Lamentation on Deploying AI

2026-10-05 · Ho Chanh Nghia

Originally published on Substack: https://hochanhnghia.substack.com/p/lamentation-on-deploying-ai

I decided to deploy AI. So now I deal with AI’s problems.

Most of my issues with AI came from this tokenism stuff. AI tokenizes to think, and tokens are demanding — every token drags the full model weights through VRAM, which is why the whole industry now races for the fastest VRAM on earth. The arms race isn’t about compute. It’s about feeding the token habit.

Meanwhile the CPU ran on ones and zeros for ages and did compute fine. I always wanted to go back to CPU and RAM. I’m not sure why people think you need a GPU to compute. Build around ones and zeros without the token process and you’re faster — the bottleneck was never the compute.

Under the old company design — instructions and process — drift and hallucination are near impossible. Write the instructions, follow the process, check the work. The machine only drifts when I let it generate instead of instructing it.

The entire AI problem is an AI-generated problem. The hallucination, the VRAM race, the throttle — none of it exists outside the token paradigm. It’s the industry’s own exhaust. And the solutions it sells are things humans already had, doable by other means.

People read books and aggregate them all the time. Some aggregate words. Some aggregate higher-order concepts. It was never an unstructured mess, and the aggregation was never new. What the machine adds is speed on the mechanical pass. What counts, what’s true — that stays mine.

That’s the lament. I decided to deploy AI, so I deal with its problems: problems it generated itself, selling me solutions I already had, to do work people already did.