👀 quite interesting…. • Model evaluation: 50% • Data cleaning and organization: 40% • System integration: 8% • Model training: 2% The point isn't that training doesn't matter. It's that most people dramatically underestimate the importance of data and evaluation. Working with founders building AI cameras, robotics, and edge AI systems, we've seen the same thing. Teams often spend weeks debating which model to use, while spending far less time defining success, improving labels, or understanding failure cases. A better model can give you incremental gains. a better feedback loop can change the trajectory of the entire product. As foundation models become increasingly accessible, the real moat may not be the model itself, but the quality of the data and the discipline behind how it's evaluated.