Great machines, sure, but it's still (to me) a compromise. Right now I have a Macbook pro for portable use and a full-on workstation at home with a 20870Ti in it for playing with ML and CUDA, and for gaming.

I love the idea of having one machine and just plugging in to an external GPU caddy when at home for the full capability. Unfortunately MacOS and Nvidia cards aren't friends right now...

And the laptops that have got discrete GPUs are both heavier than the ideal (I'd like a portable machine coming in at 1kg or less) and far less capable than the 2080Ti.

It's not rocket science. I've run a 30% overclock on a machine with 80% daily uptime on air and saw it BSOD like twice over 4 years. Also as some other people in this thread have mentioned, modern CPUs will overclock themselves, and if you undervolt you can get quite a lot of headroom, especially if you get lucky with binning. With that headroom and some hackery (Throttlestop is often all you need) you can get modern chips to hold all core turbos indefinitely. YMMV but I've managed to hold a stable 35% increase in sustained processing power without running too close to the thermal limits.

The risk of breaking anything by overclocking is really low anymore - you'll run into instability long before you damage anything. The best argument against overclocking IMHO is that the time and expense is generally not worth it. With a CPU overclock you might gain 10% performance while consuming 30% more power, after investing time and fancy cooling solution. Gone are the days of rapid improvement when you could occasionally get a stable 33% overclock with no real downside.