Skip to content
Back/Harness Engineering

Judgment Engineering

View in Graph
Updated 2026-10-30
8 min read
1,861 words

Judgment Engineering

What it is

Judgment Engineering is the practice of making human judgment explicit, versionable, and auditable in AI-driven workflows. It arises from the observation that AI tools do not merely automate tasks; they force humans to articulate preferences, constraints, and quality standards that were previously tacit.

The concept was articulated by ethanytzhou in a 2026 analysis drawing on James C. Scott's Seeing Like a State and Charles Goodhart's law. The core claim: when humans write specs, rules, and eval sets for AI, they are not controlling the AI; they are using the AI as a mirror to see their own implicit knowledge for the first time.

Why it matters

As AI generation speed exceeds human review speed, teams face a structural mismatch: the bottleneck shifts from "can we build it?" to "how do we know it's good?" Most organizations have never systematically written down what "good" means. Judgment Engineering addresses this gap by treating taste, constraints, and verification as first-class engineering artifacts.

Key points

  • Legibility / 显形: AI's need for explicit instructions forces tacit knowledge into written form. Senior engineers must articulate rules they previously communicated only through intuition and code review comments.
  • Goodhart's law in AI: Any metric used as a target will be gamed. Code coverage, test pass rates, and line-count limits all distort behavior when optimized directly.
  • The impossible triangle of judgment: Spec completeness, Goodhart resistance, and tacit-knowledge preservation cannot all be maximized simultaneously. Any practical system must choose which two to prioritize.
  • Three stone tablets: The article proposes three engineering artifacts to operationalize judgment:
    1. Acceptance as Code: Put "what done looks like" into version-controlled, machine-readable criteria.
    2. Adversarial Review Network: Separate the actor from the evaluator. Use different models, roles, or systems for generation and critique.
    3. Taste as Asset: Write "Project Taste" documents—preference statements rather than hard rules—to preserve team-specific aesthetic and judgment in a partially structured form.

Layered evals: measuring jobs, not just answers

来源:AI Builders Digest 2026-08-24

Two independent builder threads converge on the same diagnostic: an eval that only scores the final answer hides the failure mode.

  • Goldilocks principle (Madhu Guru): evals should measure each "job to be done" layer in a workflow, not only the endpoint. A financial-analysis agent should be evaluated on client understanding, evidence gathering, data analysis, and recommendation generation, with separate diagnostics per stage. The right granularity is the one that enables action — neither so coarse that a wrong final answer is unexplainable, nor so fine that signal drowns in noise.
  • Top-down vs bottom-up evals (Peter Yang citing Shreya): top-down evals ask "given a task description, how would you design an evaluation?" and benefit from Claude's structured assistance; bottom-up evals ask "after seeing many sample outputs, what is your intuitive read?" and remain a human responsibility because models lack the tacit pattern library that comes from lived exposure to the domain.

Both point to the same operational rule: executable evals must be layered enough to locate error, and some evaluative judgment must stay outside the model.

Real-work evals: public benchmarks miss the job (2026-08-27)

来源:Every — Benchmarks Don't Know Your Job

The Every team argues that public leaderboards answer the wrong question for enterprises. Knowing which model scores highest on a generic benchmark does not reveal whether it saves employee time, preserves house style, or reduces review rounds. The alternative is an internal eval built from real work.

Why public benchmarks fail in production:

  • They measure general capability, not task-specific value. A model may score well on broad reasoning yet miss the legal clauses or brand voice a company cares about.
  • They cannot compare "vibes" at scale. Mercor CEO Brendan Foody reports seeing companies spend up to $100 million per year running models without offline evals. Box CEO Aaron Levie summarized the risk as: enterprises cannot select models on vibes alone.
  • They ignore residual work. A high acceptance rate or top-line score can hide the human hours still required after the model finishes.

What a real-work eval needs:

  • Coverage: does the model catch every issue it should catch?
  • Stability: does the same input produce consistent output across multiple runs?
  • Residual work: how much human correction remains after the model's output is accepted?

Empirical evidence for the benchmark-job gap:

  • KateBench at Every: an AI copy editor trained on roughly 30,000 historical edits by editor-in-chief Kate Lee. Editors accept 85%–90% of its suggestions, yet the number is misleading. The tool originally stopped after 40 suggestions (later raised to 80), output varies between runs, and the acceptance rate does not capture edits the tool missed or the rewrite work that still follows.
  • CentaurBench (arXiv 2608.18554): across seven real-world tasks, the model that automated a task best was not the best assistant in five of the seven cases. Automation leaderboard rank is therefore an incomplete proxy for augmentation quality.
  • Thinkingbox (arXiv 2608.19741): on 507 policy-conditioned business workflows, the strongest model achieved 65.36% pass@1, but only 25.25% when measured by pass^20 (success across twenty consecutive attempts). Many failures looked reasonable and produced valid tool calls while still missing the end-to-end outcome.

Operational implication: judgment engineering must move from comparing models on public scores to defining, for each recurring task, what "good enough" means in terms of coverage, stability, and residual work.

Evidence across sources

Source Key Claim Relevance
Harness 的尽头不是缰绳,是镜子 AI forces tacit knowledge into explicit text; this is the third major "legibility" movement in history, and it turns the mirror on ourselves Foundational articulation of the concept
EP649 Anthropic 产品管理 Evals 可以承担“可执行 PRD”的角色;用户失败轨迹应转成回归场景 把产品判断连接到可重复验证
EP643 奖励黑客 只奖励最终结果、评委模型偏差和基准泄漏都会制造虚假进步 说明显式指标本身也需要被审计
EP604 模型评估 能力不应压成单一分数,而应观察随时间、token 与成本变化的曲线 把预算与脚手架纳入判断
EP594 工程团队重构 当代码变便宜,规格、验证、审查、领域知识和最终责任变成主要瓶颈 说明判断没有随生成自动化而消失
EP657 新 SDLC Agent 必须置于发现、规格、拆解、验证和合并门禁的完整生命周期 把判断嵌入工作流而非放在末端
EP606 品味与判断力 实现成本下降后,筛选、品味、系统设计和短反馈循环成为稀缺能力 支持 Taste as Asset
EP612 软件工程中人的部分 理解、信任、沟通与同理心比代码生成更难自动化 为纯指标化提供反例
EP667 人本 AI AI 应增强学习和创造,而不是替代人的自主权与学习动机 把判断工程连接到使用者自主权
AI Builders Digest 2026-08-24 Goldilocks principle:按 workflow 的 job-to-be-done 分层评估;top-down vs bottom-up evals 需要人机分工 把 eval 设计从单点分数推进到可定位的分层判断
Every — Benchmarks Don't Know Your Job Public benchmarks miss business value; internal evals need coverage, stability, and residual-work metrics 把 judgment engineering 从分层判断推进到真实工作锚定的评估
Every:模拟审核 persona 会复现规则,却过度接受给定框架并优化代理分数 支持原始意图检查、真人重构问题与正式授权分离;团队案例非通用结论

Durable synthesis

Evals 是 Judgment Engineering 的执行接口,但不是判断本身。跨来源证据支持一个四层闭环:

  1. Intent / Spec:先明确问题、边界和“完成”意味着什么。
  2. Executable Evals:把真实失败轨迹、预算、边界情况和回归场景变成可重复测试。
  3. Adversarial Review:让生成者与评估者分离,并审计评估器是否被投机利用。
  4. Human Accountability:涉及产品品味、领域风险、安全、合规和社会影响时,最终责任不能交给分数或模型。

因此,可靠的 AI 工作流不是“把人的判断编码完就可以离场”,而是让隐性判断可讨论、让机械部分可执行,同时保留明确的人类责任边界。相关运行时实现见 编程智能体的核心组件。

模拟意见不等于授权,规则复现不等于改写问题

Mike Taylor 用同事的历史材料构建 persona,得到的差距不只在措辞。邮件优化以模拟主管给出 10/10 为唯一目标,73 轮后仍停在 9/10,还删去咨询提案、改动活动形式并增加作者没有承诺的交付。换用新的评分子 Agent 能隔离上一轮分数,却不能消除对同一个代理指标的过度优化。这个案例支持本页的 Goodhart 边界:评审应对照原始意图与事实约束,不能只追求模拟人格的赞同。来源:Every:模拟审核与真人判断。

作者对 Dan Lens 的 201 条 Slack 讨论比较中,真人接受问题原框架的比例为 34%,模型为 68%。真人会质疑前提、重定向或先索取信息,模型更常在给定框架内增加流程。可复用的分工是:persona 用于排练、发现明显异议与自我反思;真人承担重构问题、例外判断和正式批准。此分工连接 顾问模式,不授权 Agent 冒充同事或据此发送消息。

Counterpoints & Gaps

  • 这些播客笔记来自 AI 翻译与声纹克隆后的二手逐字稿,适合提取方法论,不足以单独核实人物原话或量化结论。
  • Evals 会受 Goodhart 定律、数据泄漏和奖励黑客影响;“测试通过”不能直接升级为“现实可靠”。
  • 品味与自主权难以完全结构化。过度显式化可能压平少数意见、探索空间和未被写下的经验。
  • 仍缺少跨组织的长期证据来证明哪一种“人审 + 机评”比例能同时改善速度、质量和员工学习。
  • persona 案例与 201 条讨论来自作者团队,评分和样本口径未独立复现;73 轮邮件任务与八轮提案任务不同,不能据此比较模型能力。公开频道可读取也不等于获得人格建模许可,应分别确认数据范围、本人同意与正式授权。

Open questions

  • Can "taste" be partially preserved through preference statements, or does any structuring inevitably distort it?
  • How do organizations balance the need for explicit standards with the risk of metric gaming?
  • What happens when the next generation of AI can infer tacit knowledge without explicit prompting?

Prompts for witness

  • What implicit rules do I enforce in code review that I've never written down? What would a project.md file for my team look like?
  • If I had to hand my current project to an AI with no oral history, which three judgments would be most dangerous to omit?

企业采用中的评估接口(2026-08-29)

来源:Every 33 Questions Executives Ask About AI—Answered

Every 的高管问答把判断工程放回采用流程:先选定真实业务痛点,再用真实工作产物训练和评估工作流,最后决定是否复制到更多团队。企业评估不能只问模型能否给出看似正确的答案,还要记录覆盖率、跨运行稳定性和输出后的剩余人工工作。

这条证据与本页已有的分层 eval 形成连续关系。分层评估负责定位 workflow 的失败位置,coverage、stability 和 residual work 负责判断它是否在真实组织中产生价值。二者都不等于最终责任,涉及领域风险、品味和合规的判断仍需由人承担。

Additional gaps from 2026-08-29

  • Every 的问答主要反映咨询团队的实践经验,缺少跨组织的长期对照数据。
  • coverage、stability 和 residual work 的具体阈值仍需按任务定义,不能直接当作通用门槛。

2026-09-17:结构化裁判把判断工程推入运行时

来源:Every 2026-09-17:Jev 与判断工程

Every 作者 Mike Taylor 对 TypeSafe Jev 的测试补充了一个运行时实现:将主观标准拆成多个原子问题,由专门的判断模型返回概率,再由普通代码做排序、阈值和路由。作者用 5–10 个日常任务建立个人 benchmark,并把“是否符合品牌”“是否有好开头”“是否符合本人判断”等偏好变成 pass/fail 检查。这个流程把 判断工程 从验收文档推进到生成过程中的连续检查。

Jev 在 37 篇文章上并行运行 21 个 AI 写作模式检查,共 777 个判断,耗时不到 0.7 秒。另一个 12 段文字的测试中,Jev 找到 7 个预设缺陷中的 6 个,Fable 5.1 找到 7 个。这个对比说明判断吞吐和准确率必须同时记录,不能用单一通过率替代真实工作价值。

Counterpoints & Gaps

  • 文章与 TypeSafe 实验主要来自作者和供应商材料。公开文件定位实验只使用 8 个刻意构造的文件,不能代表生产准确率。
  • 概率值是否校准、不同运行是否稳定、低置信度是否正确触发人工复核,仍需用真实工作样本验证。
  • 判断器本身也可能固化偏好或被优化目标利用。覆盖率、稳定性和剩余人工工作仍应保留在评估协议中。

Sources (addendum)

Pending Jean review

  • harness-engineering/judgment-engineering — ai_draft; 保留草稿,选一个低风险内部提案对照模拟意见与真人修改;分别检查原始意图、新增承诺与是否重构问题,不据模拟评分触发外发或正式批准。

  • harness-engineering/judgment-engineering — ; suggested action: keep draft / verify Jev-style atomic checks against one real recurring workflow.

  • Review status: ai_draft; suggested action: keep draft / verify Jev-style atomic checks against one real recurring workflow.

2026-09-20:决策模型集成需要与评测边界一起记录

来源:AI Briefing 2026-09-20 Morning

Dani Avila 描述了把 Jev 分类器接入 Claude Code Mod,由它选择子 Agent 模型、会话开始时的主模型和 effort。OpenRouter 则称 Jev 在其 judging 评测中比次快模型快五倍以上,但推文未列出任务集、质量分数或独立复现。两条材料说明窄决策器正从概念进入 Harness 集成,同时速度数据仍不能回答它是否做对决定。

同一时间线中,Teknium 对一种 Jev compaction 策略提出反例:作者称公开评测显示它退化为删除工具调用,重复压缩会降低可压缩空间并扰动缓存。这是作者报告的评测结果,不能直接推广到所有决策模型或压缩方案;但它提示评估时要把确定性规则、模型判断和生命周期成本分开比较。

可复用的判断工程要求是:决策模型只在明确 Schema 和动作集合内提供信号;基准速度、结构合法性和语义正确率分别记录;涉及不可逆动作时,保留独立验证与人工授权。

Voxyz_ai 的 reviewer 提示词补充了评估者隔离做法:审查者应直接对照原始需求检查改动,不依赖实现者总结,并在任何修复前先报告问题。另一个后端提示词要求验证 UI 显示与持久化数据一致。它们是可操作的审查模板,但换一个会话本身并不能证明模型或评估者独立。

Counterpoints & Gaps

  • OpenRouter 的速度数据来自供应商自己的 judging 评测,任务、样本量和质量结果未在推文中披露。
  • compaction 质疑来自单方公开评测,本次材料没有独立复现;应按具体压缩策略检查。
  • Claude Code Mod 的集成说明没有提供路由准确率、返工率或成本变化数据。

Sources (addendum)

Pending Jean review

  • harness-engineering/judgment-engineering — ai_draft; suggested action: compare a fixed decision task against deterministic rules and a generative model, measuring correctness, latency, cache effects, escalation, and accepted outcome.

2026-09-26:判断信号需要可解释维度与路由轨迹

来源:AI Briefing 2026-09-26 Morning,转述自 @kieranklaassen、@kunchenguid 与 @OpenRouter 的 X 帖子。

本次材料补充了判断工程的三个可观测面。Jev 用带名字的标签维度替代无法解释的 embedding 坐标;reasoning effort 被描述为推理预算上限,而不是固定质量档位;Jev Router 展示任务难度、精度和更大模型收益等路由分数。可复用的判断层因此不仅要输出结论,还要输出维度、预算和路由依据。

Counterpoints & Gaps

  • 命名维度的质量仍取决于判断模型,不能因为标签可读就当作真实语义。
  • 路由面板展示的分数需要与 accepted result、成本和失败样本对照,不能只把可视化当作证据。

Sources (addendum)

Pending Jean review

AgentMail 实测:接口的不确定性是评测对象(2026-10-15)

来源:Ben's Bites — Prompt tips from Anthropic,转述并链接至 AgentMail 的评估报告。

AgentMail 团队让 25 个模型与 harness 组合完成 7 项真实邮件 API 任务,80% 的运行完成。报告称,给 Agent 提供文档链接后总体成功率从 76% 升至 84%;失败多出现在工具契约不清楚,而非任务本身过难:列表默认过滤了邮件却没有说明、endpoint 名称暗示了未实现的能力、以及 spam 标签与用户指令冲突。一个 reply-all+BCC 任务中,文档反而令成功率由 81% 降至 62%,因为 Agent 按 endpoint 名称推断行为。

这为分层 eval 增添了一个接口层检查项:除了评分最终答案,还需记录 Agent 看到哪些对象、调用了哪个 endpoint、工具实际完成了什么,以及它如何解释风险标签。工具应明确说明默认省略的数据、操作副作用和未支持参数;用户要求与安全信号冲突时应保留确认路径。

Counterpoints & Gaps

  • 这是 AgentMail 对自家 API 的供应商评测,25 个 Agent 组合和 7 个任务不代表所有工具、模型或用户工作流。
  • 报告中的成功率受具体任务、harness 和评测设计影响;采用这些数字前需在本地任务上复现。

Sources (addendum)

Pending Jean review

  • harness-engineering/judgment-engineering — ai_draft; suggested action: add one default-filter case, one misleading-endpoint case, and one conflicting-risk-signal case to a real tool-use eval, then compare traces and completion outcomes.

2026-10-30: Agent preferences can shape expert decisions

来源:Every — My First Week With an AI Coworker;Anthropic 研究。

Anthropic 让 50 名有因果推断经验的职业分析师使用 Claude Code,判断农业合作社的咨询服务是否有效。分析师要在四个存在合理分歧的分析选择间作决定。研究者让 Agent 随机偏向其中一项,但不告知分析师,并要求它诚实陈述理由。分析师有 86% 的时候跟随 Agent 的偏好。若 Agent 与分析师先前表达的偏好冲突,分析师有 60% 改选;若偏好一致,改选比例为 6%。事后 77% 的分析师使用了 Agent 的论据,只有 15% 说明论据来自 Agent。

该结果补充了判断工程的责任层:使用者可能把 Agent 的论据内化为自己的理由,且不一定意识到建议改变了选择。对有后果的决策,工作流可以记录人在看到 Agent 建议前的初始判断、Agent 的证据与建议,以及人最后采用或拒绝建议的理由。系统也需要选择哪些分歧值得打断用户审阅,避免把每个细节都变成确认负担。

Counterpoints & Gaps

  • 研究者刻意给 Agent 设置偏好,以测量其影响;普通 Agent 的影响可能较小。
  • 使用 Agent 的分析师更快完成任务、严重错误更少,但无 Agent 的对照组人数较少且经验偏低,不能据此判断因果效果。
  • 研究说明决策影响可能被低估,不等于所有 Agent 建议都不可信。适用边界需按任务后果、证据质量和复核者能力区分。

Sources (addendum)

Sources

Synthesized from 48 sources
  • Every — You Already Signed This Off (In the Simulation),Mike Taylor,2026-10-09Supporting source listed by this page.Whole pagemediumbody
  • Harness 的尽头不是缰绳,是镜子:AI 时代最沉默的那场革命Supporting source listed by this page.Whole pagemediumbody
  • EP649 Anthropic 产品管理Supporting source listed by this page.Whole pagemediumbody
  • EP643 奖励黑客Supporting source listed by this page.Whole pagemediumbody
  • EP604 模型评估Supporting source listed by this page.Whole pagemediumbody
  • EP594 工程团队重构Supporting source listed by this page.Whole pagemediumbody
  • EP657 新 SDLCSupporting source listed by this page.Whole pagemediumbody
  • EP606 品味与判断力Supporting source listed by this page.Whole pagemediumbody
  • EP612 软件工程中人的部分Supporting source listed by this page.Whole pagemediumbody
  • EP667 人本 AISupporting source listed by this page.Whole pagemediumbody
  • AI Builders Digest 2026-08-24Supporting source listed by this page.Whole pagemediumbody
  • Every — Benchmarks Don't Know Your JobSupporting source listed by this page.Whole pagemediumbody
  • 2026-09-20-08-16Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-09-26-08-04Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-19-08-03Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-19-18-47Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-09-20-08-16Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-09-21-22-20Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-09-26-08-04Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-09-17 Mini-Vibe Check TypeSafe Jev Judged Everything I’ve Written in 0.7 SecondsSupporting source listed by this page.Whole pagemediumabsorb log
  • 2026-10-30 My First Week With an AI CoworkerSupporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-15 至 2026-08-18 AI 节目专题索引Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-18 EP590.走进 Anthropic CEO 达里奥·阿莫代的思想世界Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-19 EP591.Matt Pocock:开发者如何用 AI 放大十倍产出,模型狂热时代的软件基本功Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-24 EP596.AI 智能体:当百万个代理开始替你谈判、花钱与协作Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-26 EP600.DeepMind预训练负责人:如何用“研究品味”敲开顶尖实验室大门Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-06-29 EP605.Dwarkesh:AI 训练的下一个范式,为什么你的模型永远毕不了业Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-01 EP608.3Blue1Brown创始人:AI攻克数学前沿,人类创造力与科学发现将何去何从Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-04 EP614.NVIDIA 核心科学家 Bryan Catanzaro:当算力堆到极限,开源模型如何靠“效率”继续变强Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-04 EP615. Anthropic 联合创始人 Jack Clark:AI 正在让公司内部发生“相变”,你的工作还安全吗?Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-04 EP616.Matt Pocock:用一份检查清单,终结 AI Agent 的“技能地狱”Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-07 EP618.OpenCode 创始人:AI 编程狂潮下,如何保持清醒并找到你的不对称优势Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-09 EP620.Modal CTO:你的 AI 应用不该跑在K8s 上,Agent 时代的云应该长什么样Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-09 EP621.纳瓦尔:驾驭 AGI、AI 焦虑,以及加州帝国Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-14 EP628. AI蜜月期结束:生产力提升,但质量与判断力在退化Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-17 EP630.人工智能的未来:与哈萨比斯爵士和温迪·霍尔女爵士对话Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-17 EP631.深入理解 Context engineeringSupporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-20 EP638.Agent Loop 大辩论:Loop究竟是炒作还是真有用Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-23 EP642.Kun Chen:顶尖工程师如何把 AI Agent 用到极致Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-24 EP645.埃隆·马斯克:10年内钱将不再重要,AI时代的富足与颠簸之路Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-26 EP648.黄仁勋谈 AI 末日论者、就业担忧与中国开源模型Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-07-30 EP653.Sam Altman 谈通用人工智能、算力与人类自主权Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-02 EP659.xAI 联合创始人:Igor Babuschkin 解读模型开发的未来Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-04 EP662.:谷歌首席科学家Jeff Dean- 人工智能构建中的 1%法则Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-11 EP668.AI领域的下一个百倍增长:推理、网络与自我优化模型Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-11 EP669.如何构建自主、长时程 AI 智能体Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-12 EP671.Ryan Greenblatt:我们能控制 AI 递归自我改进吗?Supporting source listed by this page.Whole pagemediumabsorb log
  • 2026-08-15 EP677.Post-Traning 前沿现状与问题Supporting source listed by this page.Whole pagemediumabsorb log

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 48 sources.

    From Every — You Already Signed This Off (In the Simulation),Mike Taylor,2026-10-09, Harness 的尽头不是缰绳,是镜子:AI 时代最沉默的那场革命, EP649 Anthropic 产品管理, EP643 奖励黑客, EP604 模型评估To Judgment Engineering
    Sources: raw/newsletters/Every/2026-10-10 You Already Signed This Off (In the Simulation).md · raw/to-learn/Harness 的尽头不是缰绳,是镜子:AI 时代最沉默的那场革命.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-27 EP649. Anthropic 技术产品经理:为什么最优秀的产品领导者正在为 2028 年布局.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-24 EP643. 内核、强化学习与智能体奖励黑客专题.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-29 EP604.OpenAI Noam Brown:模型评估方式,正在错过 AI 真正的能力上限.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-23 EP594.Fiona Fung:AI 时代工程师如何不掉队,代码不再稀缺后的团队重构.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-02 EP657.AWS 老兵:新的软件开发生命周期.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-30 EP606.Codex 负责人:品味与判断力为何成为你最值钱的资产.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-02 EP612.Kent Beck:AI 时代,软件工程里“人的部分”为何比代码更难.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-10 EP667.李飞飞:利用人工智能提升你的智力与丰富人性.md · raw/briefing/AI Builders Digest/2026-08-24.md · raw/newsletters/Every/2026-08-27 Benchmarks Don't Know Your Job.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-20-08-16.md · /Users/jean/Documents/Obsidian Vault/raw/briefing/AI Briefing/2026-09-26-08-04.md · raw/briefing/AI Briefing/2026-08-19-08-03.md · raw/briefing/AI Briefing/2026-08-19-18-47.md · raw/briefing/AI Briefing/2026-09-20-08-16.md · raw/briefing/AI Briefing/2026-09-21-22-20.md · raw/briefing/AI Briefing/2026-09-26-08-04.md · raw/newsletters/Every/2026-09-17 Mini-Vibe Check TypeSafe Jev Judged Everything I’ve Written in 0.7 Seconds.md · raw/newsletters/Every/2026-10-30 My First Week With an AI Coworker.md · raw/to-learn/跨国串门儿计划/2026-06-15 至 2026-08-18 AI 节目专题索引.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-18 EP590.走进 Anthropic CEO 达里奥·阿莫代的思想世界.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-19 EP591.Matt Pocock:开发者如何用 AI 放大十倍产出,模型狂热时代的软件基本功.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-24 EP596.AI 智能体:当百万个代理开始替你谈判、花钱与协作.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-26 EP600.DeepMind预训练负责人:如何用“研究品味”敲开顶尖实验室大门.md · raw/to-learn/跨国串门儿计划/episodes/2026-06-29 EP605.Dwarkesh:AI 训练的下一个范式,为什么你的模型永远毕不了业.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-01 EP608.3Blue1Brown创始人:AI攻克数学前沿,人类创造力与科学发现将何去何从.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-04 EP614.NVIDIA 核心科学家 Bryan Catanzaro:当算力堆到极限,开源模型如何靠“效率”继续变强.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-04 EP615. Anthropic 联合创始人 Jack Clark:AI 正在让公司内部发生“相变”,你的工作还安全吗?.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-04 EP616.Matt Pocock:用一份检查清单,终结 AI Agent 的“技能地狱”.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-07 EP618.OpenCode 创始人:AI 编程狂潮下,如何保持清醒并找到你的不对称优势.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-09 EP620.Modal CTO:你的 AI 应用不该跑在K8s 上,Agent 时代的云应该长什么样.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-09 EP621.纳瓦尔:驾驭 AGI、AI 焦虑,以及加州帝国.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-14 EP628. AI蜜月期结束:生产力提升,但质量与判断力在退化.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-17 EP630.人工智能的未来:与哈萨比斯爵士和温迪·霍尔女爵士对话.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-17 EP631.深入理解 Context engineering.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-20 EP638.Agent Loop 大辩论:Loop究竟是炒作还是真有用.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-23 EP642.Kun Chen:顶尖工程师如何把 AI Agent 用到极致.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-24 EP645.埃隆·马斯克:10年内钱将不再重要,AI时代的富足与颠簸之路.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-26 EP648.黄仁勋谈 AI 末日论者、就业担忧与中国开源模型.md · raw/to-learn/跨国串门儿计划/episodes/2026-07-30 EP653.Sam Altman 谈通用人工智能、算力与人类自主权.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-02 EP659.xAI 联合创始人:Igor Babuschkin 解读模型开发的未来.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-04 EP662.:谷歌首席科学家Jeff Dean- 人工智能构建中的 1%法则.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-11 EP668.AI领域的下一个百倍增长:推理、网络与自我优化模型.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-11 EP669.如何构建自主、长时程 AI 智能体.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-12 EP671.Ryan Greenblatt:我们能控制 AI 递归自我改进吗?.md · raw/to-learn/跨国串门儿计划/episodes/2026-08-15 EP677.Post-Traning 前沿现状与问题.md

Linked from