AI Slop 与 Craft 危机
What it is
"Slop" refers to AI-generated output that is technically functional but generic, incoherent, or devoid of human judgment. The term captures a systemic risk: as AI lowers the cost of production to near zero, the volume of mediocre work rises, and the human taste required to distinguish good from bad becomes the scarce resource.
Why it matters
When anyone can generate a complete product in an afternoon, the barrier to creation drops, but the barrier to quality rises. Without deliberate craft — the final 10% where a human applies taste — AI output converges toward sameness. This affects software, design, writing, and any creative field where AI is deployed as a production tool.
Key points
- Software lost its soul before AI arrived. Ryo Lu (Cursor Design Lead) traces the erosion of software personality to the rise of growth metrics and A/B testing, which standardized design systems and flattened product character. AI accelerates this trend but did not start it.
- "Overcooking" is a new form of design debt. When the cost of adding features drops to near zero, individually sensible decisions accumulate into collective incoherence. A dashboard with sparklines on every number, confirmation modals on every action, and animated illustrations on every empty state is not wrong in any single choice, but wrong as a whole. Attention is finite; overload dissipates focus.
- Most "new ideas" are repackaged old concepts. Ryo Lu argues that many AI-generated features are new stickers on old concepts. The thinking does not go deeper; it duplicates itself into confusion. The solution is not more tools that make more slop, but the discipline to see through chaos and cut everything that does not serve the core purpose.
- Craft lives in the last 10%. Peter Yang (Product, Roblox) notes that AI is useful for generation, but craft is in the manual application of taste to make something you can be proud of. Many people never bother with this step. This echoes Kieran Klaassen's "polish" phase in Compound Engineering: the human judgment that elevates output from functional to exceptional.
- The risk is drowning, not displacement. Ryo Lu's core fear is not that AI will replace quality practitioners, but that it will flood the market with mediocrity, making it harder for crafted work to surface.
- Slop as experimental fuel. Mitchellh argues that slop is what enables fast parallel experimentation. The etiquette is not to avoid slop, but to understand when you are in "exploration mode" (where slop is valuable) versus "delivery mode" (where craft matters). This reframes slop from a problem to a phase.
- Quality requirements follow dependency, not the builder's identity. Ben Tossell accepts rough code for personal, disposable tools only if it works, is secure, and performs adequately. The tldraw internal-tool guide he links marks the next boundary: once colleagues rely on the result, existing review, identity, secrets management, persistence, monitoring, and ownership become part of delivery. Cheap generation reduces entry cost; it does not prove that those obligations disappeared.
Evidence across sources
| Source | Key Claim | Relevance |
|---|---|---|
| AI 简报 2026-03-31 | Software personality eroded by metrics and A/B testing before AI | Pre-AI root cause of homogenization |
| AI 简报 2026-04-01 | AI risks drowning quality practitioners in mediocrity | Framing the "drowning" risk |
| AI Builders Digest 2026-04-23 | "Overcooking": individually sensible decisions accumulate into incoherence; craft is the last 10% | Mechanism (overcooking) and solution (craft) |
| AI & I Podcast — The AI Sandwich | Without human polish, AI output becomes "all slop, all the same" | Independent validation from engineering practice |
| AI Briefing 2026-05-08 evening | Mitchellh: slop enables fast parallel experimentation; the skill is knowing which phase you are in | Counter-narrative: slop as a productive phase, not just a problem |
| AI Briefing 2026-05-09 morning | Mitchellh's concrete examples: internal API/GUI built entirely with slop; overnight agent-generated dozens of plugins; compares to Terraform 0.1's 3 weak providers vs today's 100 slop-generated providers | Demonstrates slop as velocity multiplier with clear guardrails (no unreviewed PRs, no customer delivery without transparency) |
| Nabeel Qureshi — What Makes Art Great? | AI slop is easy to compress (low information content); great art is hard to compress. AI picks "obvious" word choices. True craft requires Surprise, Echoes, and Depth | Formal framework for diagnosing why AI output feels dull; bridges aesthetics and engineering judgment |
| AI Briefing 2026-09-01 Afternoon | DESIGN.md, code reduction, Ponytail, and AI-generated software maintenance all frame slop as an interface between taste, constraints, and review | Adds text-governed design/coding practices for reducing slop |
| Ben’s Bites 2026-10-13 | Ben distinguishes disposable personal tools from customer-critical infrastructure; tldraw explains the path from a local prototype to a maintained team application | Adds dependency and ownership to the exploration/delivery distinction, rather than judging quality by who wrote the code |
2026-09-01: Slop reduction as textual governance
来源:AI Briefing 2026-09-01 Afternoon
本次 briefing 把 slop 的治理从“人类最后 10% polish”推进到更工程化的文本治理。Guillermo Rauch 的 DESIGN.md 观点、Matt Pocock 的“减少代码”、Voxyz 对 Ponytail skill 的介绍,以及 Mole 项目的 AI 生成代码维护经验都指向同一个模式:AI 输出质量不能只靠事后品味筛选,还要把品味、边界、复用顺序和删减原则写成 agent 可读取、可复用、可审查的规则。
这里的关键不是给 agent 更多模板,而是减少不必要生成。Ponytail 的顺序是先找已有实现,再用平台/标准库,再复用依赖,最后才写新代码;Mole 的经验是用架构文档、回归测试、模块规则和 CI 检查把真实故障沉淀下来;Pocock 的观点则把 code review 定义为持续减少代码和圈复杂度。Slop 因此不只是审美失败,也是约束没有进入工作流的结果。
Open questions
- Can "taste" be systematized or taught, or is it inherently intuitive and experiential?
- Will platform-level curation (app stores, search rankings) adapt to filter slop, or will the flood overwhelm existing quality signals?
- If AI eventually learns to mimic individual taste profiles, does craft become a training-data problem rather than a human-judgment problem?
Prompts for witness
- What is the "overcooking" pattern in your own work or products? Where have you added features that made sense in isolation but weakened the whole?
- What is the last 10% of craft in your current project that you have been skipping because AI made the first 90% so easy?
- If you had to define a single "taste standard" for your team's AI-generated output, what would it be and how would you enforce it?
Related
- harness-engineering/compound-engineering — The "polish" step as an engineering methodology
- product-trends/ai-founder-guide-to-taste — Taste as a strategic asset in creative industries
- product-trends/living-software-framework — Expectation models for software that evolves vs. static tools
- product-trends/design-is-understanding-not-output — Design quality as a function of problem understanding
2026-09-03: Infinite Slop 是探索模式的 slop 案例
来源:Ben’s Bites — Build your ideas
Infinite Slop 把 AI 视频生成做成类似 Twitch 的实时频道:观众在聊天中决定提示词,系统持续生成画面。Ben’s Bites 报道其首日有 37,000 名观众,并记录了一个页面快照中的 7,735 名观看者和实时 prompt 列表。这里的 slop 不是把粗糙内容伪装成成品,而是把低成本、低承诺的生成当作探索和共同玩耍的实验层。
它补充了本页已有的“slop 可能是实验燃料”判断,但也划出边界:探索层可以快速试错,进入决策、发布或长期资产层时仍要经过事实、质量、版权和人工品味检查。观看人数是易波动的页面指标,首日观众数则是 newsletter 的转述,二者都不是内容质量的证明。
Counterpoints & Gaps
- 实验模式仍可能制造注意力噪音、版权风险和资源浪费,不能因为低成本就免除治理。
- 当前材料没有提供观众留存、内容质量或商业可持续性的长期数据。
Sources (addendum)
Pending Jean review
- product-trends/software-soul-ai-slop —
ai_draft; suggested action: define when an exploratory AI output must pass a quality, rights, and human-taste gate before reuse.
2026-09-07 afternoon: focus and measurement as anti-slop controls
来源:AI Briefing 2026-09-07 Afternoon
Shu 的时间线材料把 slop 问题压到注意力层。AI 让产出变得更容易,但也让平庸内容更容易充满工作空间。软件和创作的稀缺资源因此不是生成速度,而是专注、筛选和把最后质量做好的能力。
Paweł Huryn 的 AI writing slop 盲测提供了治理方向:slop 不只能靠主观感受识别,也可以用固定任务、盲测配对、模型间比较和一致性过滤来测量。该方法本身仍有样本和评委偏差,但它把“AI 味”从抱怨推进到可讨论的评测对象。
Counterpoints & Gaps
- Shu 的材料是创作判断,不是定量研究。
- Paweł 的 ranking 不能直接作为模型写作质量结论。当前可吸收的是评测结构。
Sources (addendum)
Pending Jean review
- product-trends/software-soul-ai-slop —
ai_draft; suggested action: define one local writing task where focus, human taste, and blind comparison all matter.
2026-09-07 evening: 产出速度上升后,专注和删减成为质量控制
来源:AI Briefing 2026-09-07 Evening
Rohit Negi 把 AI 新闻过载描述为注意力和学习方向的问题。Voxyz_ai 的删除提示则把反 AI 腔具体化为逐词删减、保留必要限定条件和幽默,并删除只重复前文的段尾总结。两条材料共同说明,抵抗 slop 不只是增加生成能力,还要把专注、筛选和删减放进交付流程。
这也延续了本页的 exploration mode 与 delivery mode 区分。快速生成可以服务探索,但进入发布或长期资产阶段时,需要用人工判断和可复用的删改规则降低噪音。
Counterpoints & Gaps
- 新闻过载判断来自个人体验,不能直接量化注意力或学习效果。
- 删除规则可能误删限定条件,需要事实审查和作者复核。
Sources (addendum)
Pending Jean review
- product-trends/software-soul-ai-slop —
ai_draft; suggested action: define an exploration-to-delivery gate with attention budget, deletion pass, factuality review, and human taste check.
2026-10-11: AI 起草把理解成本转移给评审者
来源:AI Briefing 2026-10-11 Afternoon;@guansi 的长文(708 likes)。
这条个人观察补充了 AI slop 的组织成本:AI 能快速生成术语完整、结构漂亮的方案,但作者未必真正理解内容。评审者仍要逐项查依据、判断是否适配当前项目,并确认 AI 按反馈改稿后没有引入新问题。生成成本下降后,验证与多轮复审的成本可能转移到资深审查者身上。
这里吸收的是问题机制,不把单条推文当作普遍测量结果;原文是作者与同事交流后的经验描述。
Counterpoints & Gaps
- 该观察没有给出样本、工时或跨团队数据,不能据此量化评审负担。
- 方案质量仍需结合项目风险和具体证据判断,不能仅凭“用了 AI”推断作者不理解。
- Ben 的个人工具经验和 tldraw 的交付指南没有比较不同代码质量下的长期维护工时。代理能修复问题是可用路径,不是维护成本已归零的证据;处理敏感数据的个人工具也不能省去权限检查。
Sources (addendum)
- AI Briefing 2026-10-11 Afternoon — @guansi, AI-assisted proposals and reviewer burden.
Pending Jean review
- product-trends/software-soul-ai-slop —
ai_draft; suggested action: keep the reviewer-burden mechanism as a useful hypothesis or wait for a second source with measured review effort. - product-trends/software-soul-ai-slop —
ai_draft;建议为一个个人小工具列出升级为团队依赖时必须补齐的身份、审查、监测和维护责任。
2026-10:Ryo Lu 的「向均值收敛」
来源:AI Builders Digest 2026-10-04
Ryo Lu 提出 "convergence to the mean":当所有人互相参考、AI 又把模仿循环无摩擦化,产品失去差异化,产出被推向中位数("tidal wave of mid")。解法是像艺术家一样通过具体生活形成观点,而非更多地参考已有产品。该框架为 AI slop 提供了一个生成机制层面的解释。
2026-10:Good Code 重新定义之争
poteto:clean code 一直以性能为代价,agent 让修改变便宜后,2026 年的 Good Code 应重新定义为正确、能跑、可迭代;dhh:软件架构需重想;shadcn:仍在乎可读性;kinglycrow:agent 代码长期维护有风险;thorstenball:反对工艺守护者叙事。正反方并列,判断权留给使用场景的可验证性。