+ This code draws the likeness of a 1990s ray traced scene of two spheres. It doesn't quite look physically correct though. +
+ +> Code +```js + + + + +Alibaba has somewhat stunned the AI community by their releasing a relatively capable o1 competitor as an open model, mere months after the unveiling of o1. The model is called QwQ Preview, or QwQ-32B-Preview if you will. You can learn more about it on google.
-Let's briefly evaluate its coding chops, using the Q4_K_M variant running via Ollama.
+Let's briefly evaluate its coding chops, using the Q4_K_M variant running via Ollama. Note that I won't be comparing it to o1 directly since my OAI API credit is at 0 and I don't want to put more in there right now.
@@ -778,6 +778,68 @@
A small ray tracer. Qwen 2.5 Coder gets further than QwQ; neither gets it right.
++ This code draws the likeness of a 1990s ray traced scene of two spheres. It doesn't quite look physically correct though. +
+ + +<!DOCTYPE html>
+<html lang="en">
+<head>
+ <meta charset="UTF-8">
+ <title>Simple Scene</title>
+</head>
+<body>
+ <canvas id="sceneCanvas" width="400" height="300"></canvas>
+ <script>
+ const canvas = document.getElementById('sceneCanvas');
+ const ctx = canvas.getContext('2d');
+
+ // Draw red sphere in the middle
+ ctx.fillStyle = 'red';
+ ctx.beginPath();
+ ctx.arc(200, 150, 50, 0, Math.PI * 2);
+ ctx.fill();
+
+ // Draw blue sphere to the left
+ ctx.fillStyle = 'blue';
+ ctx.beginPath();
+ ctx.arc(100, 170, 48, 0, Math.PI * 2);
+ ctx.fill();
+
+ // Draw reflection of red sphere on blue sphere (simplified)
+ ctx.globalAlpha = 0.5; // Semi-transparent for the reflection effect
+ ctx.beginPath();
+ ctx.arc(138, 206, 19, 0, Math.PI * 2);
+ ctx.fillStyle = 'red';
+ ctx.fill();
+ </script>
+</body>
+</html>
+
+
+ In the original image, the reflection of the red sphere on the blue sphere is incorrectly placed. QwQ was the only model that clearly understood this and fixed it. Grok 2 might also have had a sniff of this issue; the other models not so much.
@@ -845,7 +907,7 @@
Both QwQ and Qwen 2.5 Coder added a game-over condition, though I'd say QwQ's is better. They both also implemented right-clicking to place flags, but since the UI framework only supports left-click, that can't be tested. Overall, I wouldn't call either version finished, but getting there.
QwQ Preview does a decent job, often rising above Qwen 2.5 Coder and sometimes matching Claude 3.5 Sonnet; but on the flip side eating up quite a few more tokens. Interesting to see the full QwQ down the road.
+QwQ Preview does a decent job, often rising above Qwen 2.5 Coder and sometimes even Claude 3.5 Sonnet; but on the flip side eating up quite a few more tokens. Interesting to see the full QwQ down the road.