Daniel Kokotajlo comments on PaLM in “Extrapolating GPT-N performance”

[Daniel Kokotajlo]

You may be interested in this image. I would be grateful for critiques; maybe I’m thinking about it wrong?