“The original GPT-3 model scored zero on the ARC-AGI benchmark.”