2048 AI Learning From Zero
By @redp314. Made with Claude Opus 5.5. Watch the original post.
Remake prompt (written by Lumpong Automated)
Make a 28-second portrait 4:5 (1080x1350) training visualization for a 2048 AI that learns from zero, built in HTML with SVG and JavaScript and rendered to an MP4. The AI should truly run in the page (reinforcement learning or n-tuple learning, no external models).
Look: a near-black page. A big bold headline 'GAME #497' in white that turns amber at the later milestones ('GAME #98,299'), a grey subline showing training time ('0.9 seconds of training.', later '12 minutes of training.') and a small circular arrow button top right. Below it a classic 2048 board on a taupe frame with beige, orange, red and yellow tiles; the 16384 tile is dark charcoal with a yellow outline. Bottom stats in small caps: SCORE, BEST TILE (amber when high), GAMES PLAYED.
Animate fast play and show milestones as the game counter jumps: Game 1 dies at 128, Game 497 first 2048, Game 3,915 first 8192, Game 98,299 first 16384, and after 2M games a 99% win rate. Speed up the counter between milestones, ease tile slides, and flash each milestone. Keep the numbers large and the pace quick.