Unless you're substituting GPT3 for AGI, GPT3 is not a super-intelligence. It's a set number of matrix multiplications that completes in polynomial time. Becaus

Unless you're substituting GPT3 for AGI, GPT3 is not a super-intelligence. It's a set number of matrix multiplications that completes in polynomial time. Because Transformer context is fixed, and given the limitations of training, adding non-determinism (tree search) means it goes off the rails and is prone to getting trapped in cycles anytime it doesn't quickly terminate.