NVIDIA CEO Jensen Huang explains how the stock market got it wrong on DeepSeek

요약 보기

딥시크 R1 오픈소스 공개로 시장이 'AI 종말'로 오인했으나, 실제로는 모델 효율화로 AI 도입이 가속되는 것이라 주장. 사전학습·사후학습·추론(reasoning) 세 스케일링 법칙 중 추론 연산 수요가 커져 엔비디아에 우호적이라 설명. CUDA 생태계와 NIM 등이 도입을 더 넓힐 것이라 언급. 구체적 가격 수치는 제시되지 않음.

다듬은 글

※ 자동 자막 원문(다듬기 전)

01. it what's really exciting and and you're probably you probably saw um what happened with deep seek mhm uh the the world's first uh reasoning model that's open sourced and it is so incredibly exciting uh the the energy around the world as a result of R1 becoming open- sourced incredible why do people think this could be a bad thing I think it's a wonderful thing well first of all I think from an investor from an investor perspective there was a there was a mental model that um the world was pre-training yeah and then inference yeah and inference was you ask an AI question and it instantly gives you an answer one-hot answer I don't know whose fault it is but obviously that Paradigm is wrong the Paradigm is is um pre-training because we want to have Foundation you need to have a basic level of foundational understanding of information in order to do the second part which is posttraining so pre-training is going to continue to be rigorous lots of data multimodal data as you were mentioning we're going to learn from of course language but language and images video sound and we're going to combine all of that together into our foundational knowledge and that that's going to keep on going the second part of it and this is the most important part actually of intelligence is we call Post trainining Yeah but this is where you learn to solve problems yeah you have foundational information you understand how vocabulary works and syntax works and grammar works and you understand how basic mathematics work and so you take this foundational knowledge you know how to apply it to solve problems M and so we call it uh you here it's post training you could use reinforcement learning human feedback that's another way of saying uh using human demonstration or you could use reinforcement learning which is self-practice you know or could um uh use reinforcement learning uh with another AI a coach right and so so there's a whole bunch of different learning paradigms that are associated with posttraining and in this Paradigm the technology has evolved tremendously in The Last 5 Years and Computing needs is intensive and so people thought that oh my gosh pre-training is a lot less um they forgot that post trainining is really quite intense yeah and then now the third scaling law is the more reasoning that you do the more thinking that you do before you answer a question the better of course there are many things that we've memorized you know how do you what's a square root of 64 me you just kind of memorize that you could reason about it but you don't have to you know and so ideally you've memorized a lot of basic things but they most of the valuable intelligence you still have to reason through and so you have to apply first principles you have to break it down step by step maybe you have to um try a whole lot of different experiment experiments and depending on the output of the results of one experiment it might inform the in you know the next experiment you do and so reasoning um is a fairly compute intensive part of and so I think the market uh responded to R1 as in oh my gosh AI is finished mhh you know it dropped out of the sky we don't need to do any Computing anymore exactly the opposite it's complete opposite because what deeps is doing is is making everybody take notice that okay there are opportunities to have the models be far more efficient than what we thought was possible and so it's expanding and it's accelerating the adoption of AI and so now we have R1 yes we have infinia data intelligence layer yeah yeah the r1s can now be able to talk to intelligent information to solve problems I mean the other thing I think is extremely enabling is the Cuda ecosystem which you clustered and nurtured and and and help really people embar on now with Cuda obj I think it is opening all kinds of possibilities because people cannot tie into this and apply it combination of Cuda obj Nims you know the inference part of it for specific Industries Life Sciences Financial Services autonomous driving and so on and so forth you take all these things you tie them together with the advances that will be made in the models and again it's an acceleration of the adoption of AI which is a wonderful thing for NVIDIA I think