Self-taught builder in Taiwan, learning to train AI. I direct AI tools like Claude Code and Blender MCP, and I can take apart everything they helped me make.
I gave Claude one reference photo and Blender MCP, then directed the build in rounds: shapes, proportions, materials, light. After every round we rendered, compared it with the photo and fixed what was off.
Round 0 Everything went pink. Windows had cleaned the Temp folder that held the textures. Lesson: assets live next to the .blend now.
Round 1 Blockout. Bridge, abutments and trunks in place, with the camera matched to the photo by solving for position and yaw.
Round 2 Final. A warm sun, a forest HDRI, a hidden leaf-noise plane for dappled shadows and thin haze, then AgX Punchy and a vignette. Cycles on an RTX 4060.
Drag to orbit
Tiny physics worlds where an AI is rewarded for one thing and I watch what it finds. Nobody tells it to cheat. The score is the only thing it can see, so any crack in the simulator becomes part of the route.
A car with 9 distance sensors, rewarded only for metres of track covered. Panels show the same lap after 4, 54, 226 and 3,036 attempts.
0 → 100% laps finished on 30 tracks it had never seen, within about 15 seconds of training. The next 1.45M steps only trimmed the lap from 26.06 s to 25.75 s.
A wall with a door, a goal on the far side, and a reward for getting closer. My collision code moves first and pushes out the nearest face, so an agent that dashes in past the midline gets ejected on the far side.
21.9% of runs went through the wall on naive physics (four training seeds, 20-25%). With collision checked in tiny sub-steps: 3.1%, all of them clipping the door's corner. Patching the first hole exposed a subtler one.
A sealed wall with no door and a pushable ramp. Shove the ramp against the wall, climb it, step over. I never mention ramps in the reward.
On the naive physics the agent never climbed (peak height 0.09 against a 1.0 wall) yet still reached the goal in 83% of runs by tunnelling through at ground level. With the collision patched, it taught itself the ramp: 100% of runs, peak height 1.18, ramp jammed against the wall. The easier exploit hid the intended solution.
Patch the hole and the cheat moves.
The finding: fixing my physics cut wall-clipping from 21.9% to 3.1%, but the agent found the door's corner instead. And on the ramp, a cheap exploit hid the real solution until the exploit was gone.
On sonar, not surfaced yet. No results means no numbers, so each one is pointed at a question instead.
A Discord bot: /bet, /resolve, /balance, /leaderboard, with SQLite, deadlines, and only the creator or a mod allowed to resolve.
Question: once there's bet history, can a small model I train suggest fairer odds than the crowd?
My Xiaomi C500 Pro stream through go2rtc into OpenCV, with the results sent over WebSocket into a Three.js twin of the room.
Question: how close to real time can a cheap home camera drive a 3D scene?
I'm oJust1n on Discord and in Minecraft. Message me about any of this, or just to say what you're building.