Yeah, there's no way they'll move away from current AMD64 ISA. uarch change is just under the hood, most parts of the interface (the ISA) will remain the same.
They probably want to trim the obsolete stuff from it, like all the initialization dance the OS has to go through until it's in x64 mode. You literally need code from the 80s to bootstrap the OS (yes, 16bit stuff!), then you skip through 2 more modes because they're useless anyways now, while everything for these modes is still present in the CPU. And then there's 32 bit compatibility mode to maintain, wondering if that's on the chopping block too, as it doesn't have to be. It's not too complicated to map user level instructions 32bit instructions to 64bit ones in microcode and the OS will have to do some trickery as well so that programs will work, but it's doable. 32bit ring0 code will be dead though.
As far as SIMD instructions go all SSE iterations are forward compatible. So removing 1-3 but leaving 4 doesn't make sense (actually if you're looking only at 4, it's not a huge instruction set and not very useful on its own). AVX was the first point where forward compatibility broke, and I guess that's where Intel wants to draw the line. AVX conceptually (not the instructions themselves) is backwards compatible to SSE but forward compatibility was lost, and when writing code that switched modes constantly (without zeroing the vectors first) many cpu cycles are wasted, and in general it's a very bad practice because it interferes with OoOE.
My guess is what they want to do is to map legacy SSE instructions to respective VEX prefix'ed ones in microcode. This is a relatively easy change which should preserve full backwards compatibility with all normal code. Stuff that actively mixes SSE and AVX without zeroing out registers will break though.
But all the talk about removing some backwards compatibility should be small stuff. And that what ticks me off, it should be, but it feels like not. It's not like this wasn't brought up before many many times, while each and every time was shot down with the same argument. And now that same argument is used the other way around... Somewhat puzzling. And currently I can think of only one explanation, and it's not a good one for the technology. So I hope I'm wrong