Elon Musk recently said Grok 4.7 needs a few more days, but what caught my attention is that Grok 4.8 is being trained with a new C++ software stack. Earlier he also mentioned rewriting the training/inference stack in C/C++, removing many software layers and mapping it more directly to hardware.
This made me wonder: as AI writes more and more code, will we gradually move away from abstractions designed mainly for human programmers?
A lot of programming languages, frameworks and software layers exist because they make programming easier for humans. But LLMs may not need the same convenience. Maybe some of these abstractions actually limit what an AI can do.
I think this may also apply to agent harnesses: we design tools and interfaces to make agents easier for humans to control, but those interfaces may not be what is best for the model itself.
Could we eventually see software move from human-oriented abstractions toward machine-oriented ones?
Elon Musk recently said Grok 4.7 needs a few more days, but what caught my attention is that Grok 4.8 is being trained with a new C++ software stack. Earlier he also mentioned rewriting the training/inference stack in C/C++, removing many software layers and mapping it more directly to hardware.
This made me wonder: as AI writes more and more code, will we gradually move away from abstractions designed mainly for human programmers?
A lot of programming languages, frameworks and software layers exist because they make programming easier for humans. But LLMs may not need the same convenience. Maybe some of these abstractions actually limit what an AI can do.
I think this may also apply to agent harnesses: we design tools and interfaces to make agents easier for humans to control, but those interfaces may not be what is best for the model itself.
Could we eventually see software move from human-oriented abstractions toward machine-oriented ones?