I used AI to make a Musk vs Altman racing game. It was buggy, but shows why vibe coding is the talk of Silicon Valley.

There are lots of vibe coding tools out there, such as Cursor and ChatGPT. I used Replit Agent, which lets you build apps with prompts in plain English.

I began by describing the type of game I wanted to build: a Mario Kart-style game in which Sam Altman races Elon Musk to achieve artificial general intelligence, or AGI.

The agent suggested a plan for the game and gave options in boxes to check and approve. The plan included building the initial prototype, adding multiplayer functionality, implementing additional power-ups and obstacles, creating additional race tracks with different themes, and adding character special abilities.

I clicked to approve the plan, and it started writing code in real time on the right-hand side of the screen.

Its first attempt looked promising. In just a few minutes, it created a main menu screen and added a text box explaining what the game was. It also decided to call the game "AGI race."

I liked that it created a second page for players to choose their character and displayed their statistics, including speed and acceleration.

But when I clicked "start race," the game didn't actually work. Instead, it displayed a weird error message that I didn't understand. I told the agent it wasn't working and copied and pasted the error message so it could fix it.

I also instructed it to make the main menu page more aesthetically pleasing and add 3D animations for the characters. It spent a few minutes making those changes before asking me to check if the game was running properly.

The main menu page looked better with the 3D characters, although it initially turned Musk and Altman into people of color. However, I realized I was too focused on how it looked and decided to get it to function properly before getting it to tweak the user interface. It made those changes and then asked if everything was loading and displaying correctly.

This time, it displayed a race track, but the game was not working. Using the keyboard shortcuts just made the cars and characters spin in circles — they were impossible to control.

After telling the Replit agent to fix the spinning issue and make the game functional, it changed the page completely to a simplified race track, but this time nothing worked.

I told it to revert to the previous version but fix the steering and add coins the characters could collect that would give them a speed boost.

It ran a whole bunch of checks and displayed its "thinking" while it reviewed its code to fix the bugs. That seemed to do the trick, as the controls were now working, and I could accelerate without spinning.

Next, I told it to make some tweaks, like labeling the coins "compute resources," "hire top talent," and "funding from investors." Also, I told it to have a visual effect for when the characters collect coins to show that it gives them a speed boost and another for when an opponent is hit with a lawsuit to show that it slows them down.

After about 20 minutes, the agent said it enhanced the game with several key improvements and told me to try the game again.

However, that made the game experience deteriorate, and the screen started blurring when I tried to play.

I encountered this issue a lot throughout my experience using Replit to create the game: I would instruct it to fix one problem, and it would lead to new ones. Much of my time was spent waiting for it to review its code and identify the errors. That was really frustrating.

I suspect that if I had coding experience, I might have been able to identify the root causes and fix them faster.

After some debugging, the game was mostly functional. However, the visual elements fell short of my expectations. I wanted the race track to resemble an actual track and feature a Silicon Valley-themed background, but I hesitated to request further major changes, fearing it might break again and cause more headaches.

For the final touches, I instructed the agent to change the game's title to "The Final Boss of AGI" and make the main menu slightly more exciting.

This is what it came up with — try out the game for yourself here.

My vibe coding takeaways

Overall, I found the experience impressive. I fully gave in to the vibes, letting the AI agent take the wheel, as it's clearly a whole lot more competent than I am at coding and app development. As Karpathy put it, "I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works."

Here are my main takeaways:

• You can just build stuff. The tool makes it easier than ever to bring ideas to life and has lowered the barrier to app development.
• Keep it simple. Asking for visual improvements or other tweaks can unexpectedly break other aspects of the projects. Sticking to straightforward prompts is the best approach to making something quickly unless you're prepared to spend more time debugging or refining the app.
• Coding knowledge still matters. While the AI is powerful, it's not flawless at troubleshooting. To get everything working exactly as intended, you need coding experience to diagnose and fix issues.
• It's great for low-risk projects. As Karpathy said, "It's not too bad for throwaway weekend projects, but still quite amusing." That sentiment perfectly sums up my experience: it's a fantastic tool for experimentation but probably not quite ready for high-stakes development.