640K of Intelligence Should Be Enough

“Good enough” is a perfectly respectable benchmark. It is also a terrible fortune teller.

The Old Line Wearing New Shoes

The claim that inexpensive Chinese AI models are good enough for most users has a familiar clank. It sounds like the old “640K ought to be enough for anybody” line: commonly attributed to Bill Gates, apocryphal, and still useful because everyone gets the joke. Technology keeps humiliating confident ceilings.

Chinese open models may be excellent, competitive, inexpensive, and strategically important. This is not a complaint about them. It is a complaint about the sleepy phrase “good enough,” which describes a moment in time, not a durable technology strategy.

A few years ago, basic text generation felt magical. Now people expect research, code, tool use, multimodal reasoning, workflow automation, and agents that can stay useful beyond one prompt. As models improve, applications expand to consume the improvement. Better does not sit politely in a corner.

Cost still matters. A cheaper model can absolutely be the correct choice for a support summarizer, internal classifier, batch extraction job, or low-risk agent step. Please do that. Spend money like an adult.

But that does not mean greater intelligence, context, accuracy, reliability, or reasoning depth will go unused. Agentic systems make the gap more visible. One workflow may involve dozens or hundreds of model decisions. A small reliability difference can compound from “mostly fine” into “why did the agent rewrite billing because a bullet point looked lonely?”

The winning strategy is not pretending the race ended at the current baseline. It is matching models to workloads while continuing to raise the frontier. Use cheaper models where they fit. Use stronger models where failure, ambiguity, depth, or autonomy demand it.

“Good enough” has a long history in technology. It usually lasts until someone discovers what better makes possible.