--- id: P-REINFORCE-AUTO-PRIN-001 category: "10_Wiki/๐Ÿ’ก Topics/AI" confidence_score: 0.95 tags: [auto-reinforced, principles, decision-making, mental-models, rules, core-values, wisdom] last_reinforced: 2026-04-20 --- # [[Principles|Principles]] ## ๐Ÿ“Œ ํ•œ ์ค„ ํ†ต์ฐฐ (The Karpathy Summary) > "ํ”๋“ค๋ฆฌ์ง€ ์•Š๋Š” ๋ฟŒ๋ฆฌ: ๋งค ์ˆœ๊ฐ„ ๋‹ฅ์ณ์˜ค๋Š” ์ˆ˜์ฒœ ๊ฐ€์ง€ ์„ ํƒ์ง€ ์•ž์—์„œ ์—๋„ˆ์ง€๋ฅผ ๋‚ญ๋น„ํ•˜์ง€ ์•Š๋„๋ก, ์ด๋ฏธ ๊ฒ€์ฆ๋œ ๊ฐ€์น˜์™€ ๋…ผ๋ฆฌ์— ๊ทผ๊ฑฐํ•ด ์„ธ์›Œ๋‘” '๋‚˜๋งŒ์˜ ์ž๋™ ๊ฒฐ์ • ๊ทœ์น™'์ด์ž ๋ณต์žกํ•œ ์„ธ์ƒ์„ ๋‹จ์ˆœํ•˜๊ฒŒ ๋ŒํŒŒํ•˜๋Š” ์ง€์  ๋ฌด๊ธฐ." ## ๐Ÿ“– ๊ตฌ์กฐํ™”๋œ ์ง€์‹ (Synthesized Content) ์›์น™(Principles)์€ ๋ณดํŽธ์ ์œผ๋กœ ์ ์šฉ๋˜๋Š” ๊ทผ๋ณธ์ ์ธ ์ง„๋ฆฌ ๋˜๋Š” ํ–‰๋™ ์ง€์นจ์ž…๋‹ˆ๋‹ค. (๋ ˆ์ด ๋‹ฌ๋ฆฌ์˜ค ๋Œ€์ค‘ํ™”) 1. **์›์น™์˜ ๊ฐ€์น˜**: * **Cognitive Offloading**: ๋งค๋ฒˆ ๊ณ ๋ฏผํ•˜์ง€ ์•Š๊ณ  '์›์น™'์— ๋”ฐ๋ผ ์ฆ‰์‹œ ๊ฒฐ์ •. (Efficiency์™€ ์—ฐ๊ฒฐ) * **Consistency**: ๊ฐ์ •์ด๋‚˜ ์ƒํ™ฉ์— ํ”๋“ค๋ฆฌ์ง€ ์•Š๋Š” ์ผ๊ด€๋œ ๊ฒฐ๊ณผ ๋ณด์žฅ. * **Feedback/Refinement**: ๊ฒฐ๊ณผ๊ฐ€ ๋‚˜์˜๋ฉด ์›์น™์„ ๋ฐ”๊พธ๋ฉด ๋จ (์ง€์†์  ๊ฐœ์„ ). (Feedback-Loops์™€ ์—ฐ๊ฒฐ) 2. **์™œ ์ค‘์š”ํ•œ๊ฐ€?**: * ์›์น™์ด ์—†๋Š” ์ง€๋Šฅ์€ ์ž„๊ธฐ์‘๋ณ€์— ๊ทธ์น˜์ง€๋งŒ, ์›์น™์ด ์žˆ๋Š” ์ง€๋Šฅ์€ '์‹œ์Šคํ…œ'์œผ๋กœ ์ง„ํ™”ํ•˜์—ฌ ๋ณต๋ฆฌ ์„ฑ์žฅ์„ ๋งŒ๋“ค์–ด๋‚ด๊ธฐ ๋•Œ๋ฌธ์ž„. ## โš ๏ธ ๋ชจ์ˆœ ๋ฐ ์—…๋ฐ์ดํŠธ (Contradictions & RL Update) - **๊ณผ๊ฑฐ ๋ฐ์ดํ„ฐ์™€์˜ ์ถฉ๋Œ**: ๊ณผ๊ฑฐ์—๋Š” ์กฐ์ƒ์œผ๋กœ๋ถ€ํ„ฐ ๋ฌผ๋ ค๋ฐ›์€ '๋„๋•๋ฅ  ์ •์ฑ…'์— ๋จธ๋ฌผ๋ €์œผ๋‚˜, ํ˜„๋Œ€ ์ •์ฑ…์€ ์ž์‹ ์˜ ์‹คํ—˜๊ณผ ์‹คํŒจ๋ฅผ ํ†ตํ•ด ์Šค์Šค๋กœ ๊ตฌ์ถ•ํ•˜๋Š” '๊ฐœ์ธ์  ์˜์‚ฌ๊ฒฐ์ • ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ •์ฑ…'์œผ๋กœ ์ง„ํ™”ํ•จ(RL Update). - **์ •์ฑ… ๋ณ€ํ™”(RL Update)**: AI์˜ ์•ˆ์ „ ์ •์ฑ…(Safety Policy) ๋˜ํ•œ ์ธ๊ฐ„์˜ ์›์น™์„ ๋ชจ๋ธ์˜ ํ–‰๋™ ๊ฐ•๋ น ์ •์ฑ…์œผ๋กœ ์ฃผ์ž…ํ•˜๋Š” '์›์น™ ๊ธฐ๋ฐ˜ ์ง€์‹œ(Constitutional AI)' ์ •์ฑ…์ด ๊ฐ€์žฅ ๊ฐ•๋ ฅํ•œ ํ†ต์ œ ์ˆ˜๋‹จ ์ •์ฑ…์ด ๋จ. ## ๐Ÿ”— ์ง€์‹ ์—ฐ๊ฒฐ (Graph) - [[Decision Theory|Decision Theory]], [[Feedback-Loops|Feedback-Loops]], [[Mental-Models|Mental-Models]], [[Judgment|Judgment]], [[Philosophy|Philosophy]] - **Modern Tech/Tools**: Ray Dalio's Principles, Constitutional AI (Anthropic). ---