It's never too late, "operating systems" as in software, are really new and will change significantly many times in our lifetime. Better not get attached to ephemeral mistakes like UNIX, Plan9, Win32 and so on.
@GrowlerEnjooyer@HSVSphere It's only useful if you are executing machine code.
The compiler/vm's semantics can provide far better safety than page protection or rings, without incurring their inherent performance penalties.
See: Singularity OS
@spandrell4 thing is we DO already put neural networks into hardware. Both Intel and AMD use it for prefetching, and by now probably for branch prediction too. Those things have to run at nanosecond-scale.
@spandrell4 1. Etching a neural network into silicon is too expensive to do every few months
2. microgpt is a few thousands of parameters, not billions
3. The KV cache there is basically nothing. Once you grow it then it becomes the bottleneck and not much faster than current hardware.