今年1月底，DeepSeek發(fā)布的R1模型對(duì)整個(gè)科技圈造成了巨大轟動(dòng)，英偉達(dá)更是應(yīng)聲下跌16.79%，市值蒸發(fā)5900億美元，創(chuàng)下美國(guó)金融史記錄。‍‍‍‍‍‍‍‍‍

英偉達(dá)發(fā)言人當(dāng)時(shí)表示：“DeepSeek是一項(xiàng)出色的人工智能進(jìn)步，也是測(cè)試時(shí)間縮放的完美例子。”盡管英偉達(dá)已經(jīng)回血，不過(guò)其CEO黃仁勛一直未公開(kāi)回應(yīng)此事。

周四，黃仁勛在一場(chǎng)訪談中首次回應(yīng)了DeepSeek，他表示投資者對(duì)DeepSeek 在人工智能領(lǐng)域取得的進(jìn)展存在誤解，這導(dǎo)致了市場(chǎng)對(duì)英偉達(dá)在股市的錯(cuò)誤反應(yīng)。DeepSeek以低成本高性能引發(fā)關(guān)注后，投資者開(kāi)始質(zhì)疑科技公司投入巨額成本建設(shè)AI基礎(chǔ)設(shè)的必要性。黃仁勛表示，市場(chǎng)的劇烈反應(yīng)源于投資者的誤讀。盡管 R1 的開(kāi)發(fā)似乎減少了對(duì)算力的依賴，但人工智能行業(yè)仍需強(qiáng)大的算力來(lái)支持模型后訓(xùn)練處理方法，這些方法能讓AI模型在后訓(xùn)練進(jìn)行推理或預(yù)測(cè)。“從投資者的角度來(lái)看，他們認(rèn)為世界分為預(yù)訓(xùn)練和推理兩個(gè)階段，而推理就是向 AI 提問(wèn)并立即得到答案。我不知道這種誤解是誰(shuí)造成的，但顯然這種觀念是錯(cuò)誤的。”黃仁勛指出，預(yù)訓(xùn)練仍然重要，但后處理才是“智能最重要的部分”，也是“學(xué)習(xí)解決問(wèn)題的關(guān)鍵環(huán)節(jié)”。此外，黃仁勛還認(rèn)為R1開(kāi)源后，全球范圍內(nèi)展現(xiàn)出的熱情令人難以置信，“這是一件極其令人興奮的事情”。黃仁勛訪談主要環(huán)節(jié)實(shí)錄：

黃仁勛：

What's really exciting and you probably saw，what happened with DeepSeek.

The world's first reasoning model that's open sourced，and it is so incredibly exciting the energy around the world as a result of R1 becoming open sourced，incredible.

真正令人興奮的是，你可能已經(jīng)看到了，DeepSeek發(fā)生了什么。世界上第一個(gè)開(kāi)源的推理模型，這太不可思議了，因?yàn)镽1變成了開(kāi)源的，全球都因此而充滿了能量，真是不可思議。

訪問(wèn)者：

Why do people think this could be a bad thing?I think it's a wonderful thing.

為什么人們認(rèn)為這可能是一件壞事呢?我認(rèn)為這是一件美好的事情。

黃仁勛：

Well,first of all,I think from an investor from an investor perspective,there was a mental model that，the world was pretraining,and then inference.And inference was，you ask an AI question and it instantly gives you an answer，one shot answer.

I don't know whose fault it is,but obviously that paradigm is wrong.The paradigm is 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 post training.So pretraining is continue to be rigorous.

The second part of it and this is the most important part actually of intelligence is we call post training，but this is where you learn to solve problems.You have foundational information.You understand how vocabulary works and syntax work and grammar works,and you understand how basic mathematics work，and so you take this foundational knowledge you now have to apply it to solve problems.

首先，我認(rèn)為從投資者的角度來(lái)看，過(guò)去存在一種思維模型是，世界是預(yù)先訓(xùn)練好的，然后是推理。推理就是你問(wèn)AI一個(gè)問(wèn)題，它立即給你一個(gè)答案，一次性回答。我不知道這是誰(shuí)的錯(cuò)，但顯然這種模式是錯(cuò)誤的。

正確的模式應(yīng)該是先進(jìn)行預(yù)訓(xùn)練，因?yàn)槲覀兿胍幸粋€(gè)基礎(chǔ)，你需要對(duì)信息有一個(gè)基本的理解水平，以便進(jìn)行第二個(gè)部分，也就是后期訓(xùn)練。所以預(yù)訓(xùn)練要繼續(xù)保持嚴(yán)謹(jǐn)。第二部分實(shí)際上是智能最重要的部分，我們稱之為后訓(xùn)練，但這是你學(xué)習(xí)解決問(wèn)題的地方，你已經(jīng)掌握了基礎(chǔ)知識(shí)，你明白詞匯是如何工作的，句法是如何工作的，語(yǔ)法是如何工作的，你明白了基本數(shù)學(xué)是如何工作的，所以你現(xiàn)在必須應(yīng)用這些基礎(chǔ)知識(shí)來(lái)解決實(shí)際問(wèn)題……

So there's a whole bunch of differentlearning paradigms that are associated with post training,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，pretraining is a lot less，they forgot that post training is really quite intense.

因此后訓(xùn)練有一系列很多不同的學(xué)習(xí)模式，在這種模式下，技術(shù)在過(guò)去五年里取得了巨大的進(jìn)步，計(jì)算需求非常大，所以人們認(rèn)為，哦天那，預(yù)訓(xùn)練要少得多。但是他們忘記了后訓(xùn)練其實(shí)相當(dāng)大。

And then now the 3rd scaling law is ，the more reasoning that you do,the more thinking that you do before you answer a question.And so reasoning is a fairly compute intensive part of.And so I think the market responded to R1 as 'oh my gosh AI is finished'，you know it dropped out of the sky ,we don't need to do any computing anymore.It's exactly the opposite.

現(xiàn)在第三條縮放定律是，你做的推理越多，你在回答問(wèn)題之前思考得越多，推理就會(huì)越好，這是一個(gè)計(jì)算量相當(dāng)大的過(guò)程。因此我認(rèn)為市場(chǎng)對(duì)R1的反應(yīng)是“哦我的天哪，AI到頭了"，就好像它從天而降，我們不再需要進(jìn)行任何計(jì)算了，但實(shí)際上完全相反。