You’ve been told that systems engineering is a deep moat. That it takes years of expertise to write low-level code for hardware backends. That a single developer couldn’t possibly reimplement a complex machine learning framework like JAX for a different GPU architecture. Well, someone just did it. In a weekend. With an AI assistant. And the implications are terrifying.
On GitHub, developer Oleg Eterevsky (known as eterevsky) shared his project: metaljax. He used an AI agent called Fable to ‘vibe code’ a Metal backend for JAX — a massive undertaking that traditionally requires a team of specialized systems engineers months of work. The result? 20,000 lines of code, 98.4% of the original JAX test suite passing, and a 10x performance improvement over CPU on Apple Silicon. The developer’s own words: ‘With my minimal steering Fable wrote 20k LoC.’ Minimal steering. That’s the key phrase.
The bottleneck of human engineering has officially shifted from implementation to specification.
The twist is that the output is not a toy. It’s a legitimate, high-performance backend for a cutting-edge machine learning framework. This isn’t ‘Hello World’ generated by ChatGPT. This is deep-tech infrastructure, built by a single person who didn’t write most of the code.
Hardware backend development, once a secure bastion of elite engineering, is now a commodity. The value is in the prompt, not the code.
This is brilliant. It’s also a wake-up call. If you’re a developer who prides yourself on writing intricate C++ or CUDA, your strategic value is deflating. The market is about to reprice your skills. The question is no longer ‘Can AI write code?’ It’s ‘Are you ready to be the one who steers the AI, or the one replaced by it?’
FAQ
Q: Is this really a 20,000-line codebase written by AI? Isn't it just copy-paste from existing code?
A: It's a genuine port. The developer used an AI agent (Fable) to generate the code from scratch based on the JAX architecture. The result is a working backend with 98.4% test pass rate — not a trivial copy. The code is complex, low-level, and specific to Metal.
Q: What does this mean for my job as a systems engineer?
A: You need to shift from writing code to designing architectures, writing robust tests, and steering AI. Pure implementation skills are becoming a commodity. The value is now in knowing what to build and how to verify it, not in keyboard hours.
Q: Isn't this just a one-off demo? AI can't handle real production systems with edge cases and security concerns.
A: This is a harbinger, not a fluke. The gap between 'vibe coding' and production-grade code is narrowing fast. If a single developer can do this in a weekend with minimal steering, imagine what a team with AI can do in a month. The trend is clear: the bottleneck is shifting.