Kimi K3 Built the Game. I Still Had to Play It

Kimi K3, an AI coding model from Moonshot AI, built a playable first-person 3D browser game from a single prompt, but the game had bugs that only surfaced during human play, such as a drone clipping through the player and a knocked-out drone waking up. The author, who tested the game on the $199 monthly Kimi Code plan, spent less than 1% of the monthly quota (roughly $2) and found that while the model fixed the reported bugs, a human still had to identify failures, define correct behavior, and own the final result. The author also noted that generating 100 taunts took over 15 minutes due to server load, and highlighted Kimi K3's architecture (Kimi Delta Attention and Attention Residuals) which the company claims offers a 2.5x scaling efficiency improvement over Kimi K2.

Kimi K3 Built the Game. I Still Had to Play It. Kimi K3 built a playable first-person game from one prompt, but its bugs only appeared when a human played it. AI can make the first version. Somebody still owns the failures. The first build I gave Kimi K3 one prompt and asked it to build a browser game. The assignment was a first-person 3D maintenance game inside a procedurally generated server facility. The player had to collect data cores, reach a freight elevator, and descend three floors. The stack was Three.js, vanilla JavaScript, and Vite. Kimi produced a playable game in one shot. It was fun. It also had rough edges that mattered immediately: a drone could clip through my body, and hitting a knocked-out drone could wake it back up. The game looked like a finished demo until I tried to play it like a person. The code is on GitHub https://github.com/jcpsimmons/kimi-test , and you can play the game here https://kimi-test-theta.vercel.app/ . Testing still belonged to me I described the two failures, asked Kimi to fix them in parallel, restarted the development server, and tested the result. The collision bug was gone. The projectile bug was gone. The model did the repair work, but I had to see the failure, name it, and decide what correct behavior meant. Why the prompt mattered The prompt was deliberately difficult. I did not give Kimi a flat map and ask it to add a character. I asked for procedural levels with a different layout each time the random seed changed. That forced the model to keep the architecture, gameplay rules, rendering, physics, and progression in its head at the same time. What Kimi K3 is good at Kimi K3 is clearly built for long-horizon coding tasks with a clear outcome. Give it a large project and it will make a lot of decisions for you. Give it a small task and it may still respond like you asked for a department. What the first result cannot tell you A playable game is a real result. It is also the first result a human can judge. The work moves into testing, correction, security, performance, and the boring decisions that decide whether anybody can live with the software. The price of the experiment I used Kimi Code on the $199 monthly plan. After the initial specification, the first game build, the bug fixes, and a few chat tests, I had used less than 1% of the monthly quota. By rough math, the project cost about $2 on that plan. Kimi's pricing page https://www.kimi.com/help/membership/membership-pricing lists the plan and its shared quota structure. The wait was the problem. When I asked Kimi to generate 100 taunts and add them to the game, the task took slightly more than 15 minutes. That is long enough to get coffee and long enough to lose the context of the work when you come back. The servers were under heavy load, so I cannot treat that speed as a final verdict. I can say that quality stayed much more consistent than the waiting time. The engineering underneath The architecture matters here. Kimi K3 is built on Kimi Delta Attention and Attention Residuals. The company says the combination produces an approximate 2.5x improvement in overall scaling efficiency compared with Kimi K2. If that claim holds up, Kimi can do more work without demanding a miracle from the hardware. I walked through the claim in the video https://www.youtube.com/watch?v=6GUDjXh-6g8 , and Moonshot's technical write-up https://www.kimi.com/blog/kimi-k3 lays out the architecture. Somebody still owns the failures A model can produce a system quickly. Somebody still has to define the system, catch the drone walking through the player, test the fix, and decide whether the result is safe enough to ship. That person did not disappear. The job moved closer to judgment and ownership, and companies will try to pay less for both. AI is real. The people selling the labor story are the part I distrust. Kimi K3 can do work that used to require a room full of setup, code, and patience. It can also hand you a system whose failure modes appear only after you start playing it. The model earned my recommendation. It did not earn a free pass. - Dr. J