The Return of Assembly: When LLMs No Longer Need High-Level Languages (dev.to)

A developer demonstrated that a modern LLM (GitHub Copilot) can be prompted in plain English to generate working assembly for a specific target—an Apple M4 Max (ARM64) MacBook Pro—then auto-create a Python test harness to load and call the routine (validating a simple 3+5=8 example). The experiment showed the model handled low-level details (instruction selection, calling conventions, assembler syntax) without the author needing to know registers or platform quirks. The proof-of-concept is tiny and intentionally brittle, but it confirms an LLM can produce end-to-end machine code plus tests from a natural-language specification. For AI/ML and systems engineers this is significant: LLMs can treat ISA-level code like any other language, potentially enabling per-architecture models, AI-driven “compilers” that translate intent to optimized machine code, and automated micro-optimizations across CPU/GPU targets. That could blur the lines between design, coding, and compiling—letting teams describe intent rather than hand-write boilerplate or intermediates. Major caveats remain: generated assembly may be non-deterministic, hard to debug, brittle, and unsafe (segfaults, silent failures), and current demos are far from production-ready. Still, the work highlights a plausible near-term shift where models reason directly at hardware level, complementing — not immediately replacing — traditional toolchains and engineering discipline.