I gave MiniMax Code one prompt: build a 3D neon rooftop runner. And chose M3.1 Flash model, Then I left it alone. 46 min later: a playable rain-soaked city. One HTML file, no build step. Then I said “make it more cinematic.” It reworked the lighting and the controls still worked.sunset that fades to night as you speed up, flickering neon signs, shield power-ups, slide barriers, combos, touch controls. Controls unchanged. Then I told it to playtest its own game. It found a bug I’d never have caught by playing for a minute. After a few hundred meters, rooftops stacked on each other and the runner fell through the floor with no gap. Its first tests missed it too, because they only ran short sessions. MiniMax Code with M3.1 Flash fixed the track logic and added a check across 30,000 steps. Final: 96/96 checks, zero console errors. It also told me what it couldn’t verify: real GPU frame rate, since its sandbox had no GPU. MiniMax Code’s daily check-in: 400 points/day, 1,000 on day 4 and 7, 4,000 for the full week. What would you build with an AI agent? Try MiniMax Code: agent.minimax.io/download

Stanford’s AI degree content (worth over $200,000) is now FREE. The university has released full AI and Machine Learning courses on YouTube. And these aren’t watered-down versions. They’re the exact same lectures and materials used by students on campus. 2️⃣ CS231N: Deep Learning for Computer Vision | 2025 The gold-standard course for understanding how machines “see” and interpret images. Link in the comments 👇🏼 Follow for more & Save for later🔖