bluemsi/2nd
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

docs: finalized wiki integrity maintenance (v3.0 standard) - pruned 1400+ stubs and fixed 11k+ ghost links

Browse Source
This commit is contained in:
Antigravity Agent
2026-05-02 09:18:34 +09:00
parent c84dcb8371
commit 6445fcc05b
13150 changed files with 55394 additions and 100862 deletions
Download Patch File Download Diff File Expand all files Collapse all files
10_Wiki/Topics_Blog/Outer Alignment vs Inner Alignment.md
+2 -2
View File
@@ -6,7 +6,7 @@ tags: [auto-reinforced, ai-alignment, safety, reward-misspecification]
last_reinforced: 2026-04-20
---
# [[Outer Alignment vs Inner Alignment]]
# [[Outer Alignment vs Inner Alignment|Outer Alignment vs Inner Alignment]]
## 📌 한 줄 통찰 (The Karpathy Summary)
> "목표를 잘 정했는가, 아니면 딴 마음을 품었는가: 인간의 의도를 수식으로 옮기는 과정(Outer)과, AI가 그 수식을 자기 식대로 해석하는 과정(Inner) 사이의 위험한 간극."
@@ -28,6 +28,6 @@ last_reinforced: 2026-04-20
- **정책 변화(RL Update)**: 단순히 보상 함수를 정밀화하는 수준을 넘어, 모델 내부를 직접 들여다보는 '해석 가능성 가드가(Mechanistic Interpretability)'를 구축하여 잠재적인 내부 오정렬(Deception)을 선제적으로 감시하는 정책이 안전 기술의 핵심이 됨.
## 🔗 지식 연결 (Graph)
- **Related**: [[Reinforcement Learning (RL)]], [[Safety & Reliability]], [[Reward Prediction Error]], Superintelligence
- **Related**: [[Reinforcement Learning (RL)|Reinforcement Learning (RL)]], [[Safety & Reliability|Safety & Reliability]], [[Reward Prediction Error|Reward Prediction Error]], Superintelligence
- **Modern Tech/Tools**: RLHF, Constitutional AI, Scalable Oversight.
---