been reading through @axisrobotics latest post and the core idea in it flips how most people think about training data most teams treat data collection as a one time step, collect enough trajectories then start training and rarely look back. axis is arguing that's the wrong mental model entirely the point they make is that a trajectory doesnt have a fixed value. something thats essential for an early model can become completely redundant once that model improves, while a rare edge case can suddenly become the actual bottleneck as the model gets better. so the value of any piece of data depends on where the model currently stands not some fixed quality score what they built around that idea is a closed loop, collect train evaluate adapt the data then train again. instead of a human deciding upfront what data matters, the model's own failures point directly at what still needs to be collected the correction example from their earlier research fits perfectly here too. out of 660 human corrections only 161 actually mattered enough to improve the policy, the rest were basically redundant because the model had already learned that behavior what stands out to me is the reframing at the end, more data doesnt automatically mean a better model. if new data just repeats what the model already knows, youre burning compute for nothing. the real gain comes from data that sits exactly where the model is still failing