What kills performance are not memory copies, but locks. Parallel nonblocking IO and a non POSIX stdlib will bring you far away from C++ or Rust performance.
I'm pretty sure if every thread executing an LLM model had to have its own copy that that would murder performance more than any lock does, and it won't even be close.
It's cheaper to copy than to lock when the data is small, but that does not scale and it also ignores things like reader/writer locks where the data is primarily read-only, at least during the concurrent stage. Or where the work can be safely chunked up such that writes don't ever overlap which is very common in graphics
Exactly this.
(Don't get me wrong, I liked the idea behind Pony for backend code, it's much saner than Go for the same target space. But it failed to capture that market, because Go was already there. And it was never a competitor to Rust because this model is only viable for web back end tasks, not for general computing).
The pony model was also better for compute tasks, not just IO. Because it provided safe concurrency, 10x faster than go.
Think for instance about how you'd do efficient matrix multiplication of two matrices with a million row and column, in Pony, versus how it works in languages with shared memory. You'd spend a gigantic amount of time copying data for no good reason…
strong claim. care to back it up?