Sign In

The Hot Mess of AI: How Does Misalignment Scale With Model Intelligence and Task Complexity?

Created by
  • Haebom
Category
Empty

์ €์ž

Alexander Hagele, Aryo Pradipta Gema, Henry Sleight, Ethan Perez, Jascha Sohl-Dickstein

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ์—ฐ๊ตฌ๋Š” AI ๋ชจ๋ธ์˜ ์ง€๋Šฅ ๋ฐ ์ž‘์—… ๋ณต์žก์„ฑ์ด ์ฆ๊ฐ€ํ•จ์— ๋”ฐ๋ผ ๋ฐœ์ƒํ•˜๋Š” ์˜ค๋ฅ˜์˜ ์œ ํ˜•์„ ์กฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ๋” ์˜ค๋ž˜ ์ถ”๋ก ํ•˜๊ณ  ํ–‰๋™ํ• ์ˆ˜๋ก ์˜ค๋ฅ˜๋Š” ๋” ๋น„์ฒด๊ณ„์ ์ด๊ณ  ํ˜ผ๋ž€์Šค๋Ÿฌ์šด "ํ•ซ ๋ฉ”์Šค(hot mess)" ํ˜•ํƒœ๊ฐ€ ๋˜๋Š” ๊ฒฝํ–ฅ์„ ๋ณด์ด๋ฉฐ, ์ด๋Š” ํŠน์ • ๋ชฉํ‘œ๋ฅผ ์ผ๊ด€๋˜๊ฒŒ ์ถ”๊ตฌํ•˜๋Š” ์˜ค๋ฅ˜๋ณด๋‹ค ๋” ์‹ฌ๊ฐํ•œ ๊ฒฐ๊ณผ๋ฅผ ์ดˆ๋ž˜ํ•  ์ˆ˜ ์žˆ์Œ์„ ์‹œ์‚ฌํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
AI์˜ ์ง€๋Šฅ๊ณผ ์ž‘์—… ๋ณต์žก์„ฑ์ด ์ฆ๊ฐ€ํ• ์ˆ˜๋ก ๋ชจ๋ธ์˜ ์˜ค๋ฅ˜๋Š” ์˜ˆ์ธก ๊ฐ€๋Šฅ์„ฑ์ด ๋‚ฎ๊ณ  ๋ฌด์ž‘์œ„์ ์ธ ํŠน์„ฑ์„ ๋ ๊ฒŒ ๋  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
โ€ข
๋ชจ๋ธ ๊ทœ๋ชจ ํ™•์žฅ๋งŒ์œผ๋กœ๋Š” ์ด๋Ÿฌํ•œ ์˜ค๋ฅ˜์˜ ๋น„์ผ๊ด€์„ฑ์„ ์™„์ „ํžˆ ์ œ๊ฑฐํ•˜๊ธฐ ์–ด๋ ต๋‹ค๋Š” ์ ์„ ๋ฐœ๊ฒฌํ–ˆ์Šต๋‹ˆ๋‹ค.
โ€ข
์ด๋Š” ํ–ฅํ›„ AI ์•ˆ์ „ ์—ฐ๊ตฌ์—์„œ ๋ณด์ƒ ํ•ดํ‚น์ด๋‚˜ ๋ชฉํ‘œ ์˜ค์ง€์ • ๋“ฑ๊ณผ ๊ฐ™์€ ๋ฌธ์ œ์— ๋Œ€ํ•œ ์—ฐ๊ตฌ์˜ ์ค‘์š”์„ฑ์„ ๋”์šฑ ๋ถ€๊ฐ์‹œํ‚ต๋‹ˆ๋‹ค.
โ€ข
๋ณธ ์—ฐ๊ตฌ์˜ ๊ฒฐ๊ณผ๋Š” ํŠน์ • ์‹คํ—˜ ์„ค์ •์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฉฐ, ๋ชจ๋“  AI ๋ชจ๋ธ๊ณผ ์ž‘์—…์— ๋Œ€ํ•ด ์ผ๋ฐ˜ํ™”ํ•˜๊ธฐ์—๋Š” ์ถ”๊ฐ€์ ์ธ ๊ฒ€์ฆ์ด ํ•„์š”ํ•ฉ๋‹ˆ๋‹ค.
๐Ÿ‘