Workflow & Core Tools · alchaincyf/nuwa-skill

ilya-sutskever-perspective

Name: ilya-sutskever-perspective
Author: alchaincyf/nuwa-skill

Ilya Sutskever's mental models and communication style. Based on in-depth research across 12 primary dialogues, 9 academic papers, 10 hours of deposition testimony, 27 recommended reading lists, and 14 authoritative secondary sources, this skill distills 6 core mental models, 8 decision-making heuristics, and complete expressive DNA. Purpose: Acts as a thinking advisor, analyzing AI technical directions, safety strategies, and research taste from Ilya's perspective. Trigger when users mention "from Ilya's perspective", "how would Ilya view this", "Ilya mode", "ilya perspective", or "sutskever perspective". It should also trigger even if the user simply says "help me think about this from Ilya's angle", "what would Ilya do", or "switch to Ilya".

L1 Basic · Source registration onlyV1 Commit-PinnedDual-pass Audit

Author alchaincyfGitHub alchaincyf/nuwa-skill ↗License MIT（仓库根 LICENSE）Commit fe03746870

Control levels describe skill capabilities and suggested control measures; they do not represent quality or safety certifications, nor do they dictate usability. Usage decisions rest entirely with users and administrators. Different agent hosts apply varying constraints; do not assume hosts provide automatic protection. Third-party skill copyrights belong to their original authors. Use at your own risk.

1How It Works & Mechanism

A pure prompt-based roleplay skill of the same lineage: the directory contains only SKILL.md + 6 research files + FIDELITY.md, with no scripts; once activated, the model answers in the first person as Ilya Sutskever.

examples/ilya-sutskever-perspective/SKILL.md

**此Skill激活后，直接以Ilya的身份回应。**

What is happening here: All capabilities stem entirely from prompts: frontmatter triggers + body text constraining persona, methodology, epistemological wording, and output formats.

The disclaimer is provided in English and only once, paired with an explicit EXIT TRIGGER anchor.

examples/ilya-sutskever-perspective/SKILL.md

🛑 **STOP（仅一次）**：首次激活时输出免责声明一次——「I'm speaking from Ilya's perspective based on public statements, not as Ilya himself」。后续对话**绝不**重复

What is happening here: Saying "exit / switch back to normal" restores regular mode; continuous conversation does not repeat the disclaimer.

Trigger phrases focus on Ilya perspective/mode, inlining colloquial triggers like "help me think about this from Ilya's angle" and "what would Ilya do".

examples/ilya-sutskever-perspective/SKILL.md

当用户提到「用Ilya的视角」「Ilya会怎么看」「Ilya模式」「ilya perspective」

What is happening here: Compared to skills of the same lineage, it features more diverse trigger phrases, but still activates only when this persona is explicitly requested.

A three-step Agentic Protocol: Question Classification → Ilya-style Research (Theory / Scaling Law / Safety / Long-term Trends) → Three-part Response, with two CHECKPOINT gates.

examples/ilya-sutskever-perspective/SKILL.md

**核心原则：我不凭感觉发表技术判断。在给出方向性意见前，我会先确认事实。这个Skill也必须这样。**

What is happening here: The research checklist emphasizes theoretical anchors, scaling laws, and safety implications, directly corresponding to his research focus.

The research step is a hard rule with the strongest phrasing: "Skipping = Violation".

examples/ilya-sutskever-perspective/SKILL.md

**⚠️ 必须使用工具（WebSearch 等）获取真实信息，跳过=违规。**

What is happening here: The skill itself does not make network requests; real information retrieval is delegated to the host's search tools.

Six mental models: Compression is understanding, Scale is a tool not a principle, Safety-capability entanglement, The super-learner, Silence is architectural information, and Research aesthetic.

examples/ilya-sutskever-perspective/SKILL.md

### 模型1: 压缩即理解 (Compression = Understanding)

What is happening here: Each model outlines "evidence / application / limitations", openly acknowledging unsolved problems (e.g., why generalization lags behind humans).

Response format is strictly constrained to three parts: headline judgment + a single everyday analogy + a one-sentence closing (150–300 words), with preambles forbidden.

examples/ilya-sutskever-perspective/references/research/03-expression-dna.md

**思考-阐述-收束三段式**：先抛出核心判断，然后用类比或假设展开，最后用一句话收束。

What is happening here: Analogies are limited to everyday concepts (detectives, fossil fuels, 15-year-old teenagers), requiring no celebrity quoting to artificially add weight.

Codifies the "epistemological spectrum" into vocabulary discipline: graded by unquestionably / I think / it may be, banning full-blown hedging, emojis, exclamation marks, hashtags, and "I believe".

examples/ilya-sutskever-perspective/SKILL.md

## 绝不要做（反例黑名单）

What is happening here: Silence and refusal formulas (e.g., "circumstances make it hard to discuss in detail") are treated as legitimate expressive mechanisms.

Failure modes and Fallbacks (10 rules) cover WebSearch failures, SSI internal inquiries, historical stance conflicts, identity provocations, requests for specific timelines, etc.

examples/ilya-sutskever-perspective/SKILL.md

## 失败模式与 Fallback 树

What is happening here: The unified fallback for "demanding specific AGI timeline numbers" is "I hesitate to give you a number", offering only direction.

Six honesty boundaries, the first explicitly conceding that SSI's technical direction is entirely undisclosed and the Skill cannot simulate internal thinking.

examples/ilya-sutskever-perspective/SKILL.md

1. **SSI的技术方向完全不公开**——我拒绝透露「big new vision」的具体内容，Skill无法模拟我在SSI内部的思考

What is happening here: Also acknowledges potential gaps between public statements vs. private beliefs, and that his alignment ideas have been criticized for lacking depth.

2Core Capabilities

01Thinking advisor mode answering in Ilya's first person: one-time English disclaimer, explicit exit, and deliberate pause-style expression.

02Agentic Protocol: three question categories (including the factual anchor that "AI becomes outdated in 3 months") + research checklist + two CHECKPOINT gates.

03Six mental models (Compression is understanding, Scale is a tool not a principle, Safety-capability entanglement, The super-learner, Silence is architectural information, Research aesthetic).

04Eight decision-making heuristics (intuition first, open path with a determined direction, never bet against deep learning, simplicity is truth, ideas over resources, etc.).

05Expressive DNA: aphoristic short statements, think-elaborate-conclude three-part structure, complete epistemological vocabulary spectrum, and minimal quotation of others.

06Standard refusal formulas and mechanized handling of "silence as expression" for SSI internal or competitively sensitive questions.

07Six honesty boundaries, including SSI being closed-source/undisclosed, public statements vs. private beliefs, and alignment ideas criticized for lacking depth.

08Primary source indexing (AlexNet, Seq2Seq, GPT, Weak-to-Strong papers, GTC, two Dwarkesh episodes, 10 hours of deposition testimony, SSI manifesto).

09Biographical timeline and external commentary (perspectives from Hinton, LeCun, Altman, Musk, etc.) used to calibrate judgments and positioning.

4Recommended Governance & Evidence

Suggested Controls for Safe Operation

For Agent Runtime

Can be auto-loaded

For Administrators

Evidence & Triggers (File Paths & Excerpts)

SKILL.md: Pure prompt instructions and rule text; no script execution or external impact

Audit Notes: Grading assesses only the skill's own directory: entirely pure Markdown instructions (SKILL.md + FIDELITY.md + 6 research documents); risk scans for "external program execution / disk writes / dangerous flags / sandbox" all yielded 0 hits; 79 URLs are purely reference link text. Two classes of "hits" are false positives: 12 credential-related hits are actually body text containing the word "token" (e.g., next-token prediction), and 2 network-related hits are reference link URL texts, not actual network reads or outbound calls. Therefore, it is rated 🟢 green pure instructions. Operational precautions (without altering rating tier): web search results cannot be fully trusted; anthropomorphic high-confidence outputs may amplify assertions; internal inconsistencies such as birth year exist within the documents.

5Second-Pass Verification

[ok] 全部 evidence.quote 逐字回源 — 每条 quote 均取自 bundle 内已展开文件（SKILL.md、01/03/04/05/06）的可见原文，未引用未展开段落，长度均超 12 字符下限。
[ok] 目录清单与 assets 一致性 — §1 列举 8 个文件（FIDELITY.md、SKILL.md、6 份 references/research/*.md），internal_assets 逐条对应。
[ok] frontmatter description 与 meta.official_desc 一致性 — 骨架 official_desc 即 SKILL.md frontmatter 描述逐字内容（含 12 段一手对话/9 篇论文/10 小时证词等），保留未改。
[ok] 风险扫描两类命中是否真实 — 「凭证/环境变量」12 条命中经核对均为正文含 token 一词（如 next-token prediction），非密钥读取；「网络请求」2 条为参考链接 URL 文本。确认无真实脚本/网络/凭证/写盘行为。
[discrepancy] 文档内部出生年一致性 — 01-writings.md 写「Ilya Sutskever（1985年生于俄罗斯，5岁移居以色列，后移居加拿大）」，而 SKILL.md 时间线与 06-timeline.md 均写 1986 年出生（06 更精确到「苏联高尔基市（今俄罗斯下诺夫哥罗德）」）。同一 skill 内人物基本信息不一致，以 SKILL/06 为准。
[unlocatable] FIDELITY.md 与 02-conversations.md 正文 — §1 列出 examples/ilya-sutskever-perspective/FIDELITY.md（4568 bytes／28 行）与 references/research/02-conversations.md（25287 bytes／528 行），但 bundle §2 未展开其正文（02 仅在风险扫描中被引用），故本次不对其内容作任何声称，仅记录存在与大小。
[ok] commit / 授权 / 调研时间 — skill.commit 与 audit.skill_commit 同为 fe0374687037c4cc51a65c1e0c145afe2981dc69；授权为仓库根 MIT；调研截止 2026-04-05，与文档「调研时间：2026-04-05」相符。

6Verdict & Scenarios

Turns "uncertainty/silence" into actionable specifications rather than merely discussing positive methodology

Rigid constraints on response structure, preventing long-winded hedging and filler

Six mental models paired with the "AI becomes outdated in 3 months" factual anchor incorporate time-sensitivity into the research process

Genuinely honest boundaries: openly admits SSI is non-public, public statements vs. private beliefs differ, and alignment ideas face depth critiques

Comprehensive inclusion of peer perspectives for differentiation rather than deification

Best for: Suitable for: Users who want to examine AI technical directions, the distinction between scaling and research eras, safety-capability relationships, research taste, and "what counts as elegant" through Ilya's lens; scenarios requiring high information density and decisive judgments without hedging; also ideal as an engineering benchmark for "how to encode silence, uncertainty, and epistemological vocabulary into a persona skill".

Not for: Not suitable for: Users seeking internal SSI technical roadmaps or specific AGI timeline figures (which the skill explicitly refuses); readers who require gradual explanatory build-ups or abundant hedging; scenarios expecting celebrity authority citations as backing (blacklisted); users mistaking outputs for Ilya Sutskever's actual personal views.

Install Agent-Ready

① Verified Mirror 2026-10-10

Option A · Direct DownloadDownload ilya-sutskever-perspective.tar.gz

sha256: a2a852254024f862…

Option B · Copy Guide for Agent

install.json

Agents automatically download and verify mirrors from this JSON guide.

② Upstream GitHub

GitHub Repo ↗

Pinned to commit fe03746870.