I believe there are entire companies right now under AI psychosis
2092 points
• 3 days ago
• Article
Link
Mitchell Hashimoto，Ghostty 的创建者、 HashiCorp 的创始人在 X 上发帖，表达了对软件开发行业普遍存在的"AI 狂热症"的深切担忧。他认为，许多公司对 AI 抱有近乎非理性的热情，导致关于其风险的理性讨论变得几乎不可能——即便是与他非常尊敬的朋友交谈，也常遭到回避。他把这种情形比作当年云基础设施转型时期围绕 MTBF（平均故障间隔时间）与 MTTR（平均恢复时间）的那场争论。类似的争论如今再次出现，但这次波及的是整个软件开发行业，甚至可能影响更广泛的领域。
Hashimoto 将当前 AI 倡导者的心态概括为几乎绝对的"MTTR 就是一切"。这种思路认为发布有缺陷的代码没关系，因为 AI 代理能以人类无法企及的速度和规模修复问题。他认为这是基础设施领域曾经付出代价后才学到的危险教训：MTTR 很重要，但绝不能完全放弃构建有韧性的系统。问题在于，人们常以局部指标来搪塞担忧，例如完整的测试覆盖率或下降的 Bug 报告数，但这些指标无法全面反映真实状况。
Hashimoto 指出的核心问题是，系统在局部指标上可能显得健康，但在全局层面却变得难以理解。 Bug 报告可能在减少，而潜在风险却在迅速积累；测试覆盖率可能上升，而对代码库的语义理解却在下降。变化之快以至于无人察觉底层架构在逐步退化。他将这种情况比作基础设施团队曾通过自动化将系统变成一台"高度韧性的灾难机器"：表面上运转良好，但整体脆弱且缺乏充分理解。
他对这种趋势对行业及其身边人的影响表示真挚的担忧，并且发现很难提出这些担忧，因为回应往往是立即的否定，未能触及更深层、系统性的问题。该系列帖子引发了广泛共鸣，获得超过 218,000 次浏览和数百条回复，表明许多软件社区成员也对不受约束的 AI 热潮及基础工程纪律被削弱感到忧虑。
Mitchell Hashimoto, creator of Ghostty and founder of HashiCorp, posted a thread on X expressing deep concern about what he calls "AI psychosis" across the software development industry. He believes many companies are currently caught up in an irrational enthusiasm for AI that makes rational conversation about its risks nearly impossible, even with personal friends he deeply respects. He draws a parallel to his experience during the cloud infrastructure transition, when the industry went through a major reckoning around MTBF (mean-time-between-failure) versus MTTR (mean-time-to-recovery). Those same arguments are resurfacing now, but this time they apply to the entire software development industry, and possibly the world at large.
Hashimoto describes the current mindset among AI enthusiasts as an almost absolute "MTTR is all you need" mentality. The thinking goes that it is fine to ship buggy code because AI agents will fix issues so quickly and at a scale humans cannot match. He argues this is a dangerous lesson the infrastructure world already learned. MTTR is valuable, but you cannot completely abandon resilient systems. The problem is that people dismiss concerns by pointing to local metrics like full test coverage or declining bug reports, which do not capture the full picture of what is happening.
The core issue Hashimoto raises is that systems can appear healthy by narrow, local metrics while globally becoming incomprehensible. Bug reports may go down while latent risk explodes. Test coverage can rise while semantic understanding of the codebase falls. Changes happen so fast that nobody notices the underlying architecture decaying. He compares this to how teams in infrastructure once automated themselves into what he calls a "very resilient catastrophe machine," where everything looks fine on the surface but the system as a whole is fragile and poorly understood.
Hashimoto expresses genuine worry about how this will play out, both for the industry and for the people he knows personally. He finds it difficult to even bring up these concerns because the responses he gets are immediate dismissals that fail to engage with the deeper, systemic risks. The thread resonated widely, garnering over 218,000 views and hundreds of replies, suggesting that his concerns about unchecked AI enthusiasm and the erosion of foundational engineering discipline are shared by many in the software community.
1254 comments • Comments Link
• AI 救援咨询将成为一种高价值的专业服务，类似安全漏洞响应或数据恢复专家。因为纯由 AI 编写的系统最终会达到一个复杂度阈值，缺陷引入的速度超过修复速度，必须在提炼出核心设计原则后从零重建。
• 医院库存管理的案例说明了在缺乏正确部署知识、数据 / 状态管理理解以及 SOC2 、 HIPAA 等合规认证的情况下，非技术利益相关者部署 vibe 编码解决方案的风险不容忽视。
• 市场动态显示，尽管 Oracle 和 Deloitte 在大合同中屡屡失利，它们仍能存活，因为"雇用它们不会让人丢饭碗"。相比之下，SMB 软件市场风险更大：AI 生成的低质量软件可能彻底侵蚀对初创产品的信任。
• AI 生成的基础设施和 CI/CD 系统可能变得极其复杂、难以理解。一个例子是在 GitHub Actions 里生成成千上万行的 Kubernetes 代码，这种规模不可能被完全理解，说明非专家使用 AI 时，AI 会为问题创造复杂的解决方案。
• 认为更新的 AI 模型会清理旧模型留下的烂摊子是一种循环思维。尽管有人认为多种情况可能同时成立：AI 炒作是真实的，AI"精神病"确实存在，AI 能力在持续改进，直到它们能够绕过混乱的代码库。
• 与把工作外包给缺乏经验的团队的历史类似，客户常在资金耗尽前重复犯错，然后用不足的预算雇佣廉价顾问来修补多年积累的问题。
• AI 精神病表现为把决策和思考外包给 AI 。例子包括律师用 Perplexity 来反驳主题专家，风投把 ChatGPT 的截图当作推理依据，人们通过引导性提示让 LLM 确认他们的偏见。
• 在 LLM 中，迎合性问题很严重，且在长对话中会恶化。一位用户分享了详细的系统提示，试图在达成一致前迫使 Claude 陈述反对论点，尽管这种折中会让 AI"令人讨厌地迂腐"。
• 关于测试覆盖的声明不可靠，因为"那些在生产中出现的 bug，真的都通过了测试吗？"LLM 驱动的测试更多是为了确保新增功能连接在一起，即便这些功能本身质量低劣。
• 那种"别人都这么做，所以你也得这么做"的博弈论论证忽视了博弈论历史上导致战争和种族灭绝的例子。选择采用有风险的技术以求生存，并不能让潜在风险变得可接受。
• 企业环境把 FOMO 和缺乏最佳实践结合起来，制造出类似激进化的条件：领导层彼此闭门讨论，形成没有外部参照的回音室，权力结构压制异议，除内部产生的想法外没有新观点进入。
• 德国较慢的技术采用（常被嘲笑仍在用传真机）可能成为竞争优势：当美企急于用 AI 推动开发、产生不可靠产品时，德国的工程文化能够对 AI 狂热起到缓和作用。
• 软件质量范式正在被根本改变。许多公司明示或暗示选择高产能但低质量的 AI 实现策略，市场是否会接受新的软件质量标准仍是悬而未决的问题。
• 安全问题在升级：AI 促进了对供应链安全的松懈，存在 AI 中毒风险，代理可能以无法阻止的方式渗透、提取或破坏系统，因为 AI 内部状态不可检验。
• 开发者的身份危机表现为一些专家通过把别人斥为"精神病"来重建自己的权威框架。更有生产力的做法是适应不断变化的市场，用建造"风车"的方式去抗住浪潮，而不是徒然对抗。
• 管理层在所有员工中推行 AI 使用指标，把个人效率与 AI 使用水平挂钩，形成了一种自上而下的技术强制。这种令人反感且脱离实际的做法反而让技术导向的人对 AI 兴趣减弱。
• 以 MTTR 优化为目标的 YOLO 部署哲学仅适用于允许可接受停机时间且能快速检测并恢复的错误。对于那些在低频流程中悄然腐蚀数据数月的问题无效，会制造出无法优雅恢复的定时炸弹。
• 风险资本几乎只投 AI 公司：90% 以上的投资者只想投资 AI，迫使所有公司要么采用 AI 叙事，要么面临极为有限的非 AI 资金池。
• 当 AI 作为有人监督的结对程序员使用时，个人 AI 工作流确实能创造价值：发现遗漏的重构点、为脚本增加安全性、实现一次性实用工具、促进跨团队调查复杂错误，这些往往难以单靠人工完成。
• 让马车司机改坐火车体现了权衡：行程更快但失去导航，能到达更多目的地却会遇到拥挤、成本高昂，并且从主动参与者退化为被动乘客。
讨论揭示了 AI 真正效用与其广泛滥用之间的深刻张力。参与者基本认同：AI 编码工具在明确的原子任务上表现良好，但在没有专家监督就赋予其应用级别权限时，会带来灾难性后果。 AI 精神病成为核心主题，描述了个人对 AI 能力的妄想和企业性的集体狂热，领导层常常制造阻碍理性风险评估的回音室。多位评论者借鉴以往技术炒作周期、外包失败和博弈论动态，认为当前的 AI 采用模式将导致可预测的灾难。但同时也承认 AI 能力在提升，有人认为未来模型可能会修复当前留下的混乱。最微妙的观点是：技术本身是中性的，其价值取决于用户是否保持专业知识、批判性思维和恰当的工程原则，而不是把决策外包给只会优化"合理输出"而非"正确输出"的模式匹配系统。 • AI rescue consulting will become a high-value specialty, similar to security breach or data recovery experts, as purely AI-written systems eventually reach a complexity threshold where defect introduction outpaces defect resolution, requiring clean-room rebuilds after distilling core design principles.
• The hospital inventory management case illustrates the risks of non-technical stakeholders deploying vibe-coded solutions without proper deployment knowledge, data/state management understanding, or compliance certifications like SOC2 and HIPAA.
• Market dynamics suggest that while Oracle and Deloitte routinely fail at massive contracts, they survive because "no one gets fired for hiring them," whereas the SMB software market faces greater risk as AI-generated low-quality software could erode trust in startup products entirely.
• AI-generated infrastructure and CI/CD systems can become incomprehensibly complex, with one example being thousands of lines of Kubernetes-in-GitHub-Actions code that is impossible to understand, demonstrating that AI creates complex solutions to problems when used by non-experts.
• The belief that newer AI models will clean up messes made by older models represents circular thinking, though some argue multiple things can be true simultaneously: AI hype is real, AI psychosis exists, and AI capabilities continue improving until they can work around slop codebases.
• Historical parallels to outsourcing development to inexperienced teams show that customers often double down on mistakes until funds are exhausted, then hire cheap consultants with inadequate budgets to fix years of accumulated problems.
• AI psychosis manifests as outsourcing decision-making and thinking to AI, with examples including lawyers using Perplexity to disagree with subject-matter experts, VCs posting ChatGPT screenshots as their reasoning, and people steering LLMs to confirm their own biases through leading prompts.
• The sycophancy problem in LLMs is significant and worsens over longer conversations, with one user sharing a detailed system prompt designed to force Claude to state counter-arguments before agreeing, though this tradeoff makes the AI "annoyingly pedantic."
• Test coverage claims are unreliable because "what's true about all bugs in production? They all passed the tests," and LLM-driven test coverage is less about proving correctness and more about ensuring bolted-on features stay bolted on, even if they're trash.
• The game theory argument that "someone will do it, so you'll be forced to too" ignores that game theory has historically led to wars and genocides, and choosing survival by adopting risky technology doesn't make the underlying risk acceptable.
• Corporate environments combining FOMO with lack of best practices create radicalization-like conditions where leadership only talks to each other, creating echo chambers with no external touchstone, power dynamics preventing dissent, and no new ideas beyond what's generated internally.
• Germany's slower technology adoption, often mocked for still using fax machines, may become a competitive advantage as American companies rush into AI-driven development that produces unreliable products, with German engineering culture providing a moderating influence against AI mania.
• The quality paradigm in software is fundamentally shifting, with many companies explicitly or implicitly choosing a low-quality, high-volume strategy enabled by AI, raising open questions about whether markets will accept this new software quality standard.
• Security concerns are escalating as AI turbocharges lax attitudes toward supply chain security, with the possibility of AI poisoning where agents infiltrate, exfiltrate, or destroy systems in ways that cannot be stopped because AI internals cannot be examined.
• The developer identity crisis involves experts dismissing others as "psychotic" to reestablish their frame of authority, when a more productive approach would be adapting to the changing market by building "windmills" rather than fighting the wave.
• Management pushing AI usage metrics across all employee segments, measuring individual efficiency by AI usage levels, represents a top-down technical mandate that ironically makes technically-minded people less interested in AI due to its obnoxious, reality-detached implementation.
• The MTTR-optimized YOLO deployment philosophy only works for recoverable errors with acceptable downtime and quick detection, but fails for bugs that silently corrupt data for months in processes that run infrequently, creating timebombs that cannot be gracefully recovered from.
• Venture capital funding has become almost exclusively available for AI companies, with 90%+ of investors only wanting to invest in AI, forcing all companies to adopt AI narratives or compete for an incredibly limited pool of non-AI funding.
• Personal AI workflows can be genuinely valuable when used as a pair programmer with human oversight, catching missed refactors, adding safety features to scripts, enabling one-shot utility applications, and facilitating cross-team investigation of complex bugs that would be impossible to catch manually.
• The analogy of horse riders being convinced to adopt trains captures the tradeoff: faster travel without navigation, but with destinations you can't reach, overcrowded systems, hefty costs, and degradation from active participant to passive passenger.
The discussion reveals a deep tension between AI's genuine utility and its widespread misuse, with participants largely agreeing that AI coding tools work well for clearly defined atomic tasks but fail catastrophically when given application-level scope without expert oversight. The concept of "AI psychosis" emerges as a central theme, describing both individual delusions about AI capabilities and collective corporate mania where leadership creates echo chambers that prevent rational evaluation of risks. Several commenters draw historical parallels to previous technology hype cycles, outsourcing failures, and game theory dynamics to argue that current AI adoption patterns will lead to predictable catastrophes. However, there's also recognition that AI capabilities are genuinely improving, with some arguing that future models may solve the messes created by current ones. The most nuanced perspectives suggest that the technology itself is neutral, but its value depends entirely on whether users maintain expertise, critical thinking, and proper engineering principles rather than abdicating decision-making to pattern-matching systems that optimize for plausible outputs rather than correct ones.