> We have probably hit a limit to scaling LLMs through raw parameter count alone, at least we're not seeing the exponential pace.

Source?