A Reddit user spent 4 weeks, burned through 33 billion tokens, and wrote nearly 40,000 lines of code—only to fail at producing a car that actually drives. The episode forces us to revisit a question: where exactly does "Vibe Coding" (letting AI write code based on vibes) hit its limits?
What this is
The story began with a complaint post in Reddit's game dev subreddit: the OP used Claude Pro Max and Fable (an AI game development tool) to try building a pixel-art racing game, and after 4 weeks was still stuck on basics like vehicle movement.
The comment section reached a consensus—the problem wasn't the tool, but the method. The original author distilled this into 10 engineering rules, then had Claude replicate a racing demo following them. The conclusion was blunt: with the right method, AI can produce a playable version in a short time.
Industry view
Supporters argue this proves AI's value lies in "accelerating validation," not "designing for you." Letting AI write tuning tools and test code is the highest-ROI usage—a view that surfaced repeatedly in the comments.
But the dissent was equally sharp. One developer wrote: "AI makes it very easy to reinvent the wheel—until you end up with a car that has 20 wheels, each pointing a different direction." Another pointed out that 40,000 lines of code at this stage "should only be a few hundred." Stacking features before the basic framework even runs is a textbook case of engineering debt spiraling out of control.
There's another warning sign worth flagging: when the OP asked it to "reskin" the racer into a wuxia horse-racing game, the AI did deliver—but only swapped the skin, without understanding the interaction difference between riding a horse and driving a car. AI's "understanding" today still lives at the texture layer.
Impact on regular people
For enterprise IT: the "AI-driven cost reduction" narrative needs a discount. Write specs clearly first, validate in phases, and use AI for debugging rather than core design—that's the safer path.
For individual careers: knowing how to use AI is no longer scarce; what's scarce is "knowing what you actually want." The closer you sit to the requirements-definition layer, the harder you are to replace.
For the consumer market: the next year will see a flood of AI-rushed apps and games, and the gap between "runs" and "usable" will be enormous. When picking tools, "hands-on feel" is a more reliable signal than a feature checklist.