ํ™ˆ
Seiok ๐Ÿคธโ€โ™‚ Kim
์ทจ์†Œ

์ƒˆ๋กœ์šด ์‹œ์ž‘

๐ŸŒป ๋ธ”๋กœ๊ทธ ์ƒˆ ๋‹จ์žฅ ์™„๋ฃŒ! ์ด์ œ์„œ์•ผ ์ปคํ”ผ ํ•œ์ž”์˜ ์—ฌ์œ ๋ฅผ...โ˜• — ์†Œ๊ฐœ ์ด์ „ ๋ฒ„์ „์˜ ๋ธ”๋กœ๊ทธ๋Š” ์‹คํ—˜์ ์ด๊ณ  ์‹ฌ๋ฏธ์ ์ธ ์‹œ๊ฐ ํšจ๊ณผ์— ์ค‘์ ์„ ๋‘์—ˆ๋‹ค. 2017๋…„ 8์›” ์ฒ˜์Œ ๋ธ”๋กœ๊ทธ ํฌ์ŠคํŠธ๋ฅผ ์‹œ์ž‘์œผ๋กœ, ๊ธ€์„ ์“ฐ๋Š” ๊ณผ์ •์—์„œ ์ƒ๊ฐ๋„ ์ •๋ฆฌ๊ฐ€ ๋˜๊ณ , ์ปค๋ฆฌ์–ด์—๋„ ๋„์›€์ด ๋˜์—ˆ๋‹ค. ๊ทธ๋Ÿฐ ์˜๋ฏธ์—์„œ ์ด์ „ ๋ฒ„์ „์˜ ๋ธ”๋กœ๊ทธ์— ์ƒˆ์‚ผ ๊ฐ์‚ฌํ•˜๋‹ค โ€“ ๋ณต์žกํ•˜๊ณ  ์นด์˜ค์Šค ์ ์ธ ๋‚ด ์‚ถ...

ํ•ด๋งˆ ๊ณต๊ฐ„

ํ•™์Šต์ด ๊ธฐ์–ต์ด๋˜๋Š” ๊ณต๊ฐ„ — ํ•ด๋งˆ(hippocampus) ๋‚ด๋ถ€์—์„œ ์ผ์–ด๋‚˜๋Š” ํ•™์Šต ๊ณผ์ •์„ ์‹œ๊ฐํ™”ํ•œ ๋ฉ‹์ง„ ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒ ๋น„์ฃผ์–ผ๋ผ์ด์ œ์ด์…˜์„ ๋งŒ๋“ค์—ˆ๋‹ค (Sun et al., 2025) . ์ง์ ‘ ์ƒํ˜ธ์ž‘์šฉํ•  ์ˆ˜ ์žˆ๋‹ค! ํด๋ฆญํ•˜๊ณ  ๋“œ๋ž˜๊ทธํ•ด์„œ ํšŒ์ „ํ•ด๋ณด์ž. ์•„๋ฆ„๋‹ต์ง€ ์•Š์€๊ฐ€? ์˜ค๋Š˜ ์„œ์šธ๋Œ€ํ•™๊ต๋ณ‘์› ์‹ฌํฌ์ง€์—„์˜ ํ•ต์‹ฌ ์‹œ๊ฐํ™” ์ž๋ฃŒ ์ค‘ ํ•˜๋‚˜๋กœ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. Natu...

๊ฑฐ๋Œ€ ์˜๋ฃŒ ์–ธ์–ด ๋ชจ๋ธ์—์„œ ์ถ”๋ก ๊ณผ์ • ์ถ”์ ํ•˜๊ธฐ

Project Proposal for AI702, and my pesonal passion. — ์„œ๋ก  ๋Œ€๊ทœ๋ชจ ์–ธ์–ด ๋ชจ๋ธ(LLM)์€ ์˜๋ฃŒ ์งˆ์˜์‘๋‹ต(Medical Question Answering) ๊ณผ์ œ์—์„œ ํšจ๊ณผ์ ์ธ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ณ  ์žˆ์œผ๋ฉฐ, ํ‰๊ท ์ ์ธ ์˜๋ฃŒ์ง„ ์ˆ˜์ค€ ๋˜๋Š” ๊ฒฝ์šฐ์— ๋”ฐ๋ผ ์ „๋ฌธ์˜ ์ˆ˜์ค€์— ํ•„์ ํ•˜๋Š” ๊ฒฐ๊ณผ๋ฅผ ๋‹ฌ์„ฑํ•˜๊ณ  ์žˆ๋‹ค. ๋Œ€ํ‘œ์ ์ธ ์˜ˆ๋กœ๋Š” M...

ํœด๋จธ๋…ธ์ด๋“œ 2025

ํœด๋จธ๋…ธ์ด๋“œ ํ•™ํšŒ ๋ฆฌ๋ทฐ — ์œ„์น˜ ์ง€๋Šฅ(Position Intelligence) ์—์„œ ํž˜ ์ง€๋Šฅ(Force Intelligence) ์œผ๋กœ. [\begin{array}{ccc} \color{goldenrod}{\text{๋ฒ”์šฉ ๋กœ๋ด‡}} & \xrightarrow[\text{throughput}]{} & \color{springgr...

ํŠธ๋žœ์Šคํฌ๋จธ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ์š”์†Œ๋Š”?

What is the the most important unit of the transformer — ์ง๊ด€์ ์ธ ์‹œ๊ฐํ™”) (LLM Visualization โ€” Bbycroft.net, n.d.) ๋ฅผ ์‚ดํŽด๋ณด์ž. LLM์„ ์ธํ„ฐ๋ž™ํ‹ฐ๋ธŒํ•˜๊ฒŒ ์‹œ๊ฐํ™”ํ•œ ์˜ˆ์‹œ. ํŠธ๋žœ์Šคํฌ๋จธ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ ์š”์†Œ๋Š” ๋ฌด์—‡์ผ๊นŒ? ํŠธ๋žœ์Šคํฌ๋จธ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ตฌ์„ฑ...

์Šค์ผ€์ผ ๊ฐ€๋Šฅํ•œ Q-Learning

Reaction to Seohong's Post on X — ํ™•์žฅ ๊ฐ€๋Šฅํ•œ Q-Learning์„ ์ฐพ์•„์„œ ๋‚˜๋Š” ์˜ค๋žซ๋™์•ˆ ๊ฐ•ํ™”ํ•™์Šต(Reinforcement Learning)์„ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋Š” ๋ฐฉ๋ฒ•์„ ์ฐพ์•„์™”๋‹ค. ๊ทธ์ค‘ ํฅ๋ฏธ๋กœ์šด ํ”„๋กœ์ ํŠธ ์ค‘ ํ•˜๋‚˜๊ฐ€ starcraft.ai ํ”„๋กœ์ ํŠธ์ด๋‹ค. ์šด ์ข‹๊ฒŒ๋„ DeepMind๊ฐ€ ์ง„ํ–‰ํ•œ ์›Œํฌ์ˆ์— ์ฐธ์„ํ•  ๊ธฐํšŒ๊ฐ€ ์žˆ์—ˆ๊ณ ...

์—”๋น„๋””์•„ ์  ์Šจ ํ™ฉ์˜ "์ƒ๊ฐํ•˜๋Š” ๊ธฐ๊ณ„"

์  ์Šจํ™ฉ์˜ 33๋…„ ๊ฒฝ์˜์‚ฌ๋ฅผ ๋‹ด์€ ์ฒซ ๊ณต์‹ ์ž์„œ์ „ — ์  ์Šจ ํ™ฉ์€ ์–ด๋–ค ์‚ฌ๋žŒ์ธ๊ฐ€. ์  ์Šจํ™ฉ์˜ ์ž์„œ์ „ โ€œ์ƒ๊ฐํ•˜๋Š” ๊ธฐ๊ณ„(์›์ €: Thinking Machines)โ€ (Witt et al., 2025) ๋ฅผ ์ฝ๋Š” ์ค‘์ด๋‹ค. ์„œ๋ฌธ์— ์ด๋Ÿฐ ๋ง์ด ์žˆ๋‹ค. Practice even what seems impossible. Marcus Aurelius ...

์กฐ๊ฑด๋ถ€ ์ƒ์„ฑ ๋ชจ๋ธ๋ง๋งŒ์œผ๋กœ ์˜์‚ฌ๊ฒฐ์ •์ด ๊ฐ€๋Šฅํ• ๊นŒ?

2022๋…„์— ๊ผญ ๋ด์•ผ ํ•  ๋…ผ๋ฌธ ์ค‘ ํ•˜๋‚˜์ด๋‹ค. ์–ธ์  ๊ฐ€ ๋‚˜์˜ฌ๊ฒƒ์œผ๋กœ ์˜ˆ์ƒ๋œ ๋ฐฉ๋ฒ•์ด๋ž„๊นŒ... ์—ฌ๊ธฐ์„œ๋Š” ์˜์‚ฌ๊ฒฐ์ •์„ ๊ฐ•ํ™” ํ•™์Šต(RL)์ด ์•„๋‹Œ ์กฐ๊ฑด๋ถ€ ์ƒ์„ฑ ๋ชจ๋ธ๋ง์œผ๋กœ ๊ตฌ์„ฑํ•˜์˜€๋‹ค. ๊ฐœ์ธ์ ์œผ๋กœ ๋กœ๋ด‡ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ์˜ ์‹คํ—˜๋„ ์ง„ํ–‰ํ•˜๋ฏ€๋กœ ๋‘ ๋ฐฐ๋กœ ๊ด€์‹ฌ์ด ์žˆ์—ˆ๋‹ค. ์ „ํ†ต์ ์ธ RL ๋ฐฉ๋ฒ•์˜ ๋ณต์žก์„ฑ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ๋Š”์ง€ ๊ถ๊ธˆํ•˜๋‹ค. — [\renewcommand{\...

์ผ๋ฐ˜์ฃผ์˜ ์‹ ๊ฒฝ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ํ•™์Šต์ž

์ •๋ ฌ, ๊ฒ€์ƒ‰, ๋™์  ํ”„๋กœ๊ทธ๋ž˜๋ฐ, ๊ฒฝ๋กœ ์ฐพ๊ธฐ, ๊ธฐํ•˜ํ•™๊ณผ ๊ฐ™์€ ๋‹ค์–‘ํ•œ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‹คํ–‰ํ•˜๋„๋ก ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋Š” ๋‹จ์ผ ๊ทธ๋ž˜ํ”„ ์‹ ๊ฒฝ๋ง ํ”„๋กœ์„ธ์„œ. — [\renewcommand{\V}[1]{\mathbf{#1}}] ์‹ ๊ฒฝ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๋ถ„์•ผ์˜ ๋˜ ๋‹ค๋ฅธ ์ผ๋ฐ˜์ฃผ์˜ ํ•™์Šต์ž (Ibarz et al., 2022) ๊ฐ€ ๋‚˜์™”๋‹ค. Abstract ์‹ ๊ฒฝ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ถ”๋ก ์˜...

๊ฐ€ํ† (GATO) ๋…ผ๋ฌธ์„ ์ฝ์–ด๋ณด์ž!

๐Ÿง  ์•„ํฌ๊ฐ€ํ† ๊ฐ€ ์•„๋‹ˆ๋ผ ๊ฐ€ํ† ๋‹ค. ๋”ฅ๋งˆ์ธ๋“œ์—์„œ ๋‚˜์˜จ generalist agent ๋…ผ๋ฌธ. — ๊ฐ€ํ† . ๋”ฅ๋งˆ์ธ๋“œ์—์„œ ๋‚˜์˜จ generalist AI agent (Reed et al., 2022) ๋กœ, GATO์˜ ์•ฝ์ž๋Š” ์ •ํ™•ํžˆ ๋ชจ๋ฅด๊ฒ ๋‹ค. โ€œGeneralist Agent beyond the realm of Text Outputsโ€ ์ •๋„ ๋˜์ง€ ์•Š์„๊นŒ ์‹ถ...