For me, the takeaway is that training state of the art models may become cost prohibitive even for large companies. There may be a practical limit to how complex models can become. This is at least plausibly interesting to know.