The LLM Critics Are Right. I Use LLMs Anyway
257 points
• 1 day ago
• Article
Link
许多现代工程师，包括处在以本地为先（local-first）软件前沿的那些人，对大型语言模型（LLMs）感到认知失调。尽管人们普遍认可对 LLMs 的合理批评——比如它们容易生成低质内容、依赖受版权保护的素材、带来环境问题，以及由大型科技公司垄断引发的伦理困境——但许多专业人士在日常工作中仍大量使用这些工具。这就造成了一种尴尬局面：一方面专家在演讲中提醒自动化代理的风险，另一方面又在用它们写代码，这种矛盾在职业圈子里越来越常见。
问题的核心在于协作空间中信任的流失。那些曾经默认贡献者会在 pull request 中投入实质性劳动的项目，如今面临大量自动生成的低质量提交。维护者不得不通过自动关闭 PR 或者通过现实世界的人工验证来审查贡献者，以维护项目的完整性。这一变化也让传统的师徒制变得复杂：资深工程师难以分辨一位真正付出过努力的初级工程师和将任务外包给 AI 的人，这可能损害新人的培养与成长。
尽管缺点明显，许多工程师认为不能忽视 LLMs；如果有意识地使用，它们可以成为放大人类思维的强力工具。将模型部署在私人硬件上本地运行，可以降低对大公司的依赖，避免地缘政治干预或服务中断带来的影响。关键不是让机器替你思考，而是用它来磨炼、结构化并挑战自己的想法——把模型当作陪练而不是真理的裁决者，用来反复推敲概念、指出潜在陷阱或充当"魔鬼代言人"。
应对这一局面的关键在于采用以人为本且严谨的工作流程，以抵消 LLM 输出固有的被动性。比如不断向 AI 提问以确保共识、把工作严格限定在清晰的问题陈述之内，或者引入子代理（subagents）来有意尝试推翻自己的逻辑——这些做法能把一种容易生成废话的工具，转变为成熟的认知支持系统。但这也要求使用者具备足够的领域专业能力，能分辨高质量成果与劣质结果，因为 AI 往往默认选择最流行或最平均的解法，而非最具创新性或技术上最合理的方案。
归根结底，高质量工作与 AI 产生的噪音之间的区别仍取决于人的投入与责任。如果一个工程师愿意为自己的产出负责，并能充满信心地朗读它，那么这就不应被视为自动化的垃圾。既然这些工具只是放大了用户原有的意图与专业性，行业面临的挑战是通过透明且经过深思的参与来重建信任。通过分享经验并发展更好的集成模式，社区可以走出当前的失调状态，朝着更有成效、目的更明确地使用这些技术方向前进。
Many modern engineers, including those at the forefront of local-first software, find themselves in a state of cognitive dissonance regarding large language models. While there is broad consensus on the valid critiques of LLMs. such as their tendency to generate low-quality content, their reliance on copyrighted materials, environmental concerns, and the ethical dilemmas surrounding big tech monopolies. many professionals continue to utilize these tools extensively in their daily work. This creates an awkward environment where experts give presentations warning about the dangers of automated agents while simultaneously using them to code, a tension that is increasingly common in professional circles.
The core of the issue lies in the erosion of trust within collaborative spaces. Projects that once relied on the assumption that a human contributor had invested significant effort into a pull request now face a flood of automated, low-quality submissions. Maintainers are forced to resort to auto-closing PRs or vetting contributors through manual, real-life verification just to preserve the integrity of their projects. This shift also complicates the traditional mentorship model, as seniors struggle to differentiate between a junior engineer who has genuinely put in the work and one who has simply outsourced their tasks to an AI, potentially devaluing the development of new talent.
Despite these significant drawbacks, many engineers maintain that LLMs cannot be ignored and, when used with intention, act as powerful force multipliers for human thought. By focusing on models that can run locally on private hardware, developers can reduce their reliance on corporations and shield themselves from geopolitical interference or sudden service outages. The goal is to avoid letting the machine do the thinking, but rather to use it to sharpen, structure, and challenge one's own existing ideas. This involves treating the model not as an oracle of truth, but as a sparring partner that can iterate on concepts, highlight potential pitfalls, or serve as a devil's advocate.
A key to navigating this landscape is the adoption of rigorous, human-centered workflows that counter the inherent passivity of LLM-generated outputs. Techniques such as relentless interviewing of the AI to ensure a shared understanding, strictly limiting the scope of work to clear problem statements, and using subagents to intentionally attempt to break one's own logic can transform a tool prone to generating slop into a sophisticated cognitive support system. This requires the user to possess enough domain expertise to recognize high-quality results from inferior ones, as an AI will often default to the most popular or average solution rather than the most innovative or technically sound one.
Ultimately, the distinction between high-quality work and AI-generated noise remains a matter of human effort and accountability. If an engineer is willing to stand behind their output and read it aloud with full confidence, it transcends the definition of automated junk. Because these tools merely amplify the underlying intent and expertise of the user, the challenge for the industry is to foster an environment where trust is rebuilt through transparent, thoughtful engagement. By sharing these experiences and developing better patterns for integration, the community can move past the current state of dissonance toward a more productive and intentional use of these technologies.
269 comments • Comments Link
• LLMs 可以增强现有技能、结构和观点，成为快速开发的强大工具；但如果不主动锻炼基础能力，也会带来长期认知退化的担忧。
• 生产力的衡量仍有争议：人们感知到的收益往往表现为"省下的时间去做副业"，而不是组织产出的实质增长，导致个人利益与企业 ROI 之间存在落差。
• 把不常做但复杂的任务（例如编写 regex）交给 LLMs 存在风险，会削弱人们判断何时以及如何使用这些工具的直觉，实际上把关键决策过程外包掉了。
• 在 AI agent 上投入大量 token 并不能保证质量或成功；许多开发者发现，过度依赖自动化循环、缺乏深入人工审查，会导致难以维护且充斥低劣代码的代码库。
• 人们合理担忧初级员工的培养受到影响：高级工程师把琐碎任务交给 agents，而不是把这些任务当作团队新成员成长的机会。
• 批评者拒绝把 LLMs 简化为"计算器"类比：由于幻觉（hallucinations）和抽象泄露（leaky abstractions），技能验证比使用简单计算工具更困难也更重要。
• 一个务实的折中做法是把 LLM 的输出视为"draft zero"，必须通过严格的定制界面和系统提示词（system prompts）来强制执行质量，而不是盲目信任原始结果。
• 科技行业对 AI 的快速采用往往由自身利益和竞争压力驱动，这加剧了这样一种焦虑：未能整合这些工具的人最终会被更高效的同行淘汰。
• 对 LLM 整合的批评不必然等同于卢德主义（luddism）；它是划定道德边界、承担环境责任并防止机构性知识丧失的重要组成部分。
• 根本的紧张在于信任：生成代码和文本的便利削弱了传统的验证过滤，迫使专业人士决定自己到底是创造者，还是仅仅在协调自动化输出。
这场讨论反映出 LLMs 带来的效率与对技术工艺流失的恐惧之间的深刻矛盾。有人把这些工具视为适应者的必然演进；也有人看到核心技能被侵蚀、以及掩盖人类付出的低质量产出在增加。关于生产力并无共识：有人借助工具缓解倦怠、重振职业生涯；也有人花时间修复平庸的代码而感到浪费。归根结底，辩论的焦点是行业是否会为了短期速度牺牲长期稳定性与深厚专业知识——整个社区因此分裂为两派：一派拥抱"vibecoding"的未来，另一派坚持人工验证与传统精通的严守。 • LLMs amplify existing skills, structure, and opinions, making them powerful for fast-paced development while raising concerns about long-term cognitive atrophy if foundational muscles are not actively exercised.
• Measuring productivity remains controversial, as perceived gains often manifest as time saved for side projects rather than increased output for organizations, leading to a disparity between individual benefits and corporate ROI.
• The reliance on LLMs for infrequent but complex tasks, like writing regex, risks eroding the intuitive understanding of when and how to apply these tools, effectively outsourcing critical decision-making processes.
• Significant token spend on AI agents does not guarantee quality or success, as many developers find that excessive reliance on automated loops without deep manual review results in unmaintainable, "slop-filled" codebases.
• There is a valid concern that junior training is suffering, as senior engineers now delegate mundane tasks to agents instead of using them as growth opportunities for newer team members.
• The "calculator" analogy for LLMs is rejected by critics, who argue that LLM hallucinations and leaky abstractions make skill verification far more difficult and essential than with simpler computational tools.
• A pragmatic middle ground involves treating LLM output as "draft zero," requiring rigorous custom interfaces and system prompts to enforce quality, rather than trusting vanilla outputs blindly.
• The tech industry's rapid adoption of AI is often driven by self-interest and competitive pressure, fueling an anxiety that professionals who fail to integrate these tools will eventually be rendered obsolete by more efficient peers.
• Criticizing LLM integration is not necessarily luddism but a vital part of defining ethical boundaries, environmental responsibilities, and ensuring that new technology does not lead to a systemic loss of institutional knowledge.
• The fundamental tension lies in trust, as the ease of generating code and text has broken traditional verification filters, forcing professionals to decide whether they are creators or merely orchestrators of automated outputs.
The discussion reflects a deep-seated tension between the undeniable efficiency of LLMs and the fear of losing technical craftsmanship. While many view these tools as an inevitable evolution that favors the adaptable, others see a dangerous erosion of fundamental skills and an increase in low-quality output that obscures human effort. There is no consensus on productivity, as experiences range from career-saving burnout reduction to wasted time spent fixing mediocre code. Ultimately, the debate centers on whether the industry is sacrificing long-term stability and deep expertise for short-term velocity, with the community split between those who embrace the "vibecoding" future and those who maintain a stubborn commitment to manual verification and traditional mastery.