What will the aggregate improvement of GPT5 be over GPT4 in terms of metrics?

GPT-5 plus scaffolding and inference-compute ~= training compute will achieve capabilities advance >= (GPT-4 to GPT-5).

Will the performance jump from GPT4->GPT5 be less than the one from GPT3->GPT4?

Will GPT-5 be released incrementally as GPT4.x for different checkpoints from the training run?

Will growth in the maximum MW used to train AIs, slow down by more than x2 after GPT-6-like?

How much compute will be used to train GPT-5?

What will be true about GPT-5?

Will GPT5 show clear signs of diminishing returns?

Will growth in the maximum MW used to train AIs, slow down by more than x2 after GPT-5-like?

Will GPT-5 have fewer parameters than GPT-4? (1500M subsidy)