Comment by energy123
I never understood this line of reasoning. I found it much more impressive that OpenAI's ML researchers realized this is the thing and bet big on it first, than to come up with it in the first place. It's underappreciated how much talent and insight it takes to see the obvious.
The TPU architecture is the most impressive thing I care about. They developed them and have been using them internally for years. This shows they grok what they're actually doing.
There are serious philosophical problems with betting big on an interesting outcome like ChatGPT, even though it seems obvious (Google also did this of course), but creating the best architecture to do that job seems like a first-principles intelligent move, because there was no reason to keep using graphics cards except that they "did the job."