编程远未解决：Alex Ewerlöf质疑AI编码能力的现实局限

技术评论者Alex Ewerlöf撰文反驳"编程已解决"的论断，指出尽管大语言模型在代码生成方面表现突出，但现实软件开发中的维护、可靠性、安全性等非功能性需求仍未被AI妥善解决1。他认为，虽然代码创建成本已大幅降低，但这些非功能性需求却占据了软件开发成本的绝大部分1。

Ewerlöf强调LLM存在根本性的应用限制1。由于大语言模型的输出具有随机性和非确定性，即使在相同输入下也可能产生不同结果，这种特性与确定性代码的需求相悖1。更为关键的是，LLM无法被追究责任，这使其不适合医疗、金融、汽车、国防等对风险容限要求极低的领域应用1。

在论述所有权框架时，Ewerlöf提出需要三个支柱的观点：知识、授权与问责制1。他指出，只有个人软件、概念验证产品和"武装AI"三类无需完全理解代码即可运行，而大多数专业软件均需要低风险容限和完整的问责机制1。Shopify首席执行官Toby Lutke的经历印证了这一点——他在一年前鼓励员工使用AI，后来将所得结果讽刺地称为"垃圾手雷"1。

Ewerlöf最后倡导"缓即快"的开发哲学，认为花时间充分理解代码能防止代价高昂的故障，相比盲目追求速度更具经济效益1。

Alex Ewerlöf has published a detailed critique challenging the notion that coding has been solved by artificial intelligence.1 While large language models have demonstrated impressive capabilities in code generation, Ewerlöf argues that the reality of professional software development extends far beyond writing functional code.1 He contends that non-functional requirements such as maintenance, reliability, security, and scalability account for the majority of development costs, challenges that current AI systems have not adequately addressed.1

A central concern in Ewerlöf's analysis is the fundamental unsuitability of LLMs for high-risk applications.1 He notes that large language models cannot be held accountable for their outputs, and that "the worst that can be done to it is pulling the plug," making them inappropriate for sectors like healthcare, finance, automotive, and defense where low risk tolerance is essential.1 Additionally, Ewerlöf highlights a critical technical limitation: the inherent randomness and non-deterministic nature of LLM outputs means that identical inputs may produce different results, a characteristic fundamentally at odds with the deterministic behavior required of production code.1

The issue of ownership and responsibility underpins Ewerlöf's argument, which rests on three pillars: knowledge (understanding problems and technical constraints), authorization (decision-making trust), and accountability (responsibility for failures).1 He illustrates his point by referencing Shopify CEO Toby Lutke, who encouraged employees to adopt AI tools a year prior but later described the results as "slop grenades."1 Ewerlöf acknowledges that some categories of software—personal applications, proof-of-concept projects, and "armed AI" systems—may not require complete code comprehension, but he maintains that the majority of professional software demands both low risk tolerance and clear lines of accountability.1 He advocates for a "slow is fast" philosophy, arguing that investing time to understand code prevents costly failures and ultimately proves more economical than pursuing speed at any cost.1