Stochastic Architecture
by Vaughn Vernon
It seems that most software development professionals are still trying to sort out and understand the GenAI landscape. That’s nothing to be ashamed of. In fact, it’s wise to question the claims of “intelligence” and “reasoning” models that are released at a staggering pace.
Just how much trust should we and will we put in the LLMs that have been labeled “stochastic parrots”?
From source code assistance to using MCP servers, and everything in between, there are fundamental questions that require answers: How helpful and how problematic is GenAI? Can we make sound judgments that won’t propel us too far forward or leave us far behind the AI technology curve? Should we hold our breath and jump in or should we run the other way?
And to address the burning question for this context: How can GenAI be applied to software architecture?
Consider my experience, a software practitioner who has consistently resisted the hype and entered this space “kicking and screaming.” I’ve seen the plausibly believable hallucinations that are utterly wrong. I’ve experienced the destructive behavior—literally deleting many/all correct and working source code files and “refactoring” from code that I wrote to a broken mess. The parrots don’t “understand” architecture or “design” even when they want you to believe you are safe.
Out of need, I have forged the means to guide GenAI to produce good architectures in both client-heavy applications, server-based and serverless architectures, all while applying Domain-Driven Design. Even beyond this—and possibly the best of all—how to modernize and transform the legacy Big Ball of Mud train-wrecks to malleable architectures with a well-designed domain model that can help you bypass the innovator’s dilemma. Joy me to see for yourself.