Human + AI > AI: Restructuring Software Engineering for the AI Era

The author begins by referencing Dr. Tian Yuandong's 2025 year-end summary, which introduces the concept of a societal "Fermi level" — a shift where human value is no longer measured by personal labor output but by the ability to amplify AI capabilities.

Drawing from a background in software system engineering, the author notes that earlier attempts to codify human expertise relied on deterministic platforms and documentation — often ineffective because tacit knowledge remained in people's heads and documents went unread. With large language models, expert experience can now be learned and embedded directly into AI systems, enabling future humans to leverage AI as a living knowledge base.

Four Dimensions of AI's Impact on the Software World

1. Agent / AI-Native Software. Traditional deterministic software is built by hand-coding logic (loops, branches, sequences). Many real-world tasks — translation, image recognition — cannot be captured by rule-based logic. Deep learning enabled specialized neural networks; now LLMs plus agents can handle multimodal input and expand the boundary of what software can do without task-specific training.

2. AI4SE (AI for Software Engineering). Tools like Cursor and Claude Code accelerate coding (web, apps) and assist with code review, comprehension, and other development tasks, dramatically boosting developer productivity.

3. AI4System. This has two facets: (a) embedding AI capabilities (e.g., Q&A) into traditional software to enable features previously impossible with deterministic algorithms; (b) adding learning and adaptive capabilities to traditional software. The author previously explored this in "Software Adaptability Trends" (https://mp.weixin.qq.com/s?__biz=MzI3NDYwMDM4MQ==∣=2247484311&idx=1&sn=f80c976d933dc9e8b52a4234d068d0e0&scene=21#wechat_redirect).

4. SE4AI / SYSTEM4AI. Building robust runtime environments for AI software. The author references a prior piece "What Are We Talking About When We Discuss SE4AI?" (https://mp.weixin.qq.com/s?__biz=MzI3NDYwMDM4MQ==∣=2247484534&idx=1&sn=dde6584a4faac4ff57dddb44d7b12c9a&scene=21#wechat_redirect) but does not elaborate here.

Emerging Hybrid Architecture

Under LLM augmentation, software gains multimodal processing, reasoning, and code generation. However, deterministic traditional software remains indispensable. The future trend: traditional software handles deterministic execution and evolves stronger adaptability — usable by both humans and agents. Agents take on input processing, task orchestration, and output interaction, while traditional software provides the reliable deterministic core.

Actionable Directions for Senior Engineers

Facing this shift, the author proposes three concrete questions to guide future work:

Restructure legacy systems with AI. Re-examine past software (e.g., DFx fault analysis capabilities) and ask whether they can be rebuilt or replaced using AI.

Replace heuristic/rule-based algorithms with LLMs. Example: predicting the next app a user will open. The evolution went from rule-based heuristics → Bayesian learning → small neural networks → now potentially a large model. The author asks whether the latest step can be an LLM.

Fuse knowledge engineering with AI. Capture knowledge, decisions, and expertise more effectively so they become living digital experts that continuously grow.

These directions emphasize turning static, human-dependent artifacts into dynamic, AI-enhanced assets that persist and improve beyond individual contributors.