How I'd learn ML in 2025 (if I could start over)