Beyond Tools: How Large Models Are Driving Software Engineering 3.0

Reviewing the software industry’s history, each paradigm shift—waterfall and V‑model in SE 1.0, agile and DevOps in SE 2.0—re‑examined the essence and boundaries of software, improving maintainability, scalability, and delivery speed.

Software Shape Transformation: From Programs to Models

Traditional software is a deterministic system of algorithms and data structures. With the rise of large language models (LLMs), a new core element—"model"—appears, turning software into an entity that can autonomously influence system behavior.

Traditional program : algorithm‑ and data‑structure‑driven, excels at deterministic problems.

Large model : massive parameters and high‑dimensional vectors, handles uncertainty and large‑scale inputs, produces probabilistic answers, and searches for near‑optimal solutions.

This program‑plus‑model synergy pushes software toward extending human cognition.

Human‑Machine Collaborative Development Paradigm

In SE 3.0, the key question becomes how humans and LLMs complement each other. While agile and DevOps remain human‑centric, LLMs introduce new challenges:

How to efficiently co‑develop and maintain features, deciding which stages the model leads and where human judgment remains essential?

How to ensure explainability and verifiability, mitigating risks from model uncertainty?

How to achieve "self‑evolution" through adaptive, data‑driven optimization?

Agents can retrieve and understand requirements, design, code, review, test, and fix defects, dramatically boosting development efficiency.

Quality Assurance Challenges

Unpredictable model behavior makes test case design and reproducibility difficult; lack of explainability hampers fault localization.

Security and robustness are threatened by adversarial attacks, data loss, and abnormal inputs.

Massive, high‑quality data is required; new evaluation metrics must ensure completeness and privacy.

Fairness and ethics demand attention to algorithmic bias and subjective risks.

Key Considerations for Moving Toward SE 3.0

Redefine development processes: treat the large model as a core asset integrated into every stage.

Establish new assessment systems that measure both traditional software quality and intelligent output trustworthiness.

Expand team roles to include algorithm, data, and model‑training experts.

Strengthen ethical compliance to pre‑empt technical and moral hazards.

Scholars suggest that LLMs will usher software into a "quantum era," highlighting a paradigm clash between model computation and classic programming logic. The coexistence of models and programs forms the logical core of SE 3.0, extending software from pure computational capability to cognitive decision‑making.

In the next decade, human‑machine collaboration will become the norm, manifesting as paired analysis, design, coding, and testing, where the "machine" denotes one or multiple intelligent agents. Despite advanced technology, software ultimately serves people; leadership, creativity, and value judgment remain uniquely human.

Large models will have a long‑term impact far beyond current imagination, similar to the transformative effects of the Internet and agile development a decade ago. Their true value will be fully appreciated only in hindsight, as they evolve software systems from "computational extension" to "cognitive decision evolution," using uncertainty to solve certainty.