How GPT‑4 Is Driving the Leap to Software Engineering 3.0

Software Engineering 1.0

Software Engineering 1.0 originated around 1968 when the software crisis prompted Frederick P. Brooks Jr. to describe software development as a struggle of prehistoric beasts in a tar pit. The era adopted product‑oriented, structured, process‑driven, quality‑managed, phase‑clear, responsibility‑defined and documentation‑standardized practices borrowed from civil engineering.

These characteristics are illustrated in Figure 1, the cover of "The Mythical Man‑Month".

Software Engineering 2.0

Influenced by the Internet, open‑source, Agile and DevOps, Software Engineering 2.0 treats software as a service (SaaS), emphasizes value delivery, puts people before processes, adopts self‑managed teams, continuous delivery, integrates development, testing and operations, and focuses on knowledge management and fun. Figures 2 and 3 show the Agile Manifesto and the timeline of the three eras.

Software Engineering 3.0

The breakthrough of large language models such as GPT‑4 (RLHF‑trained, multimodal, 25 k token context) enables AI to generate code, detect errors, design architectures, write test cases, and even produce BDD specifications. Figures 4‑8 illustrate GPT‑4’s capabilities in requirement analysis, functional decomposition, acceptance‑criteria generation, and BDD output.

Beyond coding, GPT‑4 assists in software design by offering suggestions, identifying design patterns, analyzing and optimizing architectures, and sharing best‑practice knowledge. It also supports knowledge sharing, meeting summarization, and collaborative brainstorming, thereby boosting team productivity.

Integrating GPT‑4 into CI/CD pipelines (Figure 10) enables automated error detection, patch generation, and continuous delivery. Figure 11 depicts the new three‑ring DevOps model where model‑driven development and operations form a seamless loop.

The "Software Engineering 3.0 Manifesto" stresses that human‑AI interaction now outweighs individual developer skill, that business and development data dominate processes and tools, that code‑generating models surpass hand‑written code, and that asking good questions is more valuable than merely solving problems.