Concerning Fred Brooks "No Silver Bullet", I disagree on this conclusion:
> Modern AI has thrown a wrench into Brooks’ theory, as it actually does reduce essential complexity. You can hand AI an incomplete or contradictory specification, and the AI will fill in the gaps by cribbing from similar specifications.
In your writeup, it seems like you're arguing that LLMs can't eliminate essential complexity, and I agree that they probably can't.
But I do think they can reduce essential complexity.
As a concrete example, here's me prompting Claude 4.1 Opus to define a domain-specific language for creating computer-generated paintings.[0] I just provided the requirements and left a lot of ambiguity in the specifics.
In that example, has LLM reduced essential complexity at all?
To me, the answer is clearly yes. It wrote a spec based on my requirements. I could potentially do better if I defined it from scratch, but if the LLM-generated spec is good enough, then it likely isn't worth the cost of me doing it myself for the marginal improvement in quality.
When LLMs first came out, I felt like I had no need for them because I think I can write code better than they can, and I enjoy writing code. But as I've started experimenting with them, I'm realizing that there are some problems that I can solve with software that I don't actually enjoy implementing and I don't care that much about specifying every aspect of my program's behavior, so LLMs fit well in those situations and eliminate essential complexity that would otherwise fall in my lap.
The only complexity AI reduces is the cognitive complexity of writing it. The code itself almost certainly will not be free of Brooks' nonessential complexity, and the reader is SOL.
It's like putting on an exo suit, lifting something very heavy and putting it on a shelf, then asking your teammate to go paint it.
> Modern AI has thrown a wrench into Brooks’ theory, as it actually does reduce essential complexity. You can hand AI an incomplete or contradictory specification, and the AI will fill in the gaps by cribbing from similar specifications.
The essential part is still not adequately covered by Generative AI, and probably never will be. Here is my detailed write-up about it: https://smartmic.bearblog.dev/no-ai-silver-bullet/