Finding Patterns: From Middleware to Modern AI

writing a Star Trek Basic game from memory after a context reset, the Z80 chip and assembly programming via a disassembly and debugging monitor, buying Turbo Pascal directly from a Munich garage 500 meters from home, studying computer science with C and Turbo Pascal, building a functional interpreter for a compiler generator with visual UI at university, joining Siemens for a research project, working on the C++ standardization proposal for meta-information and reflection through source-code instrumentation, the influence on the C++ ISO standard, the path to writing the POSA (Pattern-Oriented Software Architecture) book in 1996 in parallel with the Gang of Four book, coordinating with Erich Gamma to avoid overlap, finding patterns in middleware platforms like COM, DCOM, CORBA, and RMI rather than inventing them, broker, microkernel, pipes and filters, blackboard, layers, publisher-subscriber, forwarder-receiver, client-dispatcher-server, and whole-part patterns, learning Java from Java in a Nutshell by David Flanagan, skepticism about Java replacing C++ that proved wrong, Java Competence Centers at Siemens and Sun Microsystems, Java mobile edition for handsets and the rise of enterprise Java, Siemens application server limited to stateless EJBs, embedded systems at Siemens written in C++ and rust for ICE trains and medical tomographs, Java for enterprise services and monitoring, LLM disillusionment on the Gartner hype cycle, the plateau in transformer-based models, the brute-force compute investment vs specialized models, AGI skepticism rooted in transformer architecture limits and missing context memory and consciousness, physical AI cooperation between Nvidia and Siemens, the n times square root of n attention scaling, the practical context window limits below the advertised million tokens, errors during code generation with Claude Sonnet 4.5, the keep-it-simple principle and Occam's Razor, simplicity as an art, the role of naming, applying patterns to existing systems rather than coding them as building blocks, over-abstraction of HTTP behind custom adapters as an anti-pattern, Convention over Configuration in Quarkus with JAX-RS and Model Context Protocol annotations, grounding LLMs with publicly available Java and Jakarta EE specifications written in normative RFC 2119 language, the Ivar Jacobson BCE (Boundary Control Entity) structure as a stable scaffolding for code generation, vibe coding limited by specification coverage

Prof. Dr. Michael Stal on LinkedIn: /in/drstal/