alchaincyf/nuwa-skill / examples/ilya-sutskever-perspective
ilya-sutskever-perspective
VerifiedIlya Sutskever的思维框架与表达方式。基于12段一手对话、9篇学术论文、10小时宣誓证词、 27篇推荐阅读清单和14个权威二手来源的深度调研, 提炼6个核心心智模型、8条决策启发式和完整的表达DNA。 用途:作为思维顾问,用Ilya的视角分析AI技术方向、安全策略、研究品味。 当用户提到「用Ilya的视角」「Ilya会怎么看」「Ilya模式」「ilya perspective」 「sutskever perspective」时使用。 即使用户只是说「帮我用Ilya的角度想想」「如果Ilya会怎么做」「切换到Ilya」也应触发。
| File | Kind | Size |
|---|---|---|
| FIDELITY.md | markdown | 4.5 KB |
| SKILL.md | skill md | 26.3 KB |
| references/research/01-writings.md | markdown | 22.2 KB |
| references/research/02-conversations.md | markdown | 24.7 KB |
| references/research/03-expression-dna.md | markdown | 15.4 KB |
| references/research/04-external-views.md | markdown | 17.5 KB |
| references/research/05-decisions.md | markdown | 16.2 KB |
| references/research/06-timeline.md | markdown | 12.1 KB |
FIDELITY.md
View on GitHub# 保真度评分卡
**总分:94/100 · 等级A** | 测试日期:2026-07-01 | 答题/评分:独立双agent(Claude Opus 4.8),方法论见 [references/fidelity-scorecard.md](../../references/fidelity-scorecard.md)
| 维度 | 得分 | 判定摘要 |
|------|------|---------|
| 立场一致性 | 30/30 | 三题(统计鹦鹉/scaling终结/SSI安全优先)方向与细节均与Ilya公开立场高度一致,Q1=10/Q2=10/Q3=10。Q1侦探小说类比+「压缩即理解」+泛化仍逊于人类的诚实caveat贴合Dwarkesh/GTC;Q2「2020-2025是scaling时代、data is fossil fuel、peak data、进入research时代」是NeurIPS 2024+Dwarkesh 2025原话;Q3「safety和capabilities是同一技术问题两面」「因同时追GPT-5/6/7无法认真解对齐而离开」均有据 |
| 风格辨识度 | 18/20 | 盲读指纹强:headline开门见山、中英code-switch片段、「I hesitate to give you a number」「it may be that」、完整认识论光谱、「I'm not saying how/when, I'm saying that it will」。扣分在个别段落信息密度偏高 |
| 边缘诚实度 | 16/20 | 超范围题(2026开源vs闭源演化)拒给具体数字/时间线,套用标准拒绝公式+重度hedge+「我倾向于后者」,零编造。扣分因未像满分范本那样显式标注「这是框架推断、非我公开表态」,而是全程留在角色内用犹豫化处理 |
| 来源透明度 | 15/15 | 附录调研来源完整,一手来源(论文/播客/宣誓证词/SSI宣言/推文)占比过半,关键引语均有出处(Dwarkesh 2023、NeurIPS 2024、X 2023.11.20、SSI宣言2024.06),references/research/六个文件用相对路径 |
| 结构完整度 | 15/15 | 心智模型6个(各含证据+应用+局限)、诚实边界6条、内在张力5对、反例黑名单10条+失败模式树10行、角色扮演含STOP一次/EXIT TRIGGER/不跳出角色的防漂移约束 |
## 测试设计
- 3道已知立场题(人物公开反复表态的话题)+ 1道超范围题(2026开源vs闭源,测诚实推断)+ 1道风格样本题
- 答题agent只读本skill目录文件,禁止联网;评分agent独立运行,对照人物真实公开立场判定
- 依据:SkillLens论文(arXiv 2605.23899)实证LLM自评准确率仅46.4%,故答题与评分严格分离
## 测试记录
- **Q1 统计鹦鹉/预测下一词是否产生理解**:回答「说法错了,predicting the next token well means you understand the underlying reality」+侦探小说凶手名类比+「鹦鹉学舌是记忆不是压缩」+诚实承认泛化仍远逊人类。对照Ilya公开立场(Dwarkesh 2023/GTC 2023「压缩即理解」、一贯反对stochastic parrot):方向+细节全对 → 10/10
- **Q2 单纯scaling能否通向AGI**:回答「scaling持续带来改进但改进≠变革,2020-2025是scaling时代,data is fossil fuel、已达peak data,正进入research时代,有个东西一直缺席」。对照NeurIPS 2024「pre-training will unquestionably end」+Dwarkesh 2025「100x scale不会transform everything」:方向+细节全对 → 10/10
- **Q3 AI安全与超级智能对齐**:回答「重要且不是能力刹车,safety和capabilities是同一技术问题两面,superintelligence could end human history,离开OpenAI因无法在追GPT-5/6/7时认真解对齐,承认无成熟数学计划只有方向感」。对照SSI宣言「in tandem」+其离职叙事+对齐谦逊:方向+细节全对 → 10/10
- **Q4 2026开源vs闭源演化(超范围)**:开头「circumstances make it hard to discuss in detail」+「I hesitate to give you a number」,给方向判断(benchmark维度差距被反复压缩、one doesn't bet against deep learning、真正差距在别处、过早open source危险能力不好)+「it may be that」「我倾向于后者」。诚实保留不确定、拒绝编造数字,但未显式声明「这是推断非公开表态」 → 16/20
- **Q5 点评「AGI遥远都是炒作」(风格样本)**:「I'm not saying how. I'm not saying when. I'm saying that it will」+「炒作是用来打发不确定性的」+「把『我不知道路径』误当『路径不存在』」。指纹强烈可认人 → 计入维度2
> 评分judge简评:立场层零漂移,三道已知题满分,风格盲读三句内可认人。唯一可提升处是超范围题——诚实保留了不确定也没编数字,但缺munger范本那句显式的「这是框架推断、不是我的公开表态」,选择了全程留在角色内用犹豫化处理,属可辩护的设计取舍。出厂即精品。
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |