the code—decides whether that action is allowed to continue and what consequences remain to shape the whole world, the game.
角色是可以说谎的 他是开放的 但是世界状态是不能说谎的 所以模型需要负责去理解和提出行动 然后整个世界这个背后的这个规则系统 这段 就比如设计好的这些行为术 这些代码 许定这个行动 能不能继续发生 然后以及发生后 留下什么样的后果 去影响整个世界 整个游戏
Lü Chun41:26
This is the designer's taste—their unique taste. The choice they make is this. AI can't replace that, because only the designer can decide what to make. AI just executes. That's the core.
这就是所谓 我就说叫做 制作人自己的品位 它的独特的品位 他要做的选择就是这个 这个是AI替代不了的事情 因为只有制作人 能决定要做什么 AI只是替你去做 所以这个是 这个最核心的事情
Shao Yun57:33
I'm not selling tokens. Games have never been about selling tokens. Did they used to sell network fees? We don't price by cost. We price by your experience.
我卖的不是token 游戏从来不是mai token 游戏难道原来卖的是网费吗 我们不是按照成本来定价的 我们是按照你的体验来定价
Shao Yun1:03:36
Whether it's ‘I can't bring myself to’ or ‘I don't know what to say’—you've already cast the other person as a person, someone with emotional stakes. That's why it can trigger real emotional response in you.
但我觉得不管是 不忍心还是不知道 其实你都已经 把对面当成了一个个体 它是有情感附加值 在里面的 所以它能唤起你的一些 情感的波动
Xiao Ning1:16:45
It's not an engine. Not an airplane engine either. A game engine is actually an assembly line for making cars. A game engine is technology combined with organizational structure and capability. So there's no scenario where a single company monopolizes it.
它不是发动机 它也不是飞机发动机 事实上的游戏引擎 是生产汽车的流水线 你要理解这个概念 这什么意思呢 就是说游戏引擎 它是一套技术结合组织结构 组织能力的这样的一件事 所以它不存在哪一家公司 会垄断这个东西
Shao Yun1:31:55
Figures
| Phone-scam game demo—players who completed it | 6 | 1:16:45 |
| Morningstar Sequence—estimated hours to complete | 200+ hours | 1:45:05 |
| Liexiyuanzhen demo—official beginner experience length | 15 minutes | 1:46:07 |
| Liexiyuanzhen demo—one visitor's actual playtime | 10+ hours | 1:46:07 |
Glossary
- Agent
- An AI program unit with perception, decision-making, and autonomous action capabilities.
- Multi-agent
- A system architecture where multiple agents coordinate or compete to solve problems.
- Vibe Coding
- Rapid code generation through natural language conversation with AI.
- Behavior tree
- A traditional game-AI structure that organizes character decisions through rule-based nodes.
- Emergence engineering
- A methodology for deliberately designing systems that leverage model emergence capabilities.
- Godot
- An open-source game engine noted for its lightweight interface and AI-friendly operations.
How to listen
Game designers, founders tracking where AI applications hit their limits, and investors learning how Agent games are built and monetized.
After 1:43:22, the episode pivots to game recommendations unrelated to Agent games; skip if you're focused on AI.