Continuous Learning_Startup & Investment
2.43K subscribers
513 photos
5 videos
16 files
2.74K links
We journey together through the captivating realms of entrepreneurship, investment, life, and technology. This is my chronicle of exploration, where I capture and share the lessons that shape our world. Join us and let's never stop learning!
Download Telegram
์ƒˆ๋กœ์šด ๋ธ”๋กœ๊ทธ ๊ฒŒ์‹œ๋ฌผ! ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ AI x ๊ฒŒ์ž„์˜ ๊ฐ€์žฅ ํฐ ๊ธฐํšŒ๋Š” ์ƒˆ๋กœ์šด ์นดํ…Œ๊ณ ๋ฆฌ์˜ ํ”Œ๋ ˆ์ด๋ฅผ ์—ฌ๋Š” ๊ฒƒ์ž…๋‹ˆ๋‹ค ๏ฟฝ๏ฟฝ. ์œ ๋‹ˆํ‹ฐ๋Š” AI ํผ์ŠคํŠธ ๊ฒŒ์ž„์˜ ์ดˆ๊ธฐ ๋ชจ์Šต์„ ์—ฟ๋ณด๊ณ  ์žˆ์œผ๋ฉฐ ๊ทธ ์ž ์žฌ๋ ฅ์€ ์—„์ฒญ๋‚ฉ๋‹ˆ๋‹ค. ์œ ๋‹ˆํ‹ฐ๊ฐ€ ๊ธฐ๋Œ€ํ•˜๋Š” ๋ช‡ ๊ฐ€์ง€ ๋ถ„์•ผ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค ๏ฟฝ๏ฟฝ.

1/ ์ œ๋„ˆ๋ ˆ์ดํ‹ฐ๋ธŒ ์—์ด์ „ํŠธ - ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ฒŒ์ž„ ๋‚ด ์—์ด์ „ํŠธ์— LLM์„ ์ ์šฉํ•˜์—ฌ ๋†€๋ž๋„๋ก ์‹ค์ œ์™€ ๊ฐ™์€ ํ–‰๋™์„ ๊ตฌํ˜„ํ•ฉ๋‹ˆ๋‹ค. ์†Œ์…œ ์›น์—์„œ ํ•™์Šต๋œ LLM์„ ํ†ตํ•ด ๊ฒŒ์ž„ ๋””์ž์ด๋„ˆ์˜ ์ƒ์ƒ๋ ฅ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ ์ธ๊ฐ„ ์‚ฌํšŒ์˜ ์˜ˆ์ธก ๋ถˆ๊ฐ€๋Šฅ์„ฑ์„ ๋ฐ˜์˜ํ•˜๋Š” ์ƒˆ๋กœ์šด ์ข…๋ฅ˜์˜ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ฒŒ์ž„์ด ๋“ฑ์žฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. 24์‹œ๊ฐ„ ๋‚ด๋‚ด ๋ฉ€ํ‹ฐ ์—์ด์ „ํŠธ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์„ ๋ณด๋Š” ๊ฒƒ์€ ์ฐจ์„ธ๋Œ€ ํŠธ๋ฃจ๋จผ ์‡ผ๊ฐ€ ๋  ์ˆ˜ ์žˆ์œผ๋ฉฐ, ์˜ˆ์ƒ์น˜ ๋ชปํ•œ ๋ฐฉ์‹์œผ๋กœ ๋์—†์ด ์žฌ๋ฏธ์žˆ๊ฒŒ ์ฆ๊ธธ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

2/ ๊ฐœ์ธํ™” - ๋ชจ๋“  ํ”Œ๋ ˆ์ด์–ด์—๊ฒŒ ๊ณ ์œ ํ•œ ๊ฒŒ์ž„ํ”Œ๋ ˆ์ด ๊ฒฝํ—˜์„ ์ œ๊ณตํ•˜์—ฌ ๋ชฐ์ž…๋„์™€ ์žฌํ”Œ๋ ˆ์ด ๊ฐ€๋Šฅ์„ฑ์„ ๊ทน๋Œ€ํ™”ํ•ฉ๋‹ˆ๋‹ค. ์‚ฌ๋ก€ ์—ฐ๊ตฌ: GTA 5๋Š” ํ˜„์žฌ ๊ณต์‹ ๊ฒŒ์ž„๋ณด๋‹ค ๊ฐœ์ธํ™”๋œ ๋กคํ”Œ๋ ˆ์ž‰ ์„œ๋ฒ„์—์„œ ๋” ๋งŽ์€ ํ”Œ๋ ˆ์ด์–ด๋ฅผ ๋งŒ๋‚˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

3/ AI ์Šคํ† ๋ฆฌํ…”๋ง - ๋ฌดํ•œํ•œ ์ธ๋‚ด์‹ฌ์„ ๊ฐ€์ง„ AI ์Šคํ† ๋ฆฌํ…”๋Ÿฌ์˜ ์•ˆ๋‚ด์— ๋”ฐ๋ผ ์ข‹์•„ํ•˜๋Š” IP ์„ธ๊ณ„์—์„œ ์›ํ•˜๋Š” ๋งŒํผ ์‹œ๊ฐ„์„ ๋ณด๋‚ผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

4/ ๋‹ค์ด๋‚ด๋ฏน ์›”๋“œ - ํ”Œ๋ ˆ์ด์–ด๊ฐ€ ๊ฒŒ์ž„์„ ์ง„ํ–‰ํ•˜๋ฉด์„œ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋ ˆ๋ฒจ๊ณผ ์ฝ˜ํ…์ธ ๋ฅผ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. ์Šคํ† ๋ฆฌ์™€ ์›”๋“œ๋ฅผ ๋‹ด๋‹นํ•˜๋Š” AI ๋””๋ ‰ํ„ฐ๊ฐ€ ์žˆ์œผ๋ฉด ๋งค๋ฒˆ ํ”Œ๋ ˆ์ดํ•  ๋•Œ๋งˆ๋‹ค ๋ฐ”๋€Œ๋Š” โ€œ๋„ค๋ฒ„์—”๋”ฉโ€ ๊ฒŒ์ž„์„ ๋งŒ๋“ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์žฅ๊ธฐ์ ์œผ๋กœ๋Š” ๊ฒŒ์ž„์ด ๋” ์ด์ƒ ๋ Œ๋”๋ง๋˜์ง€ ์•Š๊ณ  ์‹ ๊ฒฝ๋ง์„ ์‚ฌ์šฉํ•˜์—ฌ ๋Ÿฐํƒ€์ž„์— ์ƒ์„ฑ๋˜๋Š” ๋ฏธ๋ž˜๋ฅผ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์–ธ์  ๊ฐ€ ๋„ทํ”Œ๋ฆญ์Šค ์˜ํ™”์—์„œ โ€œ์ƒํ˜ธ์ž‘์šฉโ€œ์„ ํด๋ฆญํ•˜๊ณ  ๋ชจ๋“  ์žฅ๋ฉด์ด ์ƒ์„ฑ๋˜๊ณ  ๊ฐœ์ธํ™”๋œ ๊ฐ€์ƒ ์„ธ๊ณ„๋กœ ๋“ค์–ด๊ฐˆ ์ˆ˜ ์žˆ๊ฒŒ ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค.

5/ AI ๋ถ€์กฐ์ข…์‚ฌ - ๋ชจ๋“  ๊ฒŒ์ž„์— AI ๋™๋ฐ˜์ž๋ฅผ ํ†ตํ•ฉํ•˜์—ฌ ์›ํ•˜๋Š” ๋Œ€๋กœ ์ฝ”์น˜ํ•˜๊ณ  ํ•จ๊ป˜ ํ”Œ๋ ˆ์ดํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. UGC ์ œ์ž‘ ๋„์šฐ๋ฏธ๋ถ€ํ„ฐ ํ˜‘๋™ ํŒŒํŠธ๋„ˆ, e์Šคํฌ์ธ  ์ฝ”์น˜์— ์ด๋ฅด๊ธฐ๊นŒ์ง€ ๋ชจ๋“  ๊ฒŒ์ž„์— AI ๋ถ€์กฐ์ข…์‚ฌ๊ฐ€ ์žˆ๋Š” ๋ฏธ๋ž˜๋Š” โ€œํ˜ผ์ž์„œ๋„ ์ข‹๊ณ , AI์™€ ํ•จ๊ป˜ํ•˜๋ฉด ๋” ์ข‹๊ณ , ์นœ๊ตฌ์™€ ํ•จ๊ป˜ํ•˜๋ฉด ๋” ์ข‹๋‹คโ€œ๋Š” ๋ชจํ† ์— ๋”ฐ๋ผ ์ด๋ฃจ์–ด์งˆ ๊ฒƒ์œผ๋กœ ๋ณด์ž…๋‹ˆ๋‹ค.

์•„์ง ์ดˆ๊ธฐ ๋‹จ๊ณ„์ด์ง€๋งŒ ์œ ๋‹ˆํ‹ฐ๋Š” ๋งŽ์€ ์žฌ๋Šฅ ์žˆ๋Š” AI x ๊ฒŒ์ž„ ํŒ€์ด ๋งŒ๋“ค์–ด์ง€๊ณ  ์žˆ๋‹ค๋Š” ์‚ฌ์‹ค์— ํฐ ๊ธฐ๋Œ€๋ฅผ ๊ฑธ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ์‚ฌ๋ก€๋Š” ์œ ๋‹ˆํ‹ฐ ๋ธ”๋กœ๊ทธ์—์„œ ํ™•์ธํ•ด๋ณด์„ธ์š”!
https://a16z.com/2023/07/19/the-neverending-game-how-ai-will-create-a-new-category-of-games/
For years, the tech media had characterized John and Patrick, 34, as boyish business prodigies who monkishly shared an apartment, rode rented e-bikes to work, and preferred reading economic histories and science fiction to ostentatious displays of their growing wealth. John, a student of capitalism who encourages employees to read biographies about moguls like Larry Ellison and John Malone, has lately come into his own as a brass-knuckled manager who has assumed critical leadership roles as his company hurtles toward its next iteration. At the same time, he has been adding a new chapter to his own billionaireโ€™s journey, with a much-loved phrase serving as an epigraph: โ€œThe world is a museum of passion projects.โ€
LLM ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜ ์•„ํ‚คํ…์ฒ˜๋ž€? (RAG์˜ ์ดํ•ด์™€ ๊ธฐ์ˆ  ์Šคํƒ ํƒ์ƒ‰)

- ๋น„์ฆˆ๋‹ˆ์Šค ๋„๋ฉ”์ธ ๊ธฐ๋ฐ˜์œผ๋กœ ๋Œ€ํ˜• ์–ธ์–ด ๋ชจ๋ธ (LLM) ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์„ ๊ฐœ๋ฐœํ•˜๊ณ ์ž ํ•  ๋•Œ (์˜ˆ๋ฅผ ๋“ค์–ด, ๊ธˆ์œต์‚ฌ ๊ณ ๊ฐ ์‘๋Œ€ ์ฑ—๋ด‡) ๋ฌธ์ œ๊ฐ€ ๋˜๋Š” ์ ์€ LLM์˜ ์ •๋ณด ๋ถ€์กฑ๊ณผ ์ด๋กœ ์ธํ•œ ์ œํ•œ๋œ ๋‹ต๋ณ€ ๋Šฅ๋ ฅ์ด๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, GPT 3.5๋Š” 2021๋…„ 9์›” ์ดํ›„์˜ ๋ฐ์ดํ„ฐ๊ฐ€ ์—†์œผ๋ฏ€๋กœ ์ตœ์‹  ๋‰ด์Šค์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ์ œ๊ณตํ•  ์ˆ˜ ์—†๋‹ค.

- ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•œ ๋ฐฉ์•ˆ์œผ๋กœ LLM์„ ์ƒˆ๋กœ์šด ๋ฐ์ดํ„ฐ๋กœ ํŒŒ์ธํŠœ๋‹ํ•˜๋Š” ๊ฒƒ์ด ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ด๋Š” ์ƒ๋‹นํ•œ ๋น„์šฉ์ด ์†Œ์š”๋œ๋‹ค. ํ˜„์žฌ ๊ธฐ์ค€์œผ๋กœ FLAN-T5-XXL (11B) ๋ชจ๋ธ์„ CNN ๋ฐ์ผ๋ฆฌ ๋ฉ”์ผ ๋ฐ์ดํ„ฐ์…‹ 1๊ฐœ์— ๋Œ€ํ•ด ํŒŒ์ธํŠœ๋‹ํ•˜๋ฉด ์•ฝ 77๋งŒ ์›์ด ๋“ค๊ฒŒ ๋œ๋‹ค.

๋˜ ๋‹ค๋ฅธ ๋Œ€์•ˆ์œผ๋กœ๋Š” ์šฐ๋ฆฌ๊ฐ€ ์›ํ•˜๋Š” ์ •๋ณด๊ฐ€ ๋‹ด๊ธด ๋ฌธ์„œ๋ฅผ ์ง์ ‘ ํ”„๋กฌํ”„ํŠธ ์ฝ˜ํ…์ŠคํŠธ์— ๋„ฃ์–ด์ฃผ๊ณ  ์›ํ•˜๋Š” ์‘๋‹ต์„ ์–ป๋Š” ๊ฒƒ์ด๋‹ค. ์—ฐ๊ตฌ์— ๋”ฐ๋ฅด๋ฉด ํ”„๋กฌํ”„ํŠธ์˜ ๊ฐ€์น˜๋Š” ๋ฐ์ดํ„ฐ ์ƒ˜ํ”Œ ์•ฝ 100๊ฐœ์— ์ค€ํ•˜๋Š” ์ •๋„๋ผ๊ณ  ํ•œ๋‹ค.

- ํ•˜์ง€๋งŒ ๋ชจ๋“  ์ •๋ณด๋ฅผ ์ฝ˜ํ…์ŠคํŠธ์— ์ผ์ผ์ด ๋„ฃ์–ด์ฃผ๋Š” ๊ฒƒ์€ ํ˜„์‹ค์ ์œผ๋กœ ๋ถˆ๊ฐ€๋Šฅํ•˜๋‹ค. GPT-4๊ฐ€ ์ฒ˜๋ฆฌ ๊ฐ€๋Šฅํ•œ ์ž…๋ ฅ์€ ์ตœ๋Œ€ 50ํŽ˜์ด์ง€ ์ •๋„๋กœ ์ œํ•œ๋œ๋‹ค. ์ด๋Ÿฌํ•œ ์ƒํ™ฉ์—์„œ๋Š” ์šฐ๋ฆฌ๊ฐ€ ๊ฐ€์ง„ ์ •๋ณด๋ฅผ ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค์— ์ €์žฅํ•ด๋‘๊ณ , ์‚ฌ์šฉ์ž์˜ ์งˆ์˜๊ฐ€ ๋“ค์–ด์˜ฌ ๋•Œ (์˜ˆ๋ฅผ ๋“ค์–ด, ์ฑ—๋ด‡์„ ํ†ตํ•ด ์€ํ–‰ ๊ณ„์ขŒ๋ฅผ ๊ฐœ์„คํ•˜๋Š” ๋ฐฉ๋ฒ•์„ ๋ฌผ์—ˆ์„ ๋•Œ) ๊ด€๋ จ๋œ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜์—ฌ ํ•ด๋‹น ์ •๋ณด๊ฐ€ ๋‹ด๊ธด ๋ฌธ์„œ๋“ค์„ ํ”„๋กฌํ”„ํŠธ๋ฅผ ํ†ตํ•ด LLM์— ์ „๋‹ฌํ•˜๋Š” ๋ฐฉ์‹์ด ๋” ํšจ์œจ์ ์ด๋‹ค. ์ด๋Ÿฌํ•œ ์„œ๋น„์Šค ์•„ํ‚คํ…์ฒ˜๋ฅผ โ€˜๊ฒ€์ƒ‰ ์ฆ๊ฐ• ์ƒ์„ฑโ€™(Retrieval Augmented Generation, RAG)๋ผ๊ณ  ํ•œ๋‹ค.

- ์šฐ๋ฆฌ๋Š” ์ •๋ณด๋ฅผ ์–ป๊ธฐ ์œ„ํ•ด LLM์„ ํ™œ์šฉํ•˜๋Š”๋ฐ, LLM์—๊ฒŒ ์ •๋ณด๋ฅผ ์ž…๋ ฅํ•˜๋Š” ๋ฐฉ์‹์ด ๋‹ค์†Œ ์ด์ƒํ•ด ๋ณด์ผ ์ˆ˜ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์ •๋ณด์˜ ์ž…์ˆ˜์™€ ์ •๋ณด์˜ ์ฒ˜๋ฆฌ(ํ•ด์„)๋Š” ์™„์ „ํžˆ ๋‹ค๋ฅธ ๊ธฐ๋Šฅ์ด๋‹ค. ์ œ2์ฐจ ์„ธ๊ณ„ ๋Œ€์ „ ๋•Œ ์˜๊ตญ ์ •๋ณด๋ถ€๋Š” ๋…์ผ๊ตฐ์˜ ์•”ํ˜ธ๋ฅผ ์–ป์–ด์™€ ์•จ๋Ÿฐ ํŠœ๋ง์—๊ฒŒ ์ „๋‹ฌํ–ˆ๋‹ค. ๊ทธ๋Ÿฌ๋ฉด ํŠœ๋ง์˜ ์ฝœ๋กœ์„œ์Šค ๋จธ์‹ ์ด ์ด ์•”ํ˜ธ๋ฅผ ํ•ด์„ํ•ด ์ฃผ์—ˆ๋‹ค. ์ด๋Ÿฐ ๋ชจ์Šต์„ ๋จธ๋ฆฟ์†์— ๊ทธ๋ ค๋ณด๋ฉด ์ดํ•ด๊ฐ€ ๋น ๋ฅผ ๊ฒƒ์ด๋‹ค. ์ด๋ ‡๊ฒŒ ๋ถ„์—…์ด ๊ฐ€๋Šฅํ•œ ์ด์œ ๋Š” ๋ฐฐ์šฐ์ง€ ์•Š์€ ์ •๋ณด๋ฅผ ์ผ๋ฐ˜ํ™”ํ•˜์—ฌ ํ•ด์„ํ•˜๋Š” LLM์˜ ๋Šฅ๋ ฅ์ด ๋งค์šฐ ์šฐ์ˆ˜ํ•ด์ง„ ๊ฒฐ๊ณผ์ด๋‹ค. (์ด๋ฅผ ์ œ๋กœ์ƒท ํ•™์Šต ๋˜๋Š” ์ธ-์ฝ˜ํ…์ŠคํŠธ ํ•™์Šต์ด๋ผ๊ณ ๋„ ํ•œ๋‹ค.) ๋‹ค์‹œ ๋งํ•ด, ์–ด๋А ์ •๋„ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์ด ์žˆ๋Š” ์–ธ์–ด ๋ชจ๋ธ์ด ์•„๋‹ˆ๋ผ๋ฉด ์ด๋Ÿฌํ•œ ์•„ํ‚คํ…์ฒ˜๋ฅผ ์‚ฌ์šฉํ•˜๊ธฐ ์–ด๋ ค์šธ ๊ฒƒ์ด๋‹ค.

์ด์ „๋ถ€ํ„ฐ ์–ธ์–ด ๋ชจ๋ธ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๊ธฐ ์œ„ํ•ด ๊ฒ€์ƒ‰์„ ํ™œ์šฉํ•˜๋Š” ์‹œ๋„๊ฐ€ ๋งŽ์•˜์œผ๋ฉฐ, ๋Œ€ํ‘œ์ ์œผ๋กœ RETRO์™€ REARM์ด ์žˆ๋‹ค. ํ•˜์ง€๋งŒ ํ˜„์žฌ๋Š” RAG ์•„ํ‚คํ…์ฒ˜๊ฐ€ ์‚ฐ์—…๊ณ„์—์„œ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ๋‹ค. ์ด๋Š” ๋น„์•ฝ์ ์œผ๋กœ ํ–ฅ์ƒ๋œ ์ธ-์ฝ˜ํ…์ŠคํŠธ ํ•™์Šต ๋Šฅ๋ ฅ๊ณผ ๋ชจ๋ธ ํ•™์Šต์ด ๋”ฐ๋กœ ํ•„์š”ํ•˜์ง€ ์•Š์€ ํŽธ๋ฆฌํ•จ ๋•Œ๋ฌธ์ด๋‹ค.

- RAG ์•„ํ‚คํ…์ฒ˜์˜ ์ฃผ๋œ ์ž‘์—…ํ๋ฆ„์„ ์‚ดํŽด๋ณด์ž. ์ฒซ ๋ฒˆ์งธ ์ž‘์—… ํ๋ฆ„์€, ์šฐ๋ฆฌ๊ฐ€ ๊ฐ€์ง„ ์ •๋ณด๋ฅผ ๊ฒ€์ƒ‰ํ•˜๊ธฐ ์œ„ํ•ด DB์— ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๋Š” ์ผ์ด๋‹ค. ๊ด€๊ณ„ํ˜• ๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค(RDB)์— ์ •ํ˜•ํ™”๋œ ์ •๋ณด๋ฅผ ์ €์žฅํ•˜๊ณ , LLM์—๊ฒŒ SQL ๋ฌธ์„ ์ž‘์„ฑํ•˜์—ฌ ๊ฒ€์ƒ‰ํ•˜๋„๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•๋„ ๊ฐ€๋Šฅํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ ์šฐ๋ฆฌ๊ฐ€ ๊ฐ€์ง„ ์ •๋ณด๋Š” ์ž์—ฐ์–ด๋กœ ์ž‘์„ฑ๋œ ๋ฌธ์„œ ํŒŒ์ผ์ธ ๊ฒฝ์šฐ(์˜ˆ: PDF) ๊ฐ€ ๋” ํ”ํ•˜๋‹ค.

์ž์—ฐ์–ด๋ฅผ DB์— ์ €์žฅํ•˜๋ ค๋ฉด, ์ผ๋ฐ˜์ ์œผ๋กœ ์ˆซ์ž ํ˜•ํƒœ๋กœ ๋ณ€ํ™˜ํ•ด์•ผ ํ•œ๋‹ค. ํŠน์ • ๋‹จ์–ด๋‚˜ ๋ฌธ์žฅ๋“ค์ด ์กด์žฌํ•˜๋Š”์ง€ ์—ฌ๋ถ€๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š” ์ด์ง„๊ฐ’ ๋ฒกํ„ฐ๋กœ ๋ณ€ํ™˜ํ•  ์ˆ˜ ์žˆ์ง€๋งŒ, ๋ณด๋‹ค ํ˜„๋Œ€์ ์ธ ๊ธฐ์ˆ ์€ ์˜๋ฏธ๊ฐ€ ์œ ์‚ฌํ•œ ๋ฌธ์žฅ๋“ค์ด ์œ ์‚ฌํ•œ ๋ฒกํ„ฐ ๊ฐ’์„ ๊ฐ–๋„๋ก ๋ณ€ํ™˜ํ•˜๋Š” ๊ฒƒ์ด๋‹ค. ์ด๋ ‡๊ฒŒ ํ•จ์œผ๋กœ์จ ๋” ์‘์ถ•๋œ ๊ธธ์ด์˜ ๋ฒกํ„ฐ ํ˜•ํƒœ๋กœ ์ •๋ณด๋ฅผ ํ‘œํ˜„ํ•  ์ˆ˜ ์žˆ๊ฒŒ ๋œ๋‹ค.

โ€˜์˜๋ฏธ๊ฐ€ ์œ ์‚ฌํ•˜๋‹คโ€™๋ผ๋Š” ํ‘œํ˜„์ด ๋‚˜์™”๋Š”๋ฐ, ์ด๋Š” โ€˜์ •๋ณด์˜ ํ•ด์„โ€™ ๊ณผ์ •์ด ํ•„์š”ํ•จ์„ ๋œปํ•œ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด์„œ๋„ ์–ธ์–ด ๋ชจ๋ธ์˜ ๊ฐœ์ž…์ด ํ•„์š”ํ•˜๋‹ค. (๋ฐ˜๋“œ์‹œ LLM์ด ์•„๋‹ˆ์–ด๋„ ๋œ๋‹ค.) ์ˆ˜์ง‘ํ•œ ๋ฌธ์„œ๋“ค์„ ์ž‘์€ ์กฐ๊ฐ๋“ค๋กœ ๋ถ„ํ• ํ•˜๊ณ (์ฒญํ‚น), ์–ธ์–ด ๋ชจ๋ธ์˜ ํ•ด์„์„ ํ†ตํ•ด ๋ฒกํ„ฐ ํ˜•ํƒœ(์ž„๋ฒ ๋”ฉ)๋กœ ๋ณ€ํ™˜ํ•œ ํ›„, ์ด๋ฅผ ๋ฒกํ„ฐ DB์— ์ €์žฅํ•œ๋‹ค. ์ด ๊ณผ์ •์€ ๊ฒ€์ƒ‰ ์—”์ง„์˜ ์˜ค๋žœ ๊ธฐ์ˆ ์ด์ง€๋งŒ, ์ตœ๊ทผ์— ๋‹ค์‹œ ๊ฐ๊ด‘๋ฐ›๊ณ  ์žˆ๋‹ค.

- ๋ฌธ์„œ๋ฅผ ์ •์ œํ•˜๋Š” ๋ฐ์ดํ„ฐ ์ „์ฒ˜๋ฆฌ ์ž‘์—…(ETL)์€ ์ฃผ๋กœ Databricks, Apache Airflow์™€ ๊ฐ™์€ ๊ธฐ์กด ๋„๊ตฌ๋ฅผ ์‚ฌ์šฉํ•˜๋ฉฐ, LangChain ๋˜๋Š” LlamaIndex์˜ ๋ฌธ์„œ ๋กœ๋” ๊ธฐ๋Šฅ์„ ์ถ”๊ฐ€๋กœ ํ™œ์šฉํ•  ์ˆ˜๋„ ์žˆ๋‹ค.

์ž„๋ฒ ๋”ฉ ์ž‘์—…์€ ์ „์ฒ˜๋ฆฌ์— ํฌํ•จ๋˜๋ฉฐ, OpenAI์˜ API(text-embedding-ada-002), Cohere, ํ—ˆ๊น…ํŽ˜์ด์Šค์˜ BERT, Sentence Transformer, GPT-J ๋“ฑ์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค. ์ž„๋ฒ ๋”ฉ์€ LLM ์•ฑ์˜ ์„ฑ๋Šฅ์— ์ง์ ‘์ ์ธ ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฏ€๋กœ, ๋งค์šฐ ์‹ ์ค‘ํ•˜๊ฒŒ ์„ ํƒํ•ด์•ผ ํ•œ๋‹ค. ๋˜ํ•œ, ์šฐ์ˆ˜ํ•œ ์ž„๋ฒ ๋”ฉ์ด๋”๋ผ๋„ ์šฐ๋ฆฌ๊ฐ€ ์›ํ•˜๋Š” ์–ธ์–ด๋ฅผ ์ง€์›ํ•˜์ง€ ์•Š๋Š”๋‹ค๋ฉด ์‚ฌ์šฉํ•  ์ˆ˜ ์—†๋‹ค.

๋ฐ์ดํ„ฐ๋ฒ ์ด์Šค๋Š” RDB๊ฐ€ ์•„๋‹Œ ๋ฒกํ„ฐ DB๋ฅผ ์‚ฌ์šฉํ•ด์•ผ ํ•œ๋‹ค. ํด๋ ˆ์ดํฟ AI ์ฐฝ๋ฆฝ์ž์ธ ์นฉ ํ›„์˜Œ์ด โ€˜๊ทธ๋ž˜ํ”„ DB์˜ ํ•ด๊ฐ€ 2021๋…„์ด๋ผ๋ฉด, ๋ฒกํ„ฐ DB์˜ ํ•ด๋Š” 2023๋…„โ€™์ด๋ผ๊ณ  ๋งํ•œ ๊ฒƒ์ฒ˜๋Ÿผ ์ด ๋ถ„์•ผ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์ด ๋†’๋‹ค. Pinecone์€ ์•ˆ์ •์ ์ธ ํด๋ผ์šฐ๋“œ ํ˜ธ์ŠคํŒ… ๊ธฐ๋ฐ˜์œผ๋กœ ์„œ๋น„์Šค๋˜๋ฉฐ, Weaviate, Vespa, Qdrant๋Š” ๋‹จ์ผ ๋…ธ๋“œ ๊ธฐ๋ฐ˜์˜ ์˜คํ”ˆ์†Œ์Šค DB์ด๋‹ค. Chroma, Faiss๋Š” ๋กœ์ปฌ ๋ฒกํ„ฐ ๊ด€๋ฆฌ ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋กœ, ์—„๋ฐ€ํžˆ DB๋Š” ์•„๋‹ˆ์ง€๋งŒ (์ƒค๋”ฉ ๋ฐ ๋ณต์ œ๋ฅผ ํ†ตํ•œ ์„ฑ๋Šฅ ๋ณด์žฅ, ๋‚ด๊ฒฐํ•จ์„ฑ ํ–ฅ์ƒ, ๋ชจ๋‹ˆํ„ฐ๋ง, ์ ‘๊ทผ ์ œ์–ด, ๋ฐฑ์—…๊ณผ ์ปฌ๋ ‰์…˜ ๋“ฑ DB์˜ ์ฃผ๋œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜์ง€ ์•Š๊ธฐ ๋•Œ๋ฌธ์—) ์ž„๋ฒ ๋”ฉ ๊ฒ€์ƒ‰ ์šฉ๋„๋กœ ๊ฐ€๋ณ๊ฒŒ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค. pgvector์™€ ๊ฐ™์ด PostgreSQL OLTP ํ™•์žฅํŒ๋„ ๋“œ๋ฌผ์ง€๋งŒ ์‚ฌ์šฉ๋˜๋Š” ์˜ˆ์‹œ๊ฐ€ ์žˆ๋‹ค.

- ์ž‘์—… ํ๋ฆ„์œผ๋กœ ๋Œ์•„๊ฐ€, ๋‘ ๋ฒˆ์งธ ํ๋ฆ„์€ LLM ์•ฑ์— ์งˆ์˜๊ฐ€ ๋“ค์–ด์˜ค๋ฉด (์˜ˆ๋ฅผ ๋“ค์–ด, "์€ํ–‰ ๊ณ„์ขŒ๋ฅผ ๊ฐœ์„คํ•˜๋ ค๋ฉด ์–ด๋–ป๊ฒŒ ํ•ด์•ผ ํ•˜๋‚˜์š”?") ์ด ๋ฌธ์žฅ์„ ์ž„๋ฒ ๋”ฉ ๋ฒกํ„ฐ๋กœ ๋งŒ๋“ค๊ณ , ๋ฒกํ„ฐ DB์—์„œ ์ˆ˜์น˜์ ์œผ๋กœ ๊ฑฐ๋ฆฌ๊ฐ€ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ๋ฌธ์„œ๋“ค์„ ๊ฒ€์ƒ‰ํ•˜์—ฌ ๊ฐ€์ ธ์˜จ๋‹ค.

๊ทธ๋Ÿฌ๋‚˜ ๊ณ„์‚ฐํ•™์ ์œผ๋กœ '๊ฐ€์žฅ ๊ฐ€๊นŒ์šด' ๋ฌธ์„œ ๋ฒกํ„ฐ๋ฅผ ์ฐพ๋Š” ๊ฒƒ์€ ๋ฒกํ„ฐ์˜ ํฌ๊ธฐ (์ฐจ์›)์ด ์ปค์งˆ์ˆ˜๋ก ์–ด๋ ค์›Œ์ง„๋‹ค. ๋”ฐ๋ผ์„œ '๊ทผ์‚ฌ์ ์œผ๋กœ ๊ฐ€๊นŒ์šด' (Approximate Nearest Neighbor, ANN) ๋ฒกํ„ฐ๋ฅผ ๋น ๋ฅด๊ฒŒ ์ฐพ๋Š” ๋ฐฉ๋ฒ•์ด ์ฃผ๋กœ ์‚ฌ์šฉ๋œ๋‹ค. ๋น ๋ฅธ ๊ฒ€์ƒ‰์„ ์œ„ํ•ด ๋ฒกํ„ฐ๋Š” ๋ฏธ๋ฆฌ ์ธ๋ฑ์‹ฑ๋˜๋ฉฐ, ์ด๋ฅผ ์œ„ํ•ด ๋ฌด์ž‘์œ„ ํˆฌ์˜, PQ (Product Quantization), LSH (Locality Sensitive Hashing), HNSW (Hierarchical Navigable Small World) ๋“ฑ์˜ ๊ธฐ๋ฒ•๋“ค์ด ์ ์šฉ๋œ๋‹ค. ์ด๋Ÿฌํ•œ ๊ธฐ๋ฒ•๋“ค์˜ ์œ ๋ฌด์™€ ๊ตฌํ˜„์€ ๋ฒกํ„ฐ DB๋งˆ๋‹ค ๋‹ค๋ฅด๋ฉฐ, ์ด๋Š” LLM ์•ฑ์˜ ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ๋ฏธ์น˜๋ฏ€๋กœ ์ฃผ์˜๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ๋˜ํ•œ, ๋ฒกํ„ฐ ์ธ๋ฑ์‹ฑ ๋Œ€์‹ ์— ํŠธ๋ฆฌ๋‚˜ ์ง€์‹ ๊ทธ๋ž˜ํ”„ ๊ตฌ์กฐ ๋˜๋Š” ์ด๋“ค์˜ ์กฐํ•ฉ์„ ํ™œ์šฉํ•œ ์ธ๋ฑ์‹ฑ์„ ๊ณ ๋ คํ•˜๊ธฐ๋„ ํ•œ๋‹ค.

๋˜ํ•œ ์งˆ์˜๋ฌธ์ด ๋„ˆ๋ฌด ๊ฐ„๊ฒฐํ•˜๋‹ค๋ฉด ๊ด€๋ จ์„ฑ ์žˆ๋Š” ๋ฌธ์„œ๋ฅผ ์ฐพ์•„์˜ค๊ธฐ ์–ด๋ ค์šธ ์ˆ˜ ์žˆ๊ณ  ์ด๋ฅผ ๋ณด์™„ํ•˜๊ธฐ ์œ„ํ•œ ์—ฐ๊ตฌ๋“ค์ด ์žˆ๋‹ค. ์˜ˆ์ปจ๋Œ€, ๊ฐ€์ƒ์˜ ๋ฌธ์„œ๋ฅผ ์ž„๋ฒ ๋”ฉํ•˜์—ฌ ์งˆ์˜๋ฅผ ๋ณด๊ฐ•ํ•œ๋‹ค๊ฑฐ๋‚˜ ์ฝ˜ํ…์ŠคํŠธ ๋ฌธ์„œ ์ƒ์„ฑ์— ํด๋Ÿฌ์Šคํ„ฐ๋ง ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ ์šฉํ•œ '์ƒ์„ฑ ํ›„ ์ฝ๊ธฐ' (GenRead) ๊ธฐ๋ฒ• ๋“ฑ์ด ์žˆ๋‹ค.

- ์„ธ ๋ฒˆ์งธ ์ž‘์—…ํ๋ฆ„์œผ๋กœ ์ง„ํ–‰๋  ๋‚ด์šฉ์„ ์„ค๋ช…ํ•ด๋ณด์ž. ์งˆ์˜์™€ ๊ด€๋ จ ์žˆ๋Š” ๋ฌธ์„œ(์˜ˆ๋ฅผ ๋“ค์–ด, ๋น„๋Œ€๋ฉด ์€ํ–‰ ๊ณ„์ขŒ ๊ฐœ์„ค ์•ˆ๋‚ด ๋ฌธ์„œ)๋ฅผ ๊ฐ€์ ธ์™”๋‹ค๋ฉด, ์ด๋ฅผ ํ”„๋กฌํ”„ํŠธ ์ฝ˜ํ…์ŠคํŠธ์— ์ž…๋ ฅํ•˜๊ณ  LLM์—๊ฒŒ ๋‹ต๋ณ€์„ ์š”์ฒญํ•ด์•ผ ํ•œ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด์„œ๋Š” ์งˆ์˜์— ๋Œ€ํ•œ ์ž„๋ฒ ๋”ฉ ์ฒ˜๋ฆฌ, ๊ฒ€์ƒ‰ ์š”์ฒญ, ๊ฒ€์ƒ‰ ๊ฒฐ๊ณผ๋ฅผ ํ™œ์šฉํ•œ ํ”„๋กฌํ”„ํŠธ ์—”์ง€๋‹ˆ์–ด๋ง ์ž‘์—…์ด ํ•„์š”ํ•˜๋‹ค. ์ด๋Ÿฌํ•œ ์ž‘์—…์„ ํŠน์ • ์ฃผ์ฒด๊ฐ€ ์ค‘์•™์—์„œ ์ œ์–ดํ•ด์ฃผ๋Š”๋ฐ, ์ด๋ฅผ ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜ ํ”„๋ ˆ์ž„์›Œํฌ๋ผ๊ณ  ํ•œ๋‹ค.

์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜ ํ”„๋ ˆ์ž„์›Œํฌ ์ค‘ ๋Œ€ํ‘œ์ ์ธ ๊ฒƒ์ด Langchain์ด๋ฉฐ, LlamaIndex (์ตœ์ดˆ ๋ช…์นญ์€ GPT Index)๋„ ์ฃผ๋ชฉ๋ฐ›๊ณ  ์žˆ๋‹ค. Langchain์€ ์žฅํ™ฉํ•œ ์ธํ„ฐํŽ˜์ด์Šค์™€ ๋А๋ฆฐ ๊ตฌํ˜„์œผ๋กœ ์ธํ•ด ํŒŒ์ด์ฌ์œผ๋กœ ์ง์ ‘ ๊ฐœ๋ฐœํ•˜๋Š” ๊ฒฝ์šฐ๋„ ์žˆ๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ Langchain์€ ์•„์ง 0.0.239 ๋ฒ„์ „์œผ๋กœ ๊ฐœ๋ฐœ์ด ์ง„ํ–‰ ์ค‘์ด๋ฏ€๋กœ ๊ธฐ๋‹ค๋ ค๋ด์•ผ ํ•  ํ•„์š”๊ฐ€ ์žˆ๋‹ค. ์žฌ๋ฐŒ๊ฒŒ๋„ ChatGPT์— ํ”Œ๋Ÿฌ๊ทธ์ธ์„ ์‚ฌ์šฉํ•˜์—ฌ ์˜ค์ผ€์ŠคํŠธ๋ ˆ์ด์…˜์„ ์‹œํ—˜ํ•ด๋ณผ ์ˆ˜๋„ ์žˆ๋‹ค.

LLM์˜ ์ฑ”ํ”ผ์–ธ์€ ๋‹น์—ฐํžˆ OpenAI์˜ gpt-4, gpt-4โ€“32k์ด๋ฉฐ, 50๋ฐฐ ์ €๋ ดํ•˜๊ณ  ๋น ๋ฅธ gpt-3.5-turbo๋ฅผ ์‚ฌ์šฉํ•ด๋ณผ ์ˆ˜๋„ ์žˆ๋‹ค. ๋˜ํ•œ, Anthropic์˜ Claude๋Š” GPT-3.5 ์ˆ˜์ค€์˜ ์ •ํ™•๋„์™€ ์‹ ์†ํ•œ ์ถ”๋ก  ์†๋„๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์œผ๋ฉฐ, ํŠนํžˆ Claude 2๋Š” 100k ๊ธธ์ด์˜ ์ฝ˜ํ…์ŠคํŠธ ์ฐฝ์„ ์ง€์›ํ•œ๋‹ค๊ณ  ํ•œ๋‹ค. ์ผ๋ถ€ ๋‹จ์ˆœํ•œ ์š”์ฒญ์€ ์˜คํ”ˆ์†Œ์Šค ๋ชจ๋ธ๋กœ ๋ถ„๊ธฐํ•˜์—ฌ ๋Œ€์‘ํ•˜๋ฉด์„œ ๋น„์šฉ์„ ๋‚ฎ์ถœ ์ˆ˜๋„ ์žˆ๋‹ค.
์ด๋Ÿฐ ๊ฒฝ์šฐ๋ฅผ ๋Œ€๋น„ํ•˜์—ฌ Databricks, AnyScale, Mosaic, Modal, RunPod ํ”Œ๋žซํผ์€ ๋ชจ๋ธ ํŒŒ์ธํŠœ๋‹ ๋„๊ตฌ๋ฅผ, Hugging Face์™€ Replicate๋Š” API ์ธํ„ฐํŽ˜์ด์Šค๋ฅผ ์ œ๊ณตํ•˜๊ณ  ์žˆ๋‹ค.

์‹œ์ค‘์—๋Š” ๋‹ค์–‘ํ•œ ์ƒ์—…์šฉ ์˜คํ”ˆ์†Œ์Šค LLM ๋ชจ๋ธ์ด ์กด์žฌํ•˜๋ฉฐ(Together, Mosaic, Falcon, Mistral ๋“ฑ) ์ด๋“ค์ด GPT-3.5์˜ ์ •ํ™•๋„ ์ˆ˜์ค€์— ๋‹ค๋‹ค๋ฅธ๋‹ค๋ฉด ์‚ฐ์—…๊ณ„์—์„œ๋Š” ์˜คํ”ˆ์†Œ์Šค ์‚ฌ์šฉ๋„ ๊ณ ๋ คํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋ณด์ธ๋‹ค. ์ตœ๊ทผ Meta์—์„œ Llama2๋ฅผ ๋ฆด๋ฆฌ์ฆˆํ•˜๋ฉด์„œ LLM ์”ฌ์—๋„ ๊ฐœ์„ฑ์žˆ๊ณ  ์šฐ์ˆ˜ํ•œ ๋ชจ๋ธ์ด ์Ÿ์•„์ ธ ๋‚˜์˜ค๋Š” ์Šคํ…Œ์ด๋ธ” ๋””ํ“จ์ „ ๋ชจ๋จผํŠธ๊ฐ€ ๋ฐœ์ƒํ•˜์ง€ ์•Š์„๊นŒ ๊ทธ ๊ท€์ถ”๊ฐ€ ์ฃผ๋ชฉ๋˜๊ณ  ์žˆ๋‹ค.

- ์ฃผ๋œ ์ž‘์—…ํ๋ฆ„์— ๋Œ€ํ•œ ์ด์•ผ๊ธฐ๋Š” ๋๋‚ฌ๋‹ค. ์š”์•ฝํ•˜์ž๋ฉด, ์–ด๋–ค ์ž„๋ฒ ๋”ฉ, ๋ฒกํ„ฐ DB, ๊ทธ๋ฆฌ๊ณ  LLM์„ ์‚ฌ์šฉํ• ์ง€๊ฐ€ ์• ํ”Œ๋ฆฌ์ผ€์ด์…˜์˜ ์„ฑ๋Šฅ์„ ์ขŒ์šฐํ•˜๋Š” ํ•ต์‹ฌ ์š”์†Œ์ด๋‹ค. MLOps์™€ ๋น„์Šทํ•˜๊ฒŒ RAG ์ฃผ์œ„๋กœ๋Š” ์ˆ˜๋งŽ์€ LLMOps ๋„๊ตฌ๋“ค์˜ ์ƒํƒœ๊ณ„๊ฐ€ ํ˜•์„ฑ๋˜๊ณ  ์žˆ๋‹ค. LLM ์‘๋‹ต ์บ์‹ฑ์„ ์œ„ํ•ด Redis, GPTCache๊ฐ€, LLM ์ถœ๋ ฅ ๋ชจ๋‹ˆํ„ฐ๋ง, ์ถ”์ , ํ‰๊ฐ€๋ฅผ ์œ„ํ•ด W&B, MLflow, PromptLayer, Helicone์ด, ์ถœ๋ ฅ์˜ ์œ ํ•ด์„ฑ ๊ฒ€์ฆ์„ ์œ„ํ•ด Guardrails๊ฐ€, ๊ทธ๋ฆฌ๊ณ  ํ”„๋กฌํ”„ํŠธ ์ฃผ์ž… ๊ณต๊ฒฉ ๋ฐฉ์–ด๋ฅผ ์œ„ํ•ด Rebuff๊ฐ€ ์ œ๊ณต๋˜๊ณ  ์žˆ๋‹ค.

์ „์ฒด์ ์ธ RAG ์•„ํ‚คํ…์ฒ˜ ํ˜ธ์ŠคํŒ…์€ ์ฃผ๋กœ ํด๋ผ์šฐ๋“œ ๊ณต๊ธ‰์ž ์„œ๋น„์Šค ๋‚ด์—์„œ ์ด๋ฃจ์–ด์ง€๊ฒ ์ง€๋งŒ, ํฅ๋ฏธ๋กœ์šด ๋‘ ๊ฐ€์ง€ ์‹œ๋„๊ฐ€ ์žˆ๋‹ค. Streamship์€ LLM ์•ฑ์„ ์—”๋“œ-ํˆฌ-์—”๋“œ๋กœ ํ˜ธ์ŠคํŒ…ํ•˜๋ฉด์„œ ๋‹ค์–‘ํ•œ ๊ธฐ๋Šฅ์„ ์ œ๊ณตํ•˜๋ ค๊ณ  ํ•˜๊ณ , ๋…์ž์ ์ธ LLM ๋ชจ๋ธ์„ ๊ฐ–์ถ”๊ณ  ์žˆ๋Š” AnyScale, Mosaic์€ ๋ชจ๋ธ๊ณผ ํŒŒ์ด์ฌ ์ฝ”๋“œ๋ฅผ ํ•œ ๊ณณ์—์„œ ํ˜ธ์ŠคํŒ… ๊ฐ€๋Šฅํ•˜๋„๋ก ์ง€์›ํ•˜๊ณ  ์žˆ๋‹ค.

- ๋‹ค์Œ๋ฒˆ์—๋Š” RAG ์•„ํ‚คํ…์ฒ˜์— ๋Œ€ํ•ด AWS ์„œ๋น„์Šค๋กœ ํ•œ์ •ํ•˜์—ฌ ์ ์–ด๋ณผ ์˜ˆ์ •์ด๋‹ค.

์ฐธ์กฐ:
- Emerging Architectures for LLM Applications(https://a16z.com/2023/06/20/emerging-architectures-for-llm-applications/)
- Knowledge Retrieval Architecture for LLMโ€™s (2023)(https://mattboegner.com/knowledge-retrieval-architecture-for-llms/?fbclid=IwAR1zYHaStMU-DJuP84iZ-h9K1vMMb_qVtBsQ5FotPqAjv5Lb4fyyBG_nAoI)
- Building LLM Applications for Production(https://huyenchip.com//2023/04/11/llm-engineering.html)
- ChatGPT์˜ ์ „๋‘์—ฝ(์žฅ๊ธฐ๊ธฐ์–ต ์ €์žฅ์†Œ)์œผ๋กœ ๊ฐ๊ด‘๋ฐ›๊ณ  ์žˆ๋Š” Vector DB์— ๋Œ€ํ•ด ์•Œ์•„๋ณด์ž(https://devocean.sk.com/experts/techBoardDetail.do?ID=164964&fbclid=IwAR2XOaqC8QwGjWoThJmUG-mldeuyiIpLefJHmzd9qJE-GuFhqXlQB-ODwi8)
AI๋กœ ์ธํ•ด์„œ ์ƒˆ๋กœ์šด UX(์Œ์„ฑ, ์ด๋ฏธ์ง€, ์˜์ƒ, ํ…์ŠคํŠธ) ๋“ฑ ์ƒˆ๋กœ์šด ๊ธฐ๊ธฐ๊ฐ€ ๋“ฑ์žฅํ•  ์ˆ˜๋„ ์žˆ๊ฒ ๋„ค์š”. ์ด ์˜์ƒ์€ Iphone์„ ์ฒ˜์Œ ๊ฐœ๋ฐœํ•˜๋˜ ์‹œ๊ธฐ์— ํ™œ๋™ํ–ˆ๋˜ ์ž„์›๋“ค์˜ ์ธํ„ฐ๋ทฐ์ž…๋‹ˆ๋‹ค.

์ด ๋‚ด๋Ÿฌํ‹ฐ๋ธŒ๋Š” ์•„์ดํฐ์˜ ๊ตฌ์ƒ๋ถ€ํ„ฐ ๋Œ€์ค‘์—๊ฒŒ ๊ณต๊ฐœ๋˜๊ธฐ๊นŒ์ง€์˜ ์• ํ”Œ์˜ ์•„์ดํฐ ๊ฐœ๋ฐœ ์—ฌ์ •์— ๋Œ€ํ•ด ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ์Šคํ‹ฐ๋ธŒ ์žก์Šค๊ฐ€ ์ด๋„๋Š” ์• ํ”Œ ํŒ€์€ ์ฒ˜์Œ์— ์•„์ดํŒŸ์„ ๊ฐœ์„ ํ•˜๊ณ  ํœด๋Œ€ํฐ๊ณผ์˜ ์ž ์žฌ์  ๊ฒฝ์Ÿ์— ๋งž์„œ๊ธฐ ์œ„ํ•ด ์•„์ดํŒŸ์„ ํœด๋Œ€ํฐ๊ณผ ๊ฒฐํ•ฉํ•˜์—ฌ ์‹œ์žฅ ์ง€์œ„๋ฅผ ์œ ์ง€ํ•˜๋Š” ๊ฒƒ์„ ๋ชฉํ‘œ๋กœ ์‚ผ์•˜์Šต๋‹ˆ๋‹ค. ์ฒซ ๋ฒˆ์งธ ๋””์ž์ธ์€ ์•„์ดํŒŸ์˜ ์ธํ„ฐํŽ˜์ด์Šค์™€ ํ•˜๋“œ์›จ์–ด ํ‚ค๋ณด๋“œ๋ฅผ ํ†ตํ•ฉํ–ˆ์ง€๋งŒ, ๋…ผ๋ฆฌ์ ์ด๊ณ  ์ง๊ด€์ ์ธ ๋‹ค์ด์–ผ๋ง ๋ฉ”์ปค๋‹ˆ์ฆ˜์„ ๊ฐœ๋ฐœํ•˜๋Š” ๋ฐ ์–ด๋ ค์›€์„ ๊ฒช์—ˆ์Šต๋‹ˆ๋‹ค.

๊ทธ๋Ÿฌ๋˜ ์ค‘ ํƒœ๋ธ”๋ฆฟ์šฉ์œผ๋กœ ๊ฐœ๋ฐœํ•˜๋˜ ํ„ฐ์น˜ ๊ธฐ์ˆ ์„ ํฌ์ผ“ ์‚ฌ์ด์ฆˆ ํœด๋Œ€ํฐ์— ์ ์šฉํ•˜๋Š” ์•„์ด๋””์–ด๋ฅผ ๋– ์˜ฌ๋ฆฌ๊ฒŒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. ํŒ€์€ ๋ช‡ ๊ฐ€์ง€ ์žฅ์• ๋ฌผ๊ณผ ๊ฒฉ๋ ฌํ•œ ์ž‘์—…์— ์ง๋ฉดํ–ˆ๊ณ , ์žก์Šค๋Š” ์œ ๋งํ•œ ๊ฒฐ๊ณผ๋ฌผ์„ ์ œ์‹œํ•˜์ง€ ๋ชปํ•˜๋ฉด ํ”„๋กœ์ ํŠธ๋ฅผ ๋‹ค๋ฅธ ํŒ€์— ๋„˜๊ธฐ๊ฒ ๋‹ค๊ณ  ์œ„ํ˜‘ํ–ˆ์Šต๋‹ˆ๋‹ค.

๊ฐ€์žฅ ์–ด๋ ค์› ๋˜ ๋ถ€๋ถ„ ์ค‘ ํ•˜๋‚˜๋Š” ๊ธฐ๋Šฅ์ ์ด๊ณ  ์‚ฌ์šฉ์ž ์นœํ™”์ ์ธ ์†Œํ”„ํŠธ์›จ์–ด ํ‚ค๋ณด๋“œ๋ฅผ ์„ค๊ณ„ํ•˜๋Š” ๊ฒƒ์ด์—ˆ๋Š”๋ฐ, ์ด๋ฏธ ์ž˜ ์•Œ๋ ค์ง„ BlackBerry์™€ ๋น„๊ต๋  ๊ฒƒ์„ ์•Œ์•˜๊ธฐ ๋•Œ๋ฌธ์ž…๋‹ˆ๋‹ค. ํ‚ค๋ณด๋“œ ์†”๋ฃจ์…˜์€ ํƒ€์ดํ•‘ํ•˜๋Š” ๋™์•ˆ ๋‹ค์Œ ๊ธ€์ž๊ฐ€ ๋‹ฟ์„ ๊ฐ€๋Šฅ์„ฑ์ด ์žˆ๋Š” ์˜์—ญ์„ ํ™•๋Œ€ํ•˜๋Š” ๋™์  ํƒ€๊ฒฉ ์˜์—ญ ์‹œ์Šคํ…œ์„ ๊ฐœ๋ฐœํ•œ ํŒ€์›์œผ๋กœ๋ถ€ํ„ฐ ๋‚˜์™”์Šต๋‹ˆ๋‹ค.

์ด ํ”„๋กœ์ ํŠธ๋Š” Apple ๋‚ด๋ถ€์—์„œ๋„ ์ผ๊ธ‰ ๋น„๋ฐ€๋กœ ์œ ์ง€๋˜์–ด ์†Œํ”„ํŠธ์›จ์–ด ํŒ€๊ณผ ํ•˜๋“œ์›จ์–ด ํŒ€์ด ์„œ๋กœ์˜ ์ž‘์—… ๋‚ด์šฉ์„ ์•Œ์ง€ ๋ชปํ•˜๋Š” ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•˜์Šต๋‹ˆ๋‹ค. ๋งˆ์นจ๋‚ด ์žก์Šค๊ฐ€ ๋ผ์ด๋ธŒ ๋ฐ๋ชจ๋ฅผ ํ†ตํ•ด ์ œํ’ˆ์„ ๊ณต๊ฐœํ–ˆ๊ณ , ์ถœ์‹œ์™€ ๋™์‹œ์— ๋Œ€์ค‘์˜ ์—„์ฒญ๋‚œ ๊ด€์‹ฌ์„ ๋ฐ›์œผ๋ฉฐ ์•„์ดํฐ์ด ๊ณง ์‚ฌํšŒ์ ์œผ๋กœ ํฐ ์˜ํ–ฅ์„ ๋ฏธ์น  ๊ฒƒ์ž„์„ ์•”์‹œํ–ˆ์Šต๋‹ˆ๋‹ค.

๊ด€๋ จํ•ด์„œ๋Š” ์ด๋Ÿฐ ์ฑ…๋“ค๋„ ์ข‹๊ฒ ๋„ค์š”.
https://www.amazon.com/Creative-Selection-Inside-Apples-Process/dp/1250194466
https://www.amazon.com/Build-Unorthodox-Guide-Making-Things/dp/B09CF2YB6Z/ref=sr_1_1?crid=3MS1L5GA70UZ1&keywords=build&qid=1690167035&s=books&sprefix=build%2Cstripbooks-intl-ship%2C274&sr=1-1
Relax. Not Everyone is Out to Get You
โ€œIf you ever feel that the world is against you, you are not alone. We all have a tendency to assume that when anything goes wrong, the fault lies within some great conspiracy against us. A co-worker fails to give you a report in time? They must be trying to derail your career and beat you to a promotion. Your child drops and breaks an expensive plate? They must be trying to annoy you and waste your time. WiFi in a coffee shop not working? The staff must be lying about having it to lure you in and sample their crappy espresso. But the simple fact is that these explanations which we tend to jump to are rarely true."
Who controls your focus?
โ€œFocus was ingrained in Jobsโ€™s personality and had been honed by his Zen training. He relentlessly filtered out what he considered distractions. Colleagues and family members would at times be exasperated as they tried to get him to deal with issuesโ€”a legal problem, a medical diagnosisโ€”they considered important. But he would give a cold stare and refuse to shift his laserlike focus until he was ready.โ€
Key questions for a new business

1. The customers
1. How big?
2. How fast growing?
2. Problems?
1. How severe?
3. Solutions
1. Why are or will customer not satisfied with this current solution?
4. Your Solution
1. What do we understand that others donโ€™t?
- Could we do that better, cheaper, faster or totally different way?
- Do you understand the competition?
5. Competition
1. Do you have competition?
2. Do you have any proxy in other areas?
6. Distribution
1. Which channel is ideal to scale this idea to 10, 100x?
7. Unit Economics
1. if the difference between money spent and money earned per customer is (+), the business will make money someday.
2. Even if it's - now, if it's getting better, it's something to look forward to.
8. Why now?
1. How has this opportunity evolved?
2. Have technologies been evolved?
3. Have the regulation been changed?
9. Why you?
1. Do you want to address this problem for several years?
10. Idea space?
1. Does this area or problem have a good potential?
AI startup ํ•˜์‹œ๋Š” ๋ถ„๋“ค 7์›” 30์ผ๊นŒ์ง€ Sequioa Arc(์ดˆ๊ธฐ ์Šคํƒ€ํŠธ์—… ํˆฌ์ž ํ”„๋กœ๊ทธ๋žจ like YC)๊ฐ€ ์—ด๋ ธ๋„ค์š”.
https://www.sequoiacap.com/arc
Pros
Sequioa๋„คํŠธ์›Œํฌ๋ฅผ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ๋‹ค.
Cons
๋งŒ์•ฝ Sequioa๊ฐ€ ํ›„์†ํˆฌ์ž ์•ˆ ํ•ด์ฃผ๋ฉด ์™œ ๋ชป๋ฐ›์•˜๋Š”์ง€ ์งˆ๋ฌธ์„ ๋งŽ์ด ๋ฐ›์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
โ€œ์‹ค๋ฆฌ์ฝ˜๋ฐธ๋ฆฌ์—์„œ๋Š” ์ข์€ ๊ธฐํšŒ๋ฅผ ๋šซ๊ธฐ ์œ„ํ•ด ์„œ๋กœ ์‹ธ์šฐ๊ณ  ๊ฒฝ์Ÿํ•˜๊ธฐ๋ณด๋‹ค๋Š”, ๊ณต์œ ํ•˜๊ณ  ํ˜‘๋ ฅํ•ฉ๋‹ˆ๋‹ค. ์„ฑ๊ณตํ•œ ์‚ฌ๋žŒ์„ ์งˆํˆฌํ•˜๊ธฐ ๋ณด๋‹ค๋Š”, ๋‹ค์Œ ์„ธ๋Œ€๋ฅผ ๋Œ์–ด์ฃผ๊ณ  ๋„์™€์ค๋‹ˆ๋‹ค. ์˜ค๋Š˜ ๋งŒ๋‚œ ์ธ์—ฐ์œผ๋กœ ๋‹คํ•จ๊ป˜ ๋” ํ˜‘๋ ฅํ•˜๊ณ , ๊ณต์œ ํ•˜์—ฌ AI๋ฅผ ๋ฐœ์ „์‹œํ‚ค๊ธฐ๋ฅผ ๊ธฐ์›ํ•ฉ๋‹ˆ๋‹ค.โ€
โค1
How innovation happens in bio pharma?

Builders will need to understand both a) how to exploit the latest and greatest AI tech, and (perhaps more importantly) b) how to commercialize a product or platform in biopharma and healthcare with a defensible product and go-to-market strategy. As such, we believe teams with depth in both (scientists, AI experts, healthcare builders and operators, product and go-to-market experts) will be best poised to lead and win in this new era.
What AI startups should you keep an eye on? :robot_face: ๐Ÿ”ฅ We asked some incredible investors and founders for their picks. Here are 13 companies pushing the AI frontier ๐Ÿ‘‡

1. Alife
Alife is revolutionizing IVF with AI-powered tools. Itโ€™s enhancing decision-making at crucial stages like ovarian stimulation & embryo selection, making fertility treatments more accessible and efficient.
- Rebecca Kaden, Union Square Ventures

2. Glean
Glean uses AI to provide unified, contextual search across apps. Quickly becoming more than just a tool - itโ€™s an intuitive work assistant.
- Josh Coyne, Kleiner Perkins

3. LanceDB (YC W22)
Multi-modal AI is revolutionary, but data management remains a challenge. Lance optimizes storage and handling for this unstructured data, increasing performance, & reducing costs.
- Saar Gur, CRV

4. Abnormal Security
Weโ€™ve seen a surge in AI-powered fraud. As these sophisticated attacks rise, so does the need for AI defenses. Abnormal Security uses AI to counter AI threats.
- Saam Motamedi, Greylock

5. Factory
Factory is creating AI coding โ€œdroidsโ€ designed to take care of an engineerโ€™s annoying busywork, tackling routine tasks like code review and debugging.
- Markie Wagner, Delphi Labs

6. Lamini
Lamini LLM engine creates and fine-tunes customized, private models. It also has a neat partnership with Databricks, making it even easier to get up and running.
- Todd Jackson & James C. Wu, First Round Capital

7. Sereact
Sereact is revolutionizing warehouse automation, leveraging AI to train its robot arm. From picking electronic devices to soft fruits, its arm navigates spatial and physical nuances v well.
- Nathan Benaich, Air Street Capital

8. Mistral AI
Mistral, founded by impressive AI talent, is developing superior OS language models. A potential European challenger to OpenAI.
- Stanislas Polu, Dust

9. poolside
Poolside is another player in the AI-programming space. Its approach is to create a dedicated foundation model focused on one use case: code generation.
- Matan Grinberg, Factory

10. NewLimit
NewLimit is using ML to change the game in epigenetic reprogramming. Their approach could be transformative for treating intractable diseases.
- Simon Barnett, Dimension

11. Runway
Runway is building a new creative suite with AI. It brings professional-grade video creation to anyone. Already used by Fortune 500 and major movies productsions.
- Grace Isford, Lux Capital

12. Labelbox
By making it easy to select, annotate, and assess data, Labelbox makes it easier to experiment with using AI models like GPT-4.
- Robert Kaplan, SoftBank

13. Dust
Dust leverages LLMs for enterprise productivity. The startup is building a โ€œteam operating systemโ€ designed to augment knowledge workers.
- Konstantine Buhler, Sequoia Capital

Thereโ€™s a lot more insight and detail in the piece, linked below. Jump in and subscribe for more glimpses of the future ๐Ÿ™‚
๐Ÿ‘1
โ€œNobody tells this to people who are beginners, I wish someone told me. All of us who do creative work, we get into it because we have good taste. But there is this gap. For the first couple years you make stuff, itโ€™s just not that good. Itโ€™s trying to be good, it has potential, but itโ€™s not. But your taste, the thing that got you into the game, is still killer. And your taste is why your work disappoints you. A lot of people never get past this phase, they quit. Most people I know who do interesting, creative work went through years of this. We know our work doesnโ€™t have this special thing that we want it to have. We all go through this. And if you are just starting out or you are still in this phase, you gotta know its normal and the most important thing you can do is do a lot of work. Put yourself on a deadline so that every week you will finish one story. It is only by going through a volume of work that you will close that gap, and your work will be as good as your ambitions. And I took longer to figure out how to do this than anyone Iโ€™ve ever met. Itโ€™s gonna take awhile. Itโ€™s normal to take awhile. Youโ€™ve just gotta fight your way through.โ€

โ€• Ira Glass