I asked 6 LLMs to design 100 software systems (opinion-radar.com)

In a groundbreaking experiment, researchers conducted a "LLM Stack Census" that involved six large language models (LLMs) to design 100 distinct software architectures based solely on architectural briefs without specifying any technologies. The study aimed to identify the default software stacks that LLMs tend to generate, revealing significant insights into their design preferences. The results highlighted that PostgreSQL was the most commonly selected database, while Kubernetes emerged as the favored orchestration tool across most stacks, suggesting a default inclination towards specific technologies in AI-generated systems. This research is particularly significant for the AI/ML community as it sheds light on the tendencies and biases present in LLMs when tasked with software development. By analyzing how these models respond to constraints and the diversity of their architecture choices, developers can better understand the underlying preferences of different LLM families. The methodology offers a structured approach to evaluating AI-generated architectures, emphasizing both a census-like collection of responses and the implications for real-world applications, which can influence future AI-driven software solutions. This study not only highlights prevalent tech stacks but also invites further exploration into how LLMs could be better aligned with specific project needs and constraints.