AI: A Tool for Software Architecture?
summary
Surveyed architects name communication as the biggest challenge, and experiments show LLMs often fabricate even when given source data: JQAssistant + Neo4j plus an LLM wrongly reports Apache ActiveMQ and Artemis in a system, and an LLM summary of a YouTube video is incorrect. Historical examples (the Eliza effect) and a Nature paper find LLMs highly persuasive. Useful, constrained uses include RAG-backed reviews, drafting docs and ADRs, exam-style practice, but expect intrinsic hallucinations and avoid creating an architecture theater without code-level verification.
AI-written summary · about
chapters (8)
- 0:00Introduction and context
- 5:06AI and software architecture challenges
- 13:17AI reliability and hallucinations
- 20:04People too easily convinced by AI
- 27:03Can LLMs be reliable tools?
- 32:21Practical uses of AI in architecture
- 43:15Ethics and limitations of LLMs
- 44:38Conclusion and Q&A
related talks
14mSAG Expert Interview #03: Impact of AI in Software Architecture
01h 49mAI engineering with open access LLMs that lie, curse, and steal
47mHarnessing LLMs for Next-Level Security Analysis
24mWill an LLM replace me? A Systematic study of Gen AI's impact on data jobs
33mAI is coming for your job
40mLLMs & Data Warehouses: Reliable AI Agents for Business (Keynote)
54mAI in Testing: Hip or Hype?
44mAI Architecture Thinking: From Idea to Impact