AI/LLMs have been dramatically improving for 7+ years. There's now a lot more funding to support continued improvement. You're correct this is an "assumption", but continued improvement at the same pace (or faster) for the next 3+ years is just extrapolating a trend. Believing we've hit the top today is based on nothing at all. Continued improvement is much more likely.
You can only tell which part of the S-Curve you're on in retrospect. It's not something you can tell while you're experiencing it. Both scenarios of AI maxing out or continuing to improve are both likely.