Ask HN: You have model orders of magnitudes faster than transformer, so?

Hypothetically speaking: if you have cpu friendly ultimate "content generator" what would you do? Not saying it's the holy grail btw: so not "break satoshi's wallet", not "cure cancer" or turn mercury -> gold but say it has the power of common gpt 5+ except you don't need H100s to train it or 64gb ram but just boring old hp pc. And: context is very long + inference/generation is very fast. It may prove couple of unsolved mathematical conjecture here and there and it can make games with some effort so on. Since it's cheap and guarantees privacy (you run it locally: no SaaS). For me if it beats qwen 3.8 max or Kimi 3 I'd switch, let alone gpt 5.6 (pricey).