Tesla FSD 14.3.12 rewrites AI stack

This is less a feature drop than an infrastructure release: Tesla is co-designing training, perception, compilation, and runtime execution as one autonomy stack.

• –The MLIR rewrite targets latency and iteration speed, suggesting compiler work is becoming as important as model architecture.
• –Hard fleet examples now drive improvements for rare vehicles, complex intersections, small animals, and temporary system degradations.
• –The 20% reaction-time claim is vendor-reported; Tesla provides no benchmark setup, baseline, or independent safety data.
• –HW4-only availability highlights the growing performance gap between Tesla’s hardware generations.
• –For AI builders, the broader lesson is clear: reliable edge AI depends on tight feedback loops between data, training, models, and deployment infrastructure.