> It's not a "silver bullet" that magically eliminates complexity, but it is a new form of leverage for managing it.

I think you're right but people are treating it like the silver bullet. They're saying "actually the AI will just eliminate all the accidental complexity by being the entire software stack, from programming language to runtime environment."

So we use the LLM to write Python, and one day hope that it will just also eliminate all the compilers and codegen sitting between the language and the metal. That's silver bullet thinking.

What LLMs are doing is managing some accidental complexity, but it's adding more. "prompt engineering" and "context engineering" are accidental complexity. The special config files LLMs use are accidental complexity. The peculiarities of how the LLM sometimes hallucinates, can't answer basic questions, and behaves differently based on the time of day or how long you've been using it are accidental complexity. And what's worse, it's stochastic complexity, so even if you get your head around it, it's still not predictable.

So LLMs are not a silver bullet. Maybe they offer a new way of approaching the problem, but it's not clear to me we arrive at a new status quo with LLMs that does not also have more accidental complexity. It's like, we took out the spec sheet and added a bureaucracy. That's not any better.