Can AI Replace Programmers? Experts Debate AI's Limits in Complex Software Engineering

Overview

A Zhihu discussion gathers multiple expert perspectives on whether AI programming can replace human programmers. The consensus: AI excels at localized coding tasks but falls short on the deep contextual understanding, meta-level reasoning, and domain-specific intuition required for complex software engineering challenges.

Dr. Zhang Gang: Context and the Human-in-the-Loop Necessity

Dr. Zhang notes that while AI could theoretically produce sophisticated solutions (solutions 2, 3, 4) given sufficient context and methodology, such context is rarely available before encountering a problem. AI cannot be "on site" to identify which factor is decisive. Because AI solves problems that are not its own, it cannot close the loop; humans must remain "in the loop" at some level, sometimes at a low level where detailed solution evaluation is required.

Professor Peng Xin: Meta-Values and Deep Inquiry

Professor Peng argues that for complex problems, humans also cannot know which context matters—if they did, they would solve the problem themselves. Humans tackle complexity through prolonged exploration guided by higher-level meta-values and consciousness, using rigorous multi-angle logical verification. AI's chain-of-thought demonstrates only shallow reasoning; it is unlikely to sustain deep, value-guided inquiry.

Wang Hongzhu: Improved Human-AI Collaboration Pattern

Wang suggests a better workflow: first let AI deeply learn the project, then generate multiple candidate solutions from different angles, ask AI to justify each, and finally have the human select. AI can also search for industry best practices on similar projects and compile a report for synthesis.

Group Member Liuda: Code Aesthetics from Experience, Learnable by AI

Liuda contends that aesthetic judgment in code comes from the pain of maintaining others' code, not innate intuition. AI's evolution speed is underestimated; trained on software design books and massive code corpora, AI can master code aesthetics. Review systems can feed human review records to AI, extracting hundreds of rules—some not yet fully generalizable due to data volume.

Group Member Huichen: Historical Technology Transition Analogy

Huichen compares the question to whether IDEs replace vim, cars replace horse carriages, automatic transmissions replace manual, or EVs replace gas cars—implying a natural progression where tools augment rather than eliminate the need for human operators.

Author's Perspective: Local vs. Global Information and the "Form Arises from Mind" Principle

The author observes that the referenced answer (likely by a nginx expert) shows AI producing a workaround based on local information, while a more comprehensive architectural redesign would require global exploration. Human-AI interaction determines how much each contributes. The author emphasizes that LLM output gains meaning only through human interpretation shaped by experience. In unfamiliar domains, the author found themselves unable to ask good questions, understand problems, or judge solutions—highlighting that effective AI use demands domain expertise.

Software Complexity and the Linus Torvalds Case

Complex software exhibits opaque causal relationships, e.g., a complex use-after-free (UAF) memory issue. The author cites a recent case where AI declared a kernel memory bug "hopeless," but Linus Torvalds rebooted the kernel 18 times to isolate and fix a two-year-old round_up() issue, proving AI wrong. This illustrates that deep debugging requires human persistence and system-level intuition.

Current State: Agents as Tools, Humans as Judges

LLM-based agents are powerful tools but always require human judgment and understanding, which stem from deep domain practice and lived experience. The author attributes recent layoffs not to AI but to a lack of new business demands; AI serves as a convenient scapegoat.

Referenced Articles: The Hard Parts of Software

Linked articles argue: "Simple problems are simple however analyzed; complex problems are complex however handled. Making complex problems simple is a problem." The hardest parts of software are domain modeling and architecture design—not building demos. If AI solves those hard parts, it would constitute true intelligence. Echoing Kai-Fu Lee, most current work is "cosmetic" rather than addressing the hard parts that drive cash flow and profit: strategic analysis, decoding, and decision-making.