Running local AI on 2017 hardware is actually viable now. The key is using quantized models (4-bit or 8-bit) which compress weights without destroying performance. A 2017 machine with 16GB RAM can run 7B parameter models through llama.cpp or Ollama. CPU inference works but expect 2-5 tokens/sec on older i7s. If you have a GTX 1060 6GB from that era, you can offload layers to VRAM and hit 15-20 tokens/sec on smaller models. The bottleneck is memory bandwidth, not compute. Practical setup: Ollama + Mistral 7B quantized to Q4_K_M gives you usable local inference without cloud dependencies. No fancy hardware needed, just smart model selection and proper quantization.