Pulling fine-tuning out of the black box to make it cheaper. Very much inside baseball (badly described and motivated). Clearly no cognitive science background. Technically very interesting.
Non-standard definition of emergence (a proxy for surprize) makes this paper very misleading from a cognitive perspective. The benchmarks are an anthropomorphic mess.
Easy, straightforward paper, seminal in the scaling literature. We revisited this one after four years. The only issue missing is any notion of data quality (vs data set size). Cardinality of compute and data is a good start.
I am giving two #swsec breakfast seminars back to back mid-April. If you are in Sweden, Norway or Finland, please consider coming. Pass it on to those who may be interested.