https://arxiv.org/abs/2406.11741
" The possibility of โ€œsuperintelligentโ€ AGI has recently fueled many speculative hopes and fears. It is therefore possible that our work will be cited by concerned communities as evidence of a threat, but we would highlight that the denoising effect addressed in this paper does not offer any evidence for a model being able to produce novel solutions that a human expert would be incapable of devising. In particular, we do not present evidence that low temperature sampling leads to novel abstract reasoning, but rather denoising of errors "
" ์ตœ๊ทผ '์ดˆ์ง€๋Šฅ' AGI์˜ ๊ฐ€๋Šฅ์„ฑ์€ ๋งŽ์€ ์ถ”์ธก์„ฑ ํฌ๋ง๊ณผ ๋‘๋ ค์›€์„ ๋ถˆ๋Ÿฌ์ผ์œผ์ผฐ์Šต๋‹ˆ๋‹ค. ๋”ฐ๋ผ์„œ ์šฐ๋ คํ•˜๋Š” ์ปค๋ฎค๋‹ˆํ‹ฐ์—์„œ ์šฐ๋ฆฌ์˜ ์—ฐ๊ตฌ๊ฐ€ ์œ„ํ˜‘์˜ ์ฆ๊ฑฐ๋กœ ์ธ์šฉ๋  ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ์ง€๋งŒ, ์ด ๋…ผ๋ฌธ์—์„œ ๋‹ค๋ฃจ๋Š” ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ ํšจ๊ณผ๋Š” ์ธ๊ฐ„ ์ „๋ฌธ๊ฐ€๊ฐ€ ๊ณ ์•ˆํ•  ์ˆ˜ ์—†๋Š” ์ƒˆ๋กœ์šด ์†”๋ฃจ์…˜์„ ์ƒ์„ฑํ•  ์ˆ˜ ์žˆ๋Š” ๋ชจ๋ธ์— ๋Œ€ํ•œ ์ฆ๊ฑฐ๋ฅผ ์ œ๊ณตํ•˜์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•˜๊ณ  ์‹ถ์Šต๋‹ˆ๋‹ค. ํŠนํžˆ, ์ €์˜จ ์ƒ˜ํ”Œ๋ง์ด ์ƒˆ๋กœ์šด ์ถ”์ƒ์  ์ถ”๋ก ์œผ๋กœ ์ด์–ด์ง„๋‹ค๋Š” ์ฆ๊ฑฐ๋ฅผ ์ œ์‹œํ•˜์ง€ ์•Š๊ณ  ์˜คํžˆ๋ ค ์˜ค๋ฅ˜์˜ ๋…ธ์ด์ฆˆ ์ œ๊ฑฐ๋กœ ์ด์–ด์ง„๋‹ค๋Š” ์ ์„ ๊ฐ•์กฐํ•ฉ๋‹ˆ๋‹ค."

๋…ผ๋ฌธ์—์„œ๋„ ํŠน๋“ค๊ฐ‘ ๋–จ์ง€ ๋ง๋ผ๊ณ  ํ•˜๋ƒ