It gives people a very incorrect view of the actual emissions costs of training language models. So people think "If a tiny model generated that much emissions, larger models must be thousands of times more than this" when in actual fact, because of the errors, this isn't the

→ View original post on X — @jeffdean