The claim that tok/s independent of quality is a useful comparison (I can get thousands of tok/s on a suitable small model), and secondarily that humans can’t output “tokens” faster than than in some sense, which I am less confident about