I built a full 3D open-world racing game almost entirely with AI, and it now has real daily players. Here’s the honest breakdown of what the model nailed and where it completely fell apart.

Not a hype post. I want to talk about where we actually are, because building a real, shipped, multiplayer-ish thing with AI taught me more about the current ceiling than any benchmark did.

The project: a neon open-world street racer that runs in the browser, no install. Real 3D city you drive around, other live players on the road, a garage, an economy, the works. I directed it, but the overwhelming majority of the code was written by AI. It went from empty folder to live with actual daily players in a couple of weeks.

What the AI was genuinely great at:

• Whole self-contained systems in one shot. "Build a photo mode with orbit camera and filters," done and working.
• Boilerplate-heavy, well-trodden problems: auth, a save system, a REST API, Stripe wiring. Fast and mostly correct.
• Refactors and translations. "Turn this into an instanced mesh so it's one draw call" is the kind of tedious change it does better than I would by hand.
• Being a tireless debugging partner when I could describe the symptom precisely.

Where it fell on its face:

• Spatial and 3D reasoning. Anything involving "this object is behind that one" or "the plate is buried in the bumper" it could not see, because it can't see. I had to be its eyes constantly.
• Holding the whole system in its head. It would fix one thing and quietly break a system three files away, because it didn't truly model the interactions, only the local change.
• Performance intuition. It happily wrote code that attached a light to every streamed car and tanked the framerate. It knew the fix once I found the cause, but it did not anticipate it.
• Game feel. It cannot tell you a mechanic is boring or an economy is exploitable. That judgment is still entirely yours.

The real takeaway: the bottleneck has moved. It's no longer "can it write the code," it's "can you specify precisely, verify relentlessly, and supply the taste and the spatial judgment it lacks." AI turned me from someone who writes features into someone who directs and tests them. That's a genuinely different job, and honestly a more demanding one than people expect.

The proof it's more than a toy: it's live, people play it daily, and a few have even paid to support it. So this isn't a weekend demo that died in a folder, it's a real product carried mostly by AI code with a human holding the wheel.

Curious where others draw the line. For those of you shipping real things with AI, not demos, where does it still fall apart for you? My money's on anything requiring a mental model of state over time.