SE4AI: What Makes AI Software Engineering Different?

General Design Principles for AI Software

AI is fundamentally a type of software, so many design concerns overlap with traditional product-grade systems. The article lists functional and non-functional requirements that are not unique to AI: observability (metrics, logging), token budget protection, security compliance, user privacy handling, reliability (fault handling), availability (degradation, redundancy), performance, and security sandboxes that isolate AI instances. These are standard engineering practices required for any production system.

Unique Challenges: LLM Uncertainty

The core differentiator is the inherent uncertainty of large language models. An LLM acts as an uncertainty processor — given an input, the output can be difficult to predict. This drives specialized handling at input and output boundaries:

Safety LLMs review prompt compliance and output compliance.

Reliability mechanisms include fault diagnosis and retry logic when errors occur.

Emerging frameworks address agent orchestration, low-code workflow composition, memory engineering, context engineering, and toolbox management.

Training and Inference as Engineering Problems

Beyond application-layer concerns, model training and inference present complex infrastructure challenges: GPU resource management, distributed training frameworks, distributed inference frameworks, and operational practices akin to MLOps.

Open-Source LLM Landscape Reference

The author points to a curated landscape of open-source LLM development maintained by Ant Group, available at https://github.com/antgroup/llm-oss-landscape and a visual site at https://antoss-landscape.my.canva.site/. Two diagrams from that landscape are embedded in the article (images hosted on mmbiz.qpic.cn).

Strategic Focus Question

The piece closes by asking where to concentrate effort within the sprawling AI software stack: higher-level agent frameworks, mid-level inference frameworks, or deeper compiler-level operator optimization.