Testing Kimi AI in Real-World Dev Tasks: 3 Quick Tests (and Where It Stumbles)

#ai #webdev #programming #graphics

Kimi AI by Moonshot AI has been making waves in the developer community—mostly because of its massive context window and strong coding capabilities.

Instead of writing endless theoretical reviews, I decided to run Kimi through 3 practical, everyday tests that developers face. No perfectionism, no complex prompt engineering—just testing how the model performs out of the box, where it shines, and where its logic breaks down.

Test 1: Analyzing and Summarizing Massive Log Files & Codebases
The primary selling point of Kimi is its huge context window.

The Task: Upload a massive error log file (or a huge legacy codebase file with thousands of lines) and ask Kimi to trace an issue and propose a fix.

The Result: Kimi holds context remarkably well. Unlike standard models that start "forgetting" the beginning of a document after a few prompts, Kimi pinpoints exact lines and function dependencies accurately.

Verdict: 8/10. Great tool for code reviews, debugging unfamiliar codebases, and parsing system logs.

Test 2: The Penrose Impossible Triangle in SVG / Shader Math (Where Kimi Gets Stuck)
This is where things got really interesting—and where Kimi hit a hard wall.

The Task: Ask Kimi to calculate the exact coordinate geometry or SDF (Signed Distance Function) for a Penrose Impossible Triangle (the famous optical illusion where geometry loops infinitely).

The Problem: The Penrose Triangle is physically impossible in 3D space. To render it accurately via vector points (SVG) or shader math (GLSL), the model needs to handle complex coordinate shifts, depth tricks, and flawless edge blending.

<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 600 600" width="100%" height="100%">
  <defs>
    <radialGradient id="cyber-bg" cx="50%" cy="50%" r="75%">
      <stop offset="0%" stop-color="#0d0826"/>
      <stop offset="60%" stop-color="#050212"/>
      <stop offset="100%" stop-color="#000005"/>
    </radialGradient>

    <pattern id="grid" width="30" height="30" patternUnits="userSpaceOnUse">
      <path d="M 30 0 L 0 0 0 30" fill="none" stroke="#25124e" stroke-width="0.8" opacity="0.4"/>
    </pattern>

    <filter id="neon-glow" x="-30%" y="-30%" width="160%" height="160%">
      <feGaussianBlur stdDeviation="4" result="blur1" />
      <feGaussianBlur stdDeviation="12" result="blur2" />
      <feMerge>
        <feMergeNode in="blur2" />
        <feMergeNode in="blur1" />
        <feMergeNode in="SourceGraphic" />
      </feMerge>
    </filter>

    <linearGradient id="neon-cyan" x1="0%" y1="0%" x2="100%" y2="100%">
      <stop offset="0%" stop-color="#00f6ff"/>
      <stop offset="100%" stop-color="#0066ff"/>
    </linearGradient>

    <linearGradient id="neon-magenta" x1="0%" y1="0%" x2="100%" y2="100%">
      <stop offset="0%" stop-color="#ff007f"/>
      <stop offset="100%" stop-color="#9900ff"/>
    </linearGradient>

    <linearGradient id="neon-purple" x1="0%" y1="0%" x2="100%" y2="100%">
      <stop offset="0%" stop-color="#7928ca"/>
      <stop offset="100%" stop-color="#2b0066"/>
    </linearGradient>
  </defs>

  <rect width="600" height="600" fill="url(#cyber-bg)"/>
  <rect width="600" height="600" fill="url(#grid)"/>

  <circle cx="300" cy="300" r="180" fill="none" stroke="#ff007f" stroke-width="1.5" opacity="0.3" filter="url(#neon-glow)" stroke-dasharray="10 15"/>
  <circle cx="300" cy="300" r="140" fill="none" stroke="#00f6ff" stroke-width="1" opacity="0.2"/>

  <g filter="url(#neon-glow)" transform="translate(300, 310)">
    <polygon points="0,-190 165,95 105,95 0,-86 -45,-8 -105,-112" fill="url(#neon-cyan)" stroke="#ffffff" stroke-width="0.7" opacity="0.95"/>
    <polygon points="165,95 -165,95 -135,43 105,43 0,-138 60,-138" fill="url(#neon-magenta)" stroke="#ffffff" stroke-width="0.7" opacity="0.95"/>
    <polygon points="-165,95 0,-190 60,-138 -60,7 -135,43 -165,95" fill="url(#neon-purple)" stroke="#ffffff" stroke-width="0.7" opacity="0.95"/>

    <polyline points="0,-190 165,95 -165,95 Z" fill="none" stroke="#00f6ff" stroke-width="2" opacity="0.8"/>
    <polyline points="105,95 0,-86 -105,-112" fill="none" stroke="#ff007f" stroke-width="1.5" opacity="0.8"/>
  </g>

  <text x="300" y="530" text-anchor="middle" fill="#00f6ff" font-family="'Courier New', monospace" font-size="14" font-weight="bold" letter-spacing="6" filter="url(#neon-glow)">
    SYSTEM.PARADOX // PENROSE
  </text>
  <text x="300" y="550" text-anchor="middle" fill="#ff007f" font-family="'Courier New', monospace" font-size="10" letter-spacing="3" opacity="0.7">
    [IMPOSSIBLE GEOMETRY DETECTED]
  </text>
</svg>

First, I asked it to generate vector data for a cyberpunk-style SVG render of the illusion. It understood the structure, but the coordinate math for the overlapping faces was a mess. I had to fix the vertices manually to get a clean visual.

When I pushed further and asked for a 3D Signed Distance Function (SDF) to render this in a shader, Kimi got completely stuck on the non-Euclidean math. It understood how to make a basic bar, but failed to connect them seamlessly, leaving placeholders where the actual math should be.

// Kimi's typical code output when calculating impossible geometry:
float sdPenroseBar(vec3 p, vec3 b) {
    vec3 q = abs(p) - b;
    return length(max(q, 0.0)) + min(max(q.x, max(q.y, q.z)), 0.0);
}

float sdPenroseTriangle(vec3 p) {
    float bar1 = sdPenroseBar(p - vec3(0.0), vec3(1.0, 0.2, 0.2));
    float bar2 = sdPenroseBar(rotateY(p, 1.209), vec3(1.0, 0.2, 0.2));

    // ... calculate seamless depth blend angle here
    // TODO: fix vector overlaps for closed optical loop

    return min(bar1, bar2);
}

The Result: Abstract spatial logic and impossible 3D geometry force Kimi to drop placeholders right where you need exact numbers.

Verdict: 5/10. Abstract spatial logic forces Kimi to fail on 3D geometry. You will have to fix the math manually.

Test 3: Refactoring and Optimizing Legacy SQL / TypeScript
The Task: Take a messy SQL query with multiple JOINs or a deeply nested TypeScript function and ask Kimi to refactor it for performance and readability.

The Result: Standard programming logic is Kimi's comfort zone. It easily breaks down messy SQL into clean Common Table Expressions (CTEs) and rewrites nested loops into clean array methods in TypeScript without breaking business logic.

Verdict: 9/10. Excellent for daily refactoring tasks.

Conclusion
Kimi AI is an impressive assistant for standard dev workflows—reading large logs, refactoring code, and explaining complex systems. However, if you push it into edge-case spatial math or impossible 3D geometry, expect it to hit a limit and leave placeholders.