[michaelhartm]

They used the 6b GPT4-J, not 20B. That's what's interesting, it's a smallish large language model :).

[dragonwriter]

GPT-J, not GPT4-J.