分享我與《人民日報》英文客戶端的深度對談。在對談中，我與記者交流了對人工智能熱點問題的一些看法。

• 以大模型為代表的生成式人工智能將會顛覆式地改變?nèi)祟惖墓ぷ�、溝通、學(xué)習(xí)和娛樂。所有行業(yè)都會被人工智能觸及、改變、轉(zhuǎn)型并提效。

• 中國大模型與世界領(lǐng)先模型之間的差距已經(jīng)縮小至短短幾個月，受益于中國廣闊的市場以及中國團隊的世界頂尖工程能力和落地能力，中國 AI-First 應(yīng)用有望在今年崛起并躋身世界頂尖陣營。2025 年將會是 AI-First 應(yīng)用爆發(fā)的元年，未來新的智算中心相當(dāng)大的比例會專注于推理工作。

• 人際關(guān)系、信任、同理心和愛，是人類區(qū)別于人工智能的本質(zhì)。人工智能會創(chuàng)造出許多全新的工作。而涉及人際溝通和服務(wù)導(dǎo)向的工作，將會是未來許多人可能會投身的領(lǐng)域。

以下是采訪全文：

人民日報：您這兩年的工作重心主要是在哪些方面？

李開復(fù)：過去兩年，我在創(chuàng)新工場持續(xù)關(guān)注人工智能及其他高科技領(lǐng)域投資。此外，我還創(chuàng)辦了AI 2.0大模型獨角獸公司零一萬物，該公司致力于以中國團隊的創(chuàng)新力量，以“多快好省”的方式訓(xùn)出世界第一梯隊性能的大語言模型，賦能千行百業(yè)，驅(qū)動實體經(jīng)濟的新增長范式。

人民日報：您作為AI行業(yè)的代表人物，在近期的公開場合說到了AI 2.0時代開啟，您能跟我們詳細解釋一下什么是AI 2.0時代？為什么說AI 2.0時代開啟？

李開復(fù)：我從事人工智能相關(guān)事業(yè)已有四十多年。多年來人工智能一直在努力模仿部分人類智能，但目前僅實現(xiàn)了人類大腦所具備的通用智能中的一小部分。

在過去兩年間，我們發(fā)現(xiàn)計算機似乎有望具備與人類相同的通用智能。我所說的“通用”，是指計算機能夠像大學(xué)生一樣，全面理解人類知識的方方面面，并且能夠在其他學(xué)科中迅速進行深入學(xué)習(xí)。

大約兩年前，ChatGPT 首次展現(xiàn)出了這種通用智能。此后，在美國和中國涌現(xiàn)出許多具備此類能力的公司。

令人振奮的是，人工智能能夠從人類所撰寫的每一本書籍中汲取知識。不久的將來，人類制作的每一個視頻以及說過的每一句話，都有可能成為人工智能的學(xué)習(xí)素材，進而打造出一個超級智能大腦。這個大腦所學(xué)習(xí)的數(shù)據(jù)量，將遠遠超出任何一個人在其一生中所能掌握的知識總量。

我們現(xiàn)在已經(jīng)開始看到一些跡象，它不僅能進行概括、分析、撰寫內(nèi)容，還能進行推理和演繹，在不需要具體地被教導(dǎo)如何去做的情況下，它就可以解決非常復(fù)雜的數(shù)學(xué)和物理問題。

能夠自主學(xué)習(xí)新事物并自我迭代，大語言模型所展現(xiàn)的這種能力，使我們看到了在未來十年內(nèi)打造出AGI（通用人工智能）的希望，即創(chuàng)造出一種在總體智能水平上超越任何人類的智能體。

人民日報：眾所周知，2024年“新質(zhì)生產(chǎn)力”一度成為熱詞。AI作為新質(zhì)生產(chǎn)力中的一員，您認為當(dāng)下哪些AI應(yīng)用在實現(xiàn)經(jīng)濟效益方面已經(jīng)樹立了很好的例子？給社會生產(chǎn)和生活帶來了哪些影響？

李開復(fù)：我認為“新質(zhì)生產(chǎn)力”這一概念非常有見地。它意味著，生產(chǎn)力的提升并非僅僅依賴于勞動力的增加，而是通過運用新穎獨特且具有突破性的技術(shù)，從而實現(xiàn)價值的成倍乃至指數(shù)級增長。

人工智能不僅是這一概念的良好落地范例，甚至在我看來，它是迄今為止最佳的落地范例。因為人工智能的核心理念就是它能夠像人類一樣思考、推理、做決策、創(chuàng)造內(nèi)容，并完善我們的決策，為我們提供反饋，幫助我們在任何可以想象的領(lǐng)域取得進步。

在傳統(tǒng)行業(yè)中，律師可以借助人工智能完成大部分寫作，其工作效率可提高五到十倍；會計師的工作效率也可以增加五到十倍，因為人工智能可以承擔(dān)所有常規(guī)的數(shù)字計算工作，會計師只需指導(dǎo)人工智能該做什么；客服則有99%可以由機器處理，客戶滿意度更高，人類只需負責(zé)剩余的1%。

這種情況還適用于制造企業(yè)、房地產(chǎn)公司以及所有傳統(tǒng)行業(yè)和服務(wù)行業(yè)。我們可以借助這些新人工智能訓(xùn)練機器人，進而大大降低生產(chǎn)商品的人力成本，這正是我們對新質(zhì)生產(chǎn)力創(chuàng)造新價值的期待所在。更令人振奮的是，在其他被認為能夠增強新質(zhì)生產(chǎn)力的領(lǐng)域，借助人工智能的力量還能實現(xiàn)進一步的雙重提升。

我認為，我們應(yīng)當(dāng)認識到并且接受人工智能是一種超級智能，一種規(guī)模龐大、類似人類大腦卻又不同于人類大腦的存在。這意味著，人工智能可以成為每位知識工作者值得信賴的伙伴。它擁有比人類更大的記憶容量、更快的處理速度以及更全面的知識儲備，但或許在人類直覺、特定經(jīng)驗以及人際交往等方面有所欠缺。

讓人類專注于自己最擅長的事情，讓人工智能發(fā)揮其優(yōu)勢，將直接使我們每個人的效率提升五到十倍。因此，我認為人工智能是發(fā)展新質(zhì)生產(chǎn)力的最強大技術(shù)之一。

人民日報：我們已經(jīng)看到AI這個學(xué)科已經(jīng)在和不同的學(xué)科融合，比如神經(jīng)和認知科學(xué)、心理學(xué)、藝術(shù)繪畫等。展望未來，您覺得生成式AI的應(yīng)用場景有哪些？換言之，AI可以跟哪些產(chǎn)業(yè)行業(yè)融合產(chǎn)生怎樣的可能性？

李開復(fù)：我認為，當(dāng)我們一般性地思考新技術(shù)的到來時就會發(fā)現(xiàn)，比如早期的個人電腦和互聯(lián)網(wǎng)，或者移動設(shè)備和移動互聯(lián)網(wǎng)，以及現(xiàn)在的生成式人工智能——通常當(dāng)我們進入到新技術(shù)革命時，變革往往始于改變我們?yōu)g覽或查看內(nèi)容的方式；隨后，是生產(chǎn)這些新內(nèi)容的方式發(fā)生變化；接著，是搜索、組織和發(fā)現(xiàn)新內(nèi)容的方式得到改進；再之后，是能夠處理更豐富的內(nèi)容形式，比如視頻內(nèi)容；最后，是進行交易和獲取商業(yè)回報的方式發(fā)生變革。

對于生成式人工智能而言，也不例外。上面所說的是從任務(wù)的視角去看技術(shù)將如何演進。同時，我們也可以從人類需求的角度來思考。人類一直都有工作、溝通、學(xué)習(xí)和娛樂的需求，馬斯洛人類需求理論仍然適應(yīng)于AI 2.0時代。從這個角度再次回顧個人電腦和移動設(shè)備的發(fā)展，前兩波技術(shù)浪潮都極大地改變了我們的工作、溝通、學(xué)習(xí)和娛樂方式，因此生成式人工智能也會帶來同樣的改變。

回到你的問題，答案其實是“所有行業(yè)”。你可以回想一下，過去我們是如何溝通的，最初是人與人面對面交流，然后是通過電話，接著是即時通訊、基于互聯(lián)網(wǎng)的通話，隨后是新的社交網(wǎng)絡(luò)出現(xiàn)，現(xiàn)在我認為我們將看到人類與人工智能共同參與的全新溝通方式。

學(xué)習(xí)方式也是如此，最初是在教室里學(xué)習(xí)，然后出現(xiàn)了虛擬教師；工作方式同樣不例外，以你的工作為例，它涉及確定采訪主題、挑選采訪對象、與對方溝通安排采訪、準(zhǔn)備問題、提問、獲取答案、將答案轉(zhuǎn)化為視頻或報紙文章，就像我們現(xiàn)在正做的事。在未來，所有這些步驟都可以逐步實現(xiàn)自動化。因此，我認為所有行業(yè)都被人工智能技術(shù)觸及、改變、轉(zhuǎn)型并提效。

人民日報：國際上我們看到ChatGPT、Sora等生成式人工智能的不斷問世，國內(nèi)我們也有不少生成式人工智能模型，比如零一萬物的Yi系列模型。在您看來，國內(nèi)生成式人工智能產(chǎn)品與國際上ChatGPT、Sora這類是否有較大差距？若有，您認為有哪些差距？

李開復(fù)：是的，毫無疑問這些技術(shù)有一部分是美國人發(fā)明的，但中國人讓它們變得更高效、更實用。我認為這將是根本的區(qū)別。

我在2018年寫了一本書，名為《AI·未來》，我在書中談到了移動互聯(lián)網(wǎng)以及AI 1.0時代。這兩個時代都出現(xiàn)了同樣的情況，美國人發(fā)明了移動互聯(lián)網(wǎng)，他們開發(fā)了最初的移動互聯(lián)APP，但中國的移動互聯(lián)網(wǎng)APP在易用性方面超過美國的APP；在AI 1.0時代，（中國）也出現(xiàn)了“AI 四小龍”，以及許多計算機視覺公司、深度學(xué)習(xí)公司、自動駕駛公司，這些公司可能在創(chuàng)新力上弱于美國公司，但是它們的落地執(zhí)行力卻超過了美國公司。

同樣的情況延續(xù)到了生成式人工智能領(lǐng)域。顯然，兩年前ChatGPT問世時，中國可能落后了七年左右的時間。但在過去兩年中，中國已經(jīng)在快速學(xué)習(xí)并開發(fā)出了很多非常優(yōu)質(zhì)的大語言模型，模型性能非常接近美國頂尖模型，也許還比不上最好的模型，但已經(jīng)相當(dāng)接近。與此同時，中國模型的效率要高得多。

中國工程師確實找到了各種方法來降低成本，提出了新的算法，設(shè)計了新的模型結(jié)構(gòu)，大大加速了模型訓(xùn)練進程的同時，使其能夠在能力較差的芯片上運行，無論是國產(chǎn)還是非國產(chǎn)芯片都適配。訓(xùn)練速度更快，使用起來也就更快。這些中國模型所需的推理時間和推理成本，都比美國模型要小很多。

現(xiàn)在，DeepSeek和零一萬物等中國團隊與美國團隊之間的技術(shù)差距從兩年前的七年縮短到了現(xiàn)在短短幾個月，這是巨大的進步。訓(xùn)練成本降低了十倍甚至更多，推理成本降低了大約三十倍，這些都是由中國公司取得的令人驚嘆的進步，實際上這也讓很多美國頂尖研究人員印象深刻、刮目相看。

但我認為最關(guān)鍵的還在后頭——應(yīng)用領(lǐng)域的全面突圍。復(fù)盤過往的多次技術(shù)浪潮，應(yīng)用層在價值鏈金字塔中創(chuàng)造了最大的經(jīng)濟價值。在技術(shù)領(lǐng)域的競爭中，中國已經(jīng)具備世界頂尖工程能力和落地能力，明顯超過美國的一個方向是構(gòu)建APP，構(gòu)建滿足用戶需求、創(chuàng)造經(jīng)濟價值的應(yīng)用程序。

我們現(xiàn)在正處于這樣一個階段：無論是中國的還是美國的大語言模型，模型性能都非常優(yōu)秀，而且成本很低，尤其是中國的大語言模型，成本更為低廉。這使得那些聰明的APP開發(fā)者可以將精力集中在如何構(gòu)建人工智能APP上，而無需自身成為人工智能專家。我認為現(xiàn)在在中國，AI-First 應(yīng)用百花齊放的土壤已經(jīng)具備，那些在移動互聯(lián)網(wǎng)時代就具備優(yōu)秀APP開發(fā)能力的人，如今已經(jīng)擁有了大展身手的舞臺。

我期待2025年能成為中國 AI-First 應(yīng)用真正崛起并躋身世界頂尖陣營的一年。

人民日報：我們看到國際上有許多對華“脫鉤”的炒作或是論調(diào)，在您看來，如果對華“脫鉤”會對AI發(fā)展造成什么樣的沖擊？您曾表示中國大模型公司要走出不同于OpenAI的第二條路，所謂的“第二條路”是什么？

李開復(fù)：我認為OpenAI所走的第一條道路是，每一年半就多投入十倍以上的資金，訓(xùn)練一個參數(shù)量非常大的模型，并持續(xù)這樣做直到它能為人類所用。這被稱為Scaling Law，但這條道路對中國來說是不可行的。我認為中國更適合的道路是實用主義，注重解決問題、提高效率并創(chuàng)造價值。

正如我之前所描述的第二條道路——中國的工程師們非常擅長找到巧妙的工程解決方案，并真正實現(xiàn)垂直深度整合，讓研究員、工程師、芯片設(shè)計師共同合作，打造出非常高效的產(chǎn)品。

我認為，用一句話來描述第二條道路以及它為何取得了令美國研究人員都驚嘆的驚人成果，那就是：“需求是創(chuàng)新之母”。（Necessity is the mother of innovation.）

“需求”是指，從現(xiàn)實情況來看，我們只有美國1/3至1/50的資源，而且我們無法獲取最先進的芯片，所以我們有什么就用什么，但我們會盡力做到最好。我認為這正是中國公司和中國工程最強的地方。

需求是創(chuàng)新之母。過去，我曾經(jīng)被在北京所遇到的中國研究員身上的勤奮、愿意投身艱苦工作的精神所打動，并一直銘記至今。那是在1990年，這也是我選擇回到中國工作的原因之一。因為我認為，和具備這種職業(yè)道德的人一起，我們能夠創(chuàng)造奇跡，而這正是當(dāng)下生成式 AI 領(lǐng)域正在發(fā)生的事情。

人民日報：為滿足人工智能產(chǎn)業(yè)發(fā)展的需要，全國各地都開始建設(shè)智算中心。與傳統(tǒng)算力中心相比，您認為新一代的智算中心應(yīng)該具備怎樣的特點？

李開復(fù)：智算中心實際上承擔(dān)著兩項任務(wù)。一是幫助構(gòu)建這些模型，通常被稱為訓(xùn)練；二是幫助這些模型投入使用，這被稱為推理。我認為這兩項工作都很重要。

我樂觀地認為未來會出現(xiàn)很多優(yōu)秀的AI-First應(yīng)用程序，考慮到中國龐大的用戶數(shù)量以及我對人工智能大規(guī)模應(yīng)用的樂觀態(tài)度，我會更傾向于在推理而非訓(xùn)練上加大投入。

在過去，訓(xùn)練是智算中心被寄予厚望的主要使用方式，因為當(dāng)時基于生成式人工智能的APP并不多。未來，我樂觀地認為這類APP會越來越多。未來數(shù)據(jù)中心最大概率會被用于推理，因此我認為它們應(yīng)該配備更多推理芯片，并且被合理地部署好，以便能夠更高效地服務(wù)于全中國或至少部分區(qū)域內(nèi)的所有人。

訓(xùn)練智算中心和推理智算中心是不同的。訓(xùn)練智算中心并不側(cè)重于應(yīng)對大規(guī)模用戶使用場景，其核心在于集中大量數(shù)據(jù)并進行持續(xù)數(shù)月的模型訓(xùn)練。而推理智算中心則需要確保任何用戶隨時隨地都能訪問，響應(yīng)速度非常重要，強大的網(wǎng)絡(luò)連接也非常關(guān)鍵。當(dāng)這些新的智算中心建成時，我認為應(yīng)該有相當(dāng)大的比例應(yīng)該專注于推理工作。

人民日報：隱私和安全一直是人工智能領(lǐng)域的關(guān)注焦點，例如人工智能換臉技術(shù)所帶來的風(fēng)險。目前人工智能行業(yè)正在采取哪些措施來解決這些問題？

李開復(fù)：我認為人工智能會有不少風(fēng)險和挑戰(zhàn)，隱私只是其中之一。作為技術(shù)謹慎樂觀派，我認為，但我們不應(yīng)該過度放大這些問題，我相信新技術(shù)產(chǎn)生的問題終究可以被新技術(shù)解決。

面對這些新技術(shù)風(fēng)險，我們將需要對應(yīng)的技術(shù)解決方案，來抓獲深度偽造者，鑒別被深度偽造的視頻或圖片。這些解決方案必須通過技術(shù)手段來開發(fā)。這些技術(shù)還可以更進一步被應(yīng)用于其他場景，如辨別一些內(nèi)容是否為原創(chuàng)內(nèi)容。另一種機制是，在識別圖像時放置一個不可移除的水印，這樣你就可以知道圖片是否被篡改過。這些都是需要進一步研究的技術(shù)。

但還有許多其他擔(dān)憂，比如有人向語言模型詢問“如何制作有害的毒品或武器”怎么辦？我們?nèi)绾畏乐褂腥颂岢鲞@些問題，以及如何防止犯罪分子利用大語言模型來做壞事或制造虛假信息？我認為這些都是另外需要解決的問題。

制定法規(guī)是很有必要的，要明確使用這些技術(shù)從事非法有害行為的人將受到嚴厲懲罰，以此來阻止人們錯誤地使用這些技術(shù)。關(guān)鍵在于深思熟慮如何設(shè)置防護欄，如何通過明確且嚴厲的手段懲罰違法者來形成威懾。

此外，我個人更傾向于使用現(xiàn)有法律法規(guī)并將其擴展到人工智能領(lǐng)域，以非人工智能犯罪的懲罰方式為參考。復(fù)盤過往幾次技術(shù)革命，新技術(shù)的傳播和發(fā)展最終總是利大于弊，因此，限制新技術(shù)的廣泛傳播和發(fā)展并不是一個好主意。很多擔(dān)憂雖然是真實的，但設(shè)置防護欄和法律法規(guī)，應(yīng)該針對具體的非法行為，而不是一刀切地減緩技術(shù)的發(fā)展，因為那將會降低國家的競爭力。

人民日報：人工智能的滲透已成為不可逆轉(zhuǎn)的趨勢，人們在這個過程中可能會感到困擾或焦慮。在你看來，有哪些領(lǐng)域是人工智能無法取代人類的，你有什么建議可以幫助像我這樣的人緩解對人工智能的焦慮？

李開復(fù)：焦慮是正常的，但人工智能的廣泛傳播和持續(xù)快速迭代也是無法阻擋的。首先，我們必須將消極的焦慮轉(zhuǎn)化為積極的自我提升，而不是在焦慮之下無所作為，催生無助感。

未來有許多工作仍然會存在，就像我們看到汽車的出現(xiàn)取代了許多工作，但人類的工作總數(shù)并沒有減少。計算機、移動手機，每一項發(fā)明都取代了一些工作，但新的工作也會隨之而來。

那么哪些工作是比較安全的呢？首先，提升自己、使自己成為人工智能的老師，會是最好的工作。各行各業(yè)最頂尖的工作機會將依然存在，因為總需要有人為人工智能指明方向。

第二類比較安全的人，是那些能夠洞察人類的優(yōu)勢所在，專注于發(fā)揮這些優(yōu)勢，且愿意與人工智能合作的人。人類有一些優(yōu)勢是人工智能所不具備的。其中一點是做真正的顛覆式創(chuàng)新，創(chuàng)造以前不存在的全新概念，因為人工智能是通過數(shù)據(jù)學(xué)習(xí)的。杰出的藝術(shù)家和研究員可以繼續(xù)做出偉大的事業(yè)，這些成果可以被用于教導(dǎo)人工智能。然而，我也承認，這只是一個相對較小的群體。

還有一些可能更契合大眾需求的工作選擇。在我的幾本AI書籍中都曾提到過的最重要的幾點，就是人際關(guān)系、信任、同理心和愛。人工智能沒有情感，它無法與人建立聯(lián)系。所以我認為人們需要普遍關(guān)注這幾項能力：理解他人的能力，獲得信任的能力以及溝通和說服他人的能力。專注于所謂的軟技能，即溝通、同理心、理解、建立聯(lián)系和產(chǎn)生信任的能力，這些是人類獨有的。

在醫(yī)療行業(yè)，未來的醫(yī)生將更多地扮演富有同情心的護理者的角色，而人工智能則在后端負責(zé)確認最佳的藥物組合。醫(yī)生會問診并梳理出問題所在，這些健康問題患者不會想告訴AI，但會告訴一個他認為值得信任的人。其他需要溝通、同理心、聯(lián)系他人的職業(yè)也是如此。我認為，許多涉及人際溝通和服務(wù)導(dǎo)向的工作，將是未來許多人可能會投身的領(lǐng)域。

最后，我相信人工智能將創(chuàng)造許多新工作。今天，AI 已經(jīng)創(chuàng)造了數(shù)千萬個工作崗位，可能你沒有意識到，它被稱為人工智能數(shù)據(jù)標(biāo)注。這個工作崗位可能不會永遠持續(xù)存在，但是類似的新機會將被創(chuàng)造出來。

當(dāng)移動互聯(lián)網(wǎng)誕生時，現(xiàn)在回想起來，它也創(chuàng)造了很多新的工作崗位。線下零售店的店主、農(nóng)民，現(xiàn)在都可以通過APP來對外銷售自己的商品。隨著科技的廣泛應(yīng)用，就業(yè)市場將發(fā)生巨大變化，所創(chuàng)造的工作種類將數(shù)不勝數(shù)。我們現(xiàn)在還不知道它們是什么，但我們可以耐心等待。我敢打賭，現(xiàn)在全球已經(jīng)有數(shù)千萬數(shù)據(jù)標(biāo)注師，但是五年內(nèi)，人工智能會創(chuàng)造十倍于此的新工作機會。

我深信，人類的智慧之光終將指引我們找到未來前行的路。

本文摘編翻譯自《人民日報》英文客戶端專訪，原文如下：

Look forward to 2025 as the year where Chinese AI apps really rise up: Kai-Fu Lee

By Xu Zheqi, Cheng Weidan, Chen Lidan, Liang Peiyu and He Jiahao

Q: What have you been focusing on over the past two years?

Lee: For the last two years, I continued to make investments in AI and other fields on behalf of Sinovation Ventures. I also co-founded a company called 01.AI, which is really building large language models from China, and building models that work well for any language.

Q: As a trailblazer in the AI industry, you recently mentioned that the era of a new generation of AI has begun. Could you explain in detail what this era entails and why you believe it has started?

Lee: Yeah. I've been working on AI for over forty years. And for many years, AI has tried to emulate a little bit of human intelligence. But it only did one little sliver of the entire general brain that our brain has which we call intelligence.

In the last two years, we saw that it appears possible for computers to have that same general intelligence. And when I say general, I mean that it understands everything about human knowledge in a similar way to a college student. And then it can learn further and very quickly in any other discipline.

This general capability was made possible first by Chat-GPT about two years ago. And then both the US and China have seen more companies that have delivered such capabilities.

The excitement is that this capability for AI to learn from every book ever written. In the future soon, every video ever created, and everything ever spoken that it can create a super brain that learns from more data than any human can ever do in a human lifetime.

And we are now seeing glimpses where it can start to do not only generalization, analysis, writing content, but is able to do inference and make deductions, and solve very difficult mathematics problems and physics without ever having been taught to do so specifically.

This ability built on top of the general large language model with an ability to learn new things by itself and teach itself, gives us hope that we'll reach what's called AGI or artificial general intelligence which is overall smarter than any human being within the coming decade.

Q: "New quality productive forces" has been a buzzword in 2024 in China, with AI being a key player. Which AI applications do you think are good examples of achieving economic benefits? How have they impacted industries and people's lives?

Lee: The ideas of new quality productive forces I think are extremely insightful. It's the sense that productivity isn't just putting more labor, but rather using novel and new breakthrough technologies that can multiply or even exponentiate to see greater value being produced.

AI is not only a good example but I think by far, the best example of such technologies because the whole idea of AI is that it can do what humans do, think, reason, make decisions, create contents and refine our decisions, give us feedback, help us improve in any imaginable domain.

In traditional industries, a lawyer can be five or ten times more productive with AI doing much of the writing for the lawyer. An accountant can be five or ten times more productive, because AI does all the routine number crunching, leaving only the accountant to instruct AI what to do; and customer service can be handled 99 percent by machine with a higher level of customer satisfaction, with people only needed for one percent. This goes on, it goes to manufacturing companies, it goes to real estate companies and all traditional industries and service industries. We can have robots that are taught by these new AI, that can dramatically reduce human labor cost for producing goods, all of which lead to the expectations we have for this new quality value creation and productivity creation. And then the most exciting thing is in other areas that are largely viewed as new quality productivity enhancements, AI makes it a double enhancement.

I think the whole idea of thinking of and understanding that AI is all about another super smart, super big, human-like, but different from human brain. That means it's a partner that each of us as a knowledge worker can rely on, a partner who has a much larger memory, much faster processing, and much more complete knowledge. But maybe it lacks our intuition, maybe it lacks our particular kind of experience, maybe it lacks our human-to-human connections.

Keeping humans to do what humans do best and letting AI do what it does best leads directly to each of us, being able to be five or ten times more productive. So that I think is by far the most powerful technology that will lead to this new quality productivity gain.

Q: AI has already been integrated into various disciplines, such as neuroscience, cognitive science, psychology, and the arts. Looking ahead, what new application scenarios do you see for generative AI? What are the potential opportunities for AI to integrate with other industries?

Lee: I think if we think generally about the coming of a new technology, let's say earlier with PC and internet, or mobile and mobile internet, and now with generative AI. Usually when we enter this new technological era, it begins with how we change the way in which we browse or look at content. Then with how we produce this new content, then with how we search and organize and find new content. Then with how we deal with richer forms of content like video, then with transactions and making money.

With generative AI, it will be no exception. This is kind of at a task level, how things will improve. And also, we can think about it as human needs. Humans have always had a need to work, communicate and learn, and entertain ourselves. Again, with PC and mobile we've seen two waves in which these technology waves changed the way we communicate, work, learn and entertain ourselves, so we're going to see AI do the same.

The answer to your question really is everything.

If you think about how did we communicate in the old days, it was person-to-person, then it was through telephone, then it was through instant messaging, then it was through internet-based phone calls, then it's through new social networks. Now I think we're gonna see a brand new way of communicating with humans and AIs together. Also, learning used to be in the classroom, then you could have virtual teachers.

If you think about your job, it's about finding out what topics to have, picking out who to interview, talking to the person about arranging the interview, preparing the questions, asking the questions, getting the answer, turning the answer into a piece of video or a newspaper article, like we're doing right now. In the future, all of this can be one step at a time automated. So I think really there is no industry that will not be touched, changed and transformed and made efficient with AI technologies.

Q: Internationally, we've seen the rise of generative AI applications like Chat-GPT and Sora. In China, there are similar models such as Yi series models. From your perspective, is there a significant gap between domestic generative AI products and international ones? If so, what are the specific differences?

Lee: Yes, there's no doubt that Americans invented most of these technologies, but the Chinese made them more efficient, more usable. I think that will be the fundamental difference.

I wrote a book back in 2018 called AI Superpowers, where I talked about the mobile internet and also the AI 1.0 era. Both of which saw the same thing, that Americans invented the mobile internet. They made the first apps on mobile internet, but the Chinese mobile internet apps beat the American mobile internet apps in usability.

AI 1.0 also, there were "the four dragons," and many computer vision companies, deep learning companies, and autonomous vehicle companies that out-execute the American companies, although American companies generally out-invent the Chinese companies.

The same thing carries over. If we look at generative AI. Clearly one could also say that two years ago when Chat-GPT came out, China was probably easily seven years behind. What has happened in the last two years is that China has learned and developed all these large language models that are very very good, very close to American top models, maybe not quite as good as the best ones, but fairly close. Yet they are so much more efficient.

The Chinese engineers really found all the ways to reduce cost and come up with new algorithms, come up with new model structures, come up with faster training, make it work on lower capability chips whether domestic or not, and really made the training process much faster. When it's faster to train, it's faster to use. Using these models called inference time, compute is also a fraction of the American costs.

We are already seeing that Chinese technologies are around six months behind the US, starting two years ago, seven years behind, now six months behind, huge progress. Cheaper to train by a factor of ten or so or more. Cheaper to infer by a factor of 30 or so. These are amazing progress made by Chinese companies and it's actually made a lot of top American researchers really turn their heads and become very impressed.

But I think the best is yet to come, and the best, the single area where China clearly outshines the US in technologies is in building applications, applications that cater to users' needs and applications that create economic value.

And I think we're now at a stage where the LLMs (large language models) are very good, Chinese or American, and very cheap, in particular Chinese. All these smart application developers, who are not necessarily AI experts, can now turn their attention to how can they build an AI app. I think now hundreds of flowers can blossom in China, with all the people, who have the capability of developing great apps, who have done it in the mobile era. Now the stage is all set for them to enter.

And I look forward to 2025 as the year when Chinese AI apps really rise up and become among the best in the world.

Q: There has been rhetoric about decoupling from China on the international stage. In your opinion, what impact would decoupling have on the AI industry? You once mentioned that Chinese companies need to find a second path distinct from OpenAI. could you elaborate on what you mean by this second path?

Lee: I think the first path taken by OpenAI is every year and a half, spend 10 times more money, train a really big model, and keep going until it beats humans. That has been called the Scaling Law, which is a model not feasible for China.

I think the category that China is in is practical, get things done, make it efficient and make it valuable, so as I described earlier the second path.

The Chinese engineers are so good at finding clever engineering solutions, and doing vertical deep integration to let the researchers, the engineers, and the chip designers work together to make something very efficient.

I think the single sentence that describes the second approach and why it has led to a stunning result that even impressed the American researchers is the following sentence: Necessity is the mother of innovation.

The necessity is the reality that we have one-third to one-fiftieth as much resource, and we don't have access to the most advanced chips. So we have what we can, but let's make do with the best that we can do. This has been I think the strongest point of Chinese companies and Chinese engineering.

Necessities is the mother of innovation. That's why I really remember and I was very moved by the diligence, and willingness for hard work by the Chinese researchers I met in Beijing. This was in 1990, and that's one of the reasons I returned to China to work. Because I felt with people, with such work ethic we're going to make miracles happen, and that's exactly what has happened in generative AI today.

Q: To meet the needs of AI industry development, regional governments have been building and supporting intelligent computing centers. Compared to traditional ones, what characteristics should these new-generation AI computing centers possess?

Lee: The computing centers really perform two tasks. One is helping to make these models, which is usually called training. Secondly, helping these models to become used, which is called inference. And I think both of these are important workloads.

I think knowing how large the number of users are in China and how optimistic I am about the massive adoption of AI and how likely I think there will be so many great AI apps, I would be making a bigger bet on the inference than training.

In the past, training was the primary way that people wanted to see data centers used, because there weren't very many apps. Now that there are more and more apps, I'm optimistic about the future. I think the most likely way, these data centers will be used is through inference, so I think they should be populated with inference chips, and data centers that are well set up to service the people, in all of China, or at least regionally in a very efficient way.

The training data centers and inference data centers are different. A training data center isn't that concerned about massive connectivity, it's more concerned about getting all your data here and just keep training for two months. Inference is about anyone can access anywhere at any time, getting very fast. Response time is very important. Very strong networking is very important.

So, when these data centers are built, I think proportionately a very large number should focus on the inference workflow.

Q: Privacy and security have always been concerns regarding AI, such as risks posed by AI face-swapping technologies. what measures are the AI industry currently taking to address these issues?

Lee: I think privacy is not something that is the only issue with AI. AI actually has many issues, privacy is one of them.

I think those technologies will require technological solutions to catch the deepfake makers, to identify videos or pictures that are deepfaked, and those will have to be developed by technologies. Those technologies can be used even more in computing to find out if something is not original. Another mechanism would be, at the time of capture, placing an irremovable watermark. so that you know when the picture has been altered or not, These are new technologies that need to be invented.

But there are also many other worries, what about people who ask the language model how to make a harmful drug or weapon? How do we prevent those questions from being asked, and also how do you prevent criminals from using large language models to either do something terrible or to create misinformation. I think those are another set of issues that need to be addressed.

I think regulations are definitely needed that make it clear that people who use these technologies for illegal, harmful purposes, will be severely punished as a way to impede people from using these technologies in a wrong way.

What's important is to start thinking about the ways to protect a guardrail, the ways to create deterrence by significant and clear ways to punish offenders. Also, I think the laws and regulations should focus on the way that similar non-AI crimes are committed.

It is not a great idea to limit the proliferation of technologies. Because I think in the end, technology will do a lot of good. A lot of these concerns are real. Guardrails and regulations need to be put in place. They should be done on specific harmful, illegal acts, as opposed to generally slowing down the technologies because that will reduce the country's competitiveness.

Q: The penetration of AI at present has become an irreversible trend and people may be troubled or have anxieties in this process. In your opinion, are there any fields in which AI cannot replace humans, and what suggestions would you give to people like me to adjust AI anxiety?

Lee: Anxiety is normal, but AI’s proliferation and rapid continual improvement is unstoppable. First, we have to turn anxiety into proactive self-improvement, not turn anxiety into inaction and feeling of helplessness.

There are many jobs that will be around, just like we've seen automobiles remove a lot of jobs but total human jobs are not lacking in any means. Computers, mobile phones, every invention has replaced some jobs, but new ones will come around.

What are the types of jobs I think that are secure? First, people who are able to elevate themselves to be the masters of AI that would be the best job. The top jobs are gonna be around because someone needs to give AI direction. The second group would be those who can understand human strengths, focus on those strengths and work with AI. Some of the human strengths that AI doesn't have. I think one is the truly breakthrough innovation, inventing brand new concepts that didn't exist before. AI learns from data. Amazing artists and researchers can also continue to do great work, and in fact use it to teach AI, but I also acknowledge that's a small group of people.

What are some larger groups of people, types of things. I think one of the most important things which I've stated in all of my books is human connection, trust, empathy, and love. AI doesn't have emotion, AI doesn't connect with people. So I think people need to generally focus on their ability to understand other people, gain their trust and ability to communicate and convince other people. Focusing on so-called soft skills, the ability to communicate, empathize, understand, create connections, and render trust, these are uniquely human.

If we think about the medical profession, future physicians will be compassionate caregivers, but AI will be the back end that figures out what's the best drug combination. The physician has teased out what the issues are that maybe the patient wouldn't tell an AI, or wouldn't think to tell an AI, but would tell a trusting human.

And the list continues with all the other professions, that require communication, empathy, and connection. I think a lot of the service-oriented, human-to-human service-oriented jobs, I think will be what the large number of people might do.

And then lastly I'm sure AI will create a lot of jobs. Today AI has already created tens of millions of a category of a job, which you are probably not aware of, it's called labeling data for AI. That may not last forever, but similar opportunities will be created.

When mobile internet was created, now that we look back, it created new categories of jobs. People who have shops, the farmers who can now sell through top apps. The job market will change depending on the adoption of technology, and the types of jobs that are created will be numerous. We don't know what they are, but we can patiently wait. I bet in five years the number of new jobs created by AI will be ten times larger than the tens of millions of data laborers that exist throughout the world.

I have confidence that humans have the wisdom to figure out what to do.

https://peoplesdaily.pdnews.cn/tech/er/30048032628