agree on the code side. for decisions i'd push it a bit further though. the trap isn't really that the model does your thinking for you, it's that it agrees with the thinking you already brought to it. you can't verify a decision the way you verify code, there's no test that fails when "this answer is wrong because you wanted to hear it." so even if you don't outsource the thinking, you can walk away with a plausible-sounding agreement that feels like understanding. that's the failure mode that worries me more. replacement is honestly the easier one to spot.