You're gesturing at some fundamental limit that, if it exists, there is no evidence of peaking yet. It is not rigorous to say AI has not contributed to architecture improvement.
Transformers require massive data and compute, the smarter you want you're AI, the worse this situation gets. This isn't a hard limit but it certainly limits how quick it happens, and let me tell you, general purpose robots aren't happening anytime in the next 25 years