It's interesting that no one has really considered the possibility that an individual outside these megacorps reimplements GPT4 while they're all pausing.

We've seen several examples of CPU-optimized code (textsynth, llama.cpp) indicating that there is a lot of performance to be gained from writing optimized versions of our inference routines; I doubt it's outside the realm of possibility that a single player writes code that lets them train a GPT4+ model on a CPU with a bunch of RAM. All they have to do is find a way to write C++ that will train a 4bit model on the CPU.