https://www.wir1ed.com/story/deepmind-ceo-demis-hassabis-interview-artificial-intelligence-scale/?utm_brand=w1ired&mbid=social_twitter&utm_social-type=owned


์•„๋ž˜๋Š” ์ผ๋ถ€ AI ๋ฒˆ์—ญ



Q

In the last few years, increasing the amount of computer power and data used in training an AI model is the thing that has driven amazing advances. Sam Altman is said to be looking to raise up to $7 trillion for more AI chips. Is vastly more computer power the thing that will unlock artificial general intelligence?

์ง€๋‚œ ๋ช‡ ๋…„ ๋™์•ˆ AI ๋ชจ๋ธ ํ•™์Šต์— ์‚ฌ์šฉ๋˜๋Š” ์ปดํ“จํ„ฐ ์„ฑ๋Šฅ๊ณผ ๋ฐ์ดํ„ฐ์˜ ์–‘์„ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์ด ๋†€๋ผ์šด ๋ฐœ์ „์„ ๊ฐ€์ ธ์˜จ ์›๋™๋ ฅ์ด์—ˆ์Šต๋‹ˆ๋‹ค. ์ƒ˜ ์•ŒํŠธ๋จผ์€ ๋” ๋งŽ์€ AI ์นฉ์„ ์œ„ํ•ด ์ตœ๋Œ€ 7์กฐ ๋‹ฌ๋Ÿฌ๊นŒ์ง€ ๋ชจ๊ธˆํ•  ๊ณ„ํš์ด๋ผ๊ณ  ํ•ฉ๋‹ˆ๋‹ค. ๋” ๋งŽ์€ ์ปดํ“จํ„ฐ ์„ฑ๋Šฅ์ด ์ธ๊ณต ์ผ๋ฐ˜ ์ง€๋Šฅ์„ ์‹คํ˜„ํ•  ์ˆ˜ ์žˆ๋Š” ์—ด์‡ ์ผ๊นŒ์š”?



A

Was that a misquote? I heard someone say that maybe it was yen or something. Well, look, you do need scale; that's why Nvidia is worth what it is today. Thatโ€™s why Sam is trying to raise whatever the real number is. But I think we're a little bit different to a lot of these other organizations in that we've always been fundamental research first. At Google Research and Brain and DeepMind, we've invented the majority of machine learning techniques we're all using today, over the last 10 years of pioneering work. So thatโ€™s always been in our DNA, and we have quite a lot of senior research scientists that maybe other orgs don't have. These other startups and even big companies have a high proportion of engineering to research science.

์ž˜๋ชป ์ธ์šฉํ•œ ๊ฑด๊ฐ€์š”? ๋ˆ„๊ตฐ๊ฐ€ ์—”ํ™” ๊ฐ™์€ ๊ฑฐ๋ผ๊ณ  ๋งํ•˜๋Š” ๊ฑธ ๋“ค์—ˆ์–ด์š”. ํ•˜์ง€๋งŒ ๊ทœ๋ชจ๋ฅผ ํ‚ค์šฐ๋ ค๋ฉด ๊ทœ๋ชจ๋ฅผ ํ‚ค์›Œ์•ผ ํ•˜์ฃ . ๊ทธ๋ž˜์„œ ์˜ค๋Š˜๋‚  ์—”๋น„๋””์•„์˜ ๊ฐ€์น˜๊ฐ€ ์žˆ๋Š” ๊ฒ๋‹ˆ๋‹ค. ๊ทธ๋ž˜์„œ ์ƒ˜์€ ์‹ค์ œ ์ˆซ์ž๊ฐ€ ์–ผ๋งˆ๋“  ์˜ฌ๋ฆฌ๋ ค๊ณ  ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ํ•˜์ง€๋งŒ ์šฐ๋ฆฌ๋Š” ํ•ญ์ƒ ๊ธฐ์ดˆ ์—ฐ๊ตฌ๋ฅผ ์šฐ์„ ์‹œํ•ด ์™”๋‹ค๋Š” ์ ์—์„œ ๋‹ค๋ฅธ ๋งŽ์€ ์กฐ์ง๊ณผ ์กฐ๊ธˆ ๋‹ค๋ฅด๋‹ค๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ๊ตฌ๊ธ€ ๋ฆฌ์„œ์น˜์™€ ๋ธŒ๋ ˆ์ธ, ๋”ฅ๋งˆ์ธ๋“œ์—์„œ๋Š” ์ง€๋‚œ 10๋…„๊ฐ„์˜ ์„ ๊ตฌ์ ์ธ ์—ฐ๊ตฌ๋ฅผ ํ†ตํ•ด ์˜ค๋Š˜๋‚  ์šฐ๋ฆฌ๊ฐ€ ์‚ฌ์šฉํ•˜๊ณ  ์žˆ๋Š” ๋Œ€๋ถ€๋ถ„์˜ ๋จธ์‹ ๋Ÿฌ๋‹ ๊ธฐ์ˆ ์„ ๋ฐœ๋ช…ํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ํ•ญ์ƒ Google์˜ DNA์— ๋‚ด์žฌ๋˜์–ด ์žˆ์œผ๋ฉฐ, ๋‹ค๋ฅธ ์กฐ์ง์—๋Š” ์—†๋Š” ๋งŽ์€ ์„ ์ž„ ์—ฐ๊ตฌ ๊ณผํ•™์ž๋ฅผ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค๋ฅธ ์Šคํƒ€ํŠธ์—…์ด๋‚˜ ๋Œ€๊ธฐ์—…์€ ์—ฐ๊ตฌ ๊ณผํ•™์—์„œ ๊ณตํ•™์ด ์ฐจ์ง€ํ•˜๋Š” ๋น„์ค‘์ด ๋†’์Šต๋‹ˆ๋‹ค.



Q
Does that mean that the competition between AI companies going forward will increasingly be around tool use and agentsโ€”AI that does things rather than just chats? OpenAI is reportedly working on this.
๊ทธ๋ ‡๋‹ค๋ฉด ์•ž์œผ๋กœ AI ๊ธฐ์—… ๊ฐ„์˜ ๊ฒฝ์Ÿ์€ ์ ์  ๋” ๋„๊ตฌ ์‚ฌ์šฉ๊ณผ ์—์ด์ „ํŠธ, ์ฆ‰ ๋‹จ์ˆœํ•œ ์ฑ„ํŒ…์ด ์•„๋‹Œ ์ผ์„ ์ฒ˜๋ฆฌํ•˜๋Š” AI๋ฅผ ๋‘˜๋Ÿฌ์‹ผ ๊ฒฝ์Ÿ์ด ๋ ๊นŒ์š”? OpenAI๊ฐ€ ์ด๋ฅผ ์œ„ํ•ด ๋…ธ๋ ฅํ•˜๊ณ  ์žˆ๋Š” ๊ฒƒ์œผ๋กœ ์•Œ๋ ค์กŒ์Šต๋‹ˆ๋‹ค.

A
Probably. Weโ€™ve been on that track for a long time; thatโ€™s our bread and butter really, agents, reinforcement learning, and planning, since the AlphaGo days. [In 2016 DeepMind developed a breakthrough algorithm capable of solving complex problems and playing sophisticated games.] Weโ€™re dusting off a lot of ideas, thinking of some kind of combination of AlphaGo capabilities built on top of these large models. Introspection and planning capabilities will help with things like hallucination, I think.
It's sort of funny, if you say โ€œTake more careโ€ or โ€œLine out your reasoning,โ€ sometimes the model does better. What's going on there is you are priming it to sort of be a little bit more logical about its steps. But you'd rather that be a systematic thing that the system is doing.
This definitely is a huge area. We're investing a lot of time and energy into that area, and we think that it will be a step change in capabilities of these types of systemsโ€”when they start becoming more agent-like. Weโ€™re investing heavily in that direction, and I imagine others are as well.
์•„๋งˆ๋„์š”. ์ €ํฌ๋Š” ์•ŒํŒŒ๊ณ  ์‹œ์ ˆ๋ถ€ํ„ฐ ์—์ด์ „ํŠธ, ๊ฐ•ํ™” ํ•™์Šต, ๊ณ„ํš ์ˆ˜๋ฆฝ ๋“ฑ ์ด ๋ถ„์•ผ๋ฅผ ์˜ค๋žซ๋™์•ˆ ์—ฐ๊ตฌํ•ด ์™”์Šต๋‹ˆ๋‹ค. [2016๋…„ ๋”ฅ๋งˆ์ธ๋“œ๋Š” ๋ณต์žกํ•œ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ  ์ •๊ตํ•œ ๊ฒŒ์ž„์„ ํ”Œ๋ ˆ์ดํ•  ์ˆ˜ ์žˆ๋Š” ํš๊ธฐ์ ์ธ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ–ˆ์Šต๋‹ˆ๋‹ค.] ์ €ํฌ๋Š” ์ด๋Ÿฌํ•œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ ์œ„์— ๊ตฌ์ถ•๋œ ์•ŒํŒŒ๊ณ  ๊ธฐ๋Šฅ์˜ ์กฐํ•ฉ์„ ์ƒ๊ฐํ•˜๋ฉฐ ๋งŽ์€ ์•„์ด๋””์–ด๋ฅผ ๋– ์˜ฌ๋ฆฌ๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. ์„ฑ์ฐฐ๊ณผ ๊ณ„ํš ๋Šฅ๋ ฅ์ด ํ™˜๊ฐ๊ณผ ๊ฐ™์€ ๊ธฐ๋Šฅ์— ๋„์›€์ด ๋  ๊ฒƒ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.
"์ข€ ๋” ์ฃผ์˜ํ•ด" ๋˜๋Š” "์ถ”๋ก ์„ ์ •๋ฆฌํ•ด"๋ผ๊ณ  ๋งํ•˜๋ฉด ๋ชจ๋ธ์ด ๋” ์ž˜ํ•  ๋•Œ๊ฐ€ ์žˆ์Šต๋‹ˆ๋‹ค. ๊ทธ ๋‹จ๊ณ„์— ๋Œ€ํ•ด ์ข€ ๋” ๋…ผ๋ฆฌ์ ์œผ๋กœ ์„ค๋ช…ํ•  ์ˆ˜ ์žˆ๋„๋ก ์ค€๋น„์‹œํ‚ค๋Š” ๊ฒƒ์ด์ฃ . ํ•˜์ง€๋งŒ ๊ทธ๋ณด๋‹ค๋Š” ์‹œ์Šคํ…œ์ด ์ฒด๊ณ„์ ์œผ๋กœ ์ฒ˜๋ฆฌํ•˜๋Š” ๊ฒƒ์ด ์ข‹์Šต๋‹ˆ๋‹ค.
์ด๊ฒƒ์€ ํ™•์‹คํžˆ ๊ฑฐ๋Œ€ํ•œ ์˜์—ญ์ž…๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ์ด ๋ถ„์•ผ์— ๋งŽ์€ ์‹œ๊ฐ„๊ณผ ์—๋„ˆ์ง€๋ฅผ ํˆฌ์žํ•˜๊ณ  ์žˆ์œผ๋ฉฐ, ์ด๋Ÿฌํ•œ ์œ ํ˜•์˜ ์‹œ์Šคํ…œ์ด ์—์ด์ „ํŠธ์ฒ˜๋Ÿผ ๋˜๊ธฐ ์‹œ์ž‘ํ•˜๋ฉด ๊ทธ ๊ธฐ๋Šฅ์ด ํ•œ ๋‹จ๊ณ„ ๋” ๋ฐœ์ „ํ•  ๊ฒƒ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค. ์ €ํฌ๋Š” ๊ทธ ๋ฐฉํ–ฅ์— ๋งŽ์€ ํˆฌ์ž๋ฅผ ํ•˜๊ณ  ์žˆ์œผ๋ฉฐ ๋‹ค๋ฅธ ๊ธฐ์—…๋“ค๋„ ๋งˆ์ฐฌ๊ฐ€์ง€์ผ ๊ฒƒ์ด๋ผ๊ณ  ์ƒ๊ฐํ•ฉ๋‹ˆ๋‹ค.



์ „์ฒด 3์ค„์š”์•ฝ

์Šค์ผ€์ผ๋งŒ์œผ๋ก  ์•ˆ๋˜๊ณ  ์šฐ๋ฆฌ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜ ๊ฐœ์„ ๋ถ€๋ถ„์—์„œ ๋›ฐ์–ด๋‚˜๋‹ค
์—์ด์ „ํŠธ AI๊ฐ€ ์ƒˆ ๋ฐฉํ–ฅ์„ฑ์ด๊ณ  ์šฐ๋ฆฌ๋Š” ๊ทธ๊ฑฐ ๋‹ค ์ค€๋น„๋ผ์žˆ๋‹ค
์—์ด์ „ํŠธ AI๋Š” ์Šค์Šค๋กœ ์ธํ„ฐ๋„ท๊ณผ ํ˜„์‹ค์„ ์˜ค๊ฐ€๋ฉฐ ํ•™์Šตํ•  ๊ฒƒ์ด๋‹ค (AGI์˜ ๋ฐ”ํƒ•?)