They really do test different things. In this case he is testing video renders from Premiere Pro, which renders for tens of minutes at 100%. It's apparently an outrageously inefficient app as well as the Windows version of it renders the same output in a fraction of the time (which invalidates it as a test, really).
All modern chips thermally throttle. Run a heavy AVX2 workload and you'll often go below the so-called "base" frequency. Run one or two threads and it can dramatically "overclock". It is the nature of all current chips.
Eh. For the overwhelming majority of users I doubt this would ever be an issue. Compile a kernel. Apply a filter. Do some convolutions. Every normal thing will be ridiculously fast. Run a server or workstation style load and it won't be so good. That has been the case with laptops for time eternal.
Only ones in systems that have not been designed correctly!
The fact that the Aero x15 he tested with the 2.2Ghz base clock 6 core i7 can all core boost to 3.1Ghz in the same workload that brings the macbook pro far below it's base clock speed proves that!
P.S, it does not matter what program it is running, throttling is a hardware flaw!
Depends on the workload and amount of time spent processing the workloads.
Dave Lee ran some tests that took a half hour to complete with the CPU pretty much maxed out. I would imagine the CPU was running hot for most of that time.
[1] http://hrtapps.com/blogs/20180712/