It’s a question that’s been buzzing around the gaming world for a while: why haven’t we seen a truly groundbreaking, AI-native game yet? So many have tried to blend artificial intelligence with gaming, but the magic hasn't quite happened. As the legendary Shigeru Miyamoto once told Nintendo's former CEO Satoru Iwata, "Everything is good about this game, except it's not fun." That sentiment cuts to the core of it, doesn't it? At the end of the day, games need to be fun to be played.
I remember back in 2014, during my sophomore year at Tsinghua University. Inspired by the viral hit "2048," I created a number-matching game called "Not a Number." It actually took off on Renren, attracting nearly 20,000 players in a week. Seeing my classmates engrossed in my game during lectures was an incredible feeling. Later, I was even invited to create a problem for the National Olympiad in Informatics, a puzzle based on my elimination game. The idea was to connect numbers to form primes, palindromes, perfect numbers, or squares for points. It became quite the craze among students at Tsinghua and Peking University.
Fast forward to today, after graduating from MIT with a Ph.D. in computer graphics and co-founding Meshy AI, a leader in 3D AI technology, I find myself drawn back to my roots. Why did I get into programming and graphics in the first place? To make games. Entrepreneurship has its challenges, but with the company's success, the future is bright. Yet, no matter how far I venture, I always knew I'd return to that childhood dream: creating a game. And with Meshy being an AI company, this feels like the perfect moment to build an AI-native game.
As Satoru Iwata famously said at GDC 2015, "In my heart, I am a gamer." While I might not be on his legendary level, that sentiment resonates deeply with me.
So, has AI truly made games more fun yet? Let's look at some of the common approaches:
- World Models (like Genie 3): I spent $200 exploring Genie 3. While it's a fascinating landscape simulator, a truly fun game needs more than just pretty scenery. It requires mechanics, aesthetics, and narrative.
- AI Dialogue (Character.ai): This pulls out the NPC interaction aspect of games into a standalone product. It's popular, but not quite my cup of tea for a full gaming experience.
- AI-Generated Assets (UGC): Integrating Meshy's API to let players create their own weapons or gear. We've seen players drawn to the "AI" label, but the novelty wears off quickly if the core gameplay isn't engaging. Retention often suffers.
- AI Game Generation (Rosebud AI): Essentially "vibe coding" for games. While it can generate users, the overall quality of games on these platforms is currently too low to be a serious contender.
- AI Towns (Stanford Town): Making NPCs more lifelike. It's a good direction, but as a player, I often feel a lack of deep engagement.
- AI Story Generation: This has seen some small-scale successes and is a promising avenue.
Our team has a simple rule: we must love playing the games we make. If we wouldn't download it, if we get bored after two minutes, then something is fundamentally wrong with the direction.
Breaking the Mold: Back to First Principles
Why haven't these AI + game attempts yielded a phenomenon? Our analysis points back to the fundamental truth: for most games, the key is simply "fun." And current AI integrations haven't fundamentally made games more fun.
To break through, we need to ask: what makes games fun at their core? And how can AI enhance that essence?
Consider two games I deeply admire: Jonathan Blow's "Braid" and "The Legend of Zelda: Breath of the Wild" (BotW). Looking at their prototypes, the core gameplay mechanics – time manipulation in "Braid," or the physics-driven interactions in BotW – are evident even in their rawest forms. The true problem with many games isn't a lack of flashy graphics or witty dialogue, but a lack of engaging gameplay. It's like a restaurant with a great name and decor but mediocre food.
Gameplay and game mechanics are the heart of most games. For AI-native games to offer real value, AI must be instrumental in generating these mechanics, not just assets.
The Challenge of Generating Mechanics
Generating game mechanics is incredibly difficult. That's why many games focus on AI-generated assets rather than AI-driven mechanics. Given this challenge, and respecting the decades of expertise in the game industry, we decided to focus solely on the core problem: AI-generated game mechanics.
Learning from the Masters
To tackle this, we looked at games renowned for their mechanics, even if their visuals or stories weren't top-tier:
- Diablo 2: I have fond memories of this game. Yet, I always wished for more freedom in skill trees, for homing ice shards, or for summoned minions to summon more minions.
- Noita: This spell-crafting simulator is fascinating. You can combine spells to create incredible effects. However, it has a steep learning curve (even for an MIT Ph.D. student!) and is notoriously difficult, which can be off-putting.
- Vampire Survivors: A game that spawned a genre. Its addictive gameplay is undeniable. I often wished for more weapon variety and more freedom in combining them.
- The Legend of Zelda: Tears of the Kingdom (TotK): The "Ultrahand" and "Fuse" abilities offer a glimpse of freedom. Combining a bat eye with an arrow for a homing shot was brilliant. But the system is still rule-based. Why can't we combine any two items freely?
These games, while brilliant, often struggle with the balance between freedom, intuitiveness, and accessibility. "Braid" and "TotK" excel at intuitiveness and accessibility but offer limited freedom. "Noita" offers immense freedom but at the cost of intuitiveness and accessibility. "Diablo" and "Vampire Survivors" offer some build freedom but are still constrained by fixed rules.
The AI Mechanics Generation Model
Our goal is to achieve all three: freedom, intuitiveness, and accessibility. This is our unique opportunity. Leveraging our expertise in multi-modal large model training at Meshy AI, we believe we can build a "game mechanics dataset" to train a model that can generate an endless stream of unique mechanics, offering players a truly fresh experience every time.
This is a bold endeavor, perhaps unprecedented. But we're putting it into practice. After dedicated effort, we've developed a model that drives...
