Today might be my first real day of "vibe coding."
I studied liberal arts and now work in AI marketing. I still do not write code myself.
But today, on a domestic workstation (Hygon CPU + MetaX GPU), I got a quantized DeepSeek R1 running locally.
It sounds dramatic, but it was surprisingly straightforward.
I gave Codex one instruction: deploy DeepSeek R1 with KTransformers on this machine, and make inference work.
Then I stayed next to it and watched.
It downloaded dependencies, compiled components, tuned parameters, adapted to the GPU, and when something failed, it adjusted configs, switched quantization paths, rebuilt environments, and retried.
I did not touch the keyboard to write code.
When the terminal finally started printing inference output, I froze for a second.
Not because of speed. Honestly, performance was just okay.
What hit me was this: I had not "learned" any specific engineering trick, but I was still fully involved in the process, and we shipped the result.
Before this, domestic compute was mostly a slide-deck concept for me. Today I watched a model run in front of me.
Before this, "deploying a large model" felt like work reserved for engineers. Today my job was to define the goal clearly.
I increasingly feel the future dividing line may not be "can you code," but "can you define the problem well enough for AI to execute."
A liberal-arts background. A single domestic GPU. DeepSeek R1 running locally.
That result itself matters more than benchmark numbers.
Postscript: the speed was so poor that I checked the hardware and discovered the memory channels were not fully populated 🤦