您的浏览器禁用了JavaScript(一种计算机语言,用以实现您与网页的交互),请解除该禁用,或者联系我们。[MBI DEEP DIVES&OPENAI]:山姆·奥特曼的解释 - 发现报告

山姆·奥特曼的解释

AI智能总结
查看更多
山姆·奥特曼的解释

JensenHuanginhisrecentJoeRoganpodcasthadaninterestingstoryfromearlyyearsofNvidia: 英伟达创始⼈⻩仁勋在他最近参加的JoeRogan播客中,分享了⼀个关于英伟达早期的有趣故事: “weconvincedourselvesthatchipisgoingtobegreat.AndsoIhadtocallsomeothergentleman.SoIcalledTSMC…andIexplainedtothemwhatweweredoing.AndIexplainedtohim(MorrisChang)Ihadalotofcustomers.Ihadone,youknow,DiamondMultimedia…thedemand'sreallygreat,andwe'regoingtotapeoutachiptoyou,andIliketogodirectlytoproductionbecauseIknowitworks. “我们说服⾃⼰这款芯⽚⼀定会很棒。于是我不得不给另⼀位先⽣打电话。我打给了台积电……我向他们解释了我们在做什么。我也向他(张忠谋)说明我有很多客户。⽐如说,有⼀家⸺你知道的,DiamondMultimedia……需求真的⾮常⼤,我们会把芯⽚交给你进⾏流⽚,⽽且我希望直接进⼊量产,因为我知道它能⾏。” Andtheysaid,"Nobodyhaseverdonethatbefore.Nobodyhasevertapedoutachipthatworkedthefirsttime.Andnobodystartsoutproductionwithoutlookingatit." 他们说:“以前从来没有⼈这么做过。没有⼈第⼀次流⽚就成功的,也没有⼈在看都不看⼀眼的情况下就开始量产。”微信公众号 404K微信公众号 404K ButIknewthatifIdidn'tstarttheproduction,I'dbeoutofbusinessanyways.AndifIcouldstarttheproduction,Imighthaveachance. 但我知道,如果我不开始量产,我反正也会倒闭。⽽如果我能开始量产,我也许还有⼀线机会。 …aswewerestartingtheproduction,MorrisflewtoUnitedStates.Hedidn'tsomanywordsaskedmeso,butheaskedmeawholelotofquestionsthatwastryingto teaseoutdoIhaveanymoneybuthedidn'tdirectlyaskme…sothetruthisthatwedidn'thaveallthemoneybutwehadastrongPOfromthecustomerandifitdidn'tworksomewaferswouldhavebeenlost.I'mnotexactlysurewhatwouldhavehappenedbutwewouldhavecomeshort,itwouldhavebeenrough.” ……就在我们开始量产的时候,莫⾥斯⻜到了美国。他没怎么直接问我,但问了我⼀⼤堆问题,想旁敲侧击地弄清楚我到底有没有钱,但他没有直接问我……事实是,我们并没有所有的资⾦,但我们⼿⾥有⼀份来⾃客户的有⼒采购订单。如果不成功,⼀些晶圆就会报废。我也不太确定事情最后会变成什么样,但我们肯定会资⾦不⾜,那会⾮常艰难。” Historydoesnʼtrepeat,butitdoesrhyme.Twoandhalfdecadeslater,thereisanotherAmericanCEOwhoistryingtoconvinceeveryonethattheyhavealotofdemand!PerhapsthekeydistinctioniswhilebothNvidiaandTSMCbackthenwerehardlyafootnoteinthetechindustry,OpenAItodayisatthefrontandcenterofperhapsthemostconsequentialtechnologicalrevolutioninhistory.Iftheirdemandforecastissubstantiallyoff,thevaluedestructionin“AItrade”canbewholelotlargerthanifNvidiacouldnʼtpayTSMCinmid1990s. 历史不会简单重演,但总是押着相似的韵脚。⼆⼗五年后,⼜有⼀位美国CEO正在努⼒让所有⼈相信他们拥有巨⼤的需求!或许关键的区别在于,当年的Nvidia和TSMC在科技⾏业中⼏乎⽆⾜轻重,⽽如今的OpenAI却站在了也许是⼈类历史上最具决定性技术⾰命的舞台中央。如果他们的需求预测出现重⼤偏差,“AI交易”中造成的价值毁灭,可能远远超过20世纪90年代中期Nvidia⽆法向TSMC付款所带来的后果。微信公众号 404K微信公众号 404K InarecentappearanceonBigTechnologypodcast,SamAltmanwasaskedaboutOpenAIʼs$1.4trillion“commitment”tovariousplayersintheAIvaluechain.Thiswasagoodpodcastwiththoughtful,reflectiveanswersfromSamAltman.Itisworthlisteningtotheentireepisode,butIwillfocusprimarilyonhiscommentsregardinginfrastructurecommitment.Hereʼstheexcerptonthispoint: 在最近⼀期《BigTechnology》播客中,SamAltman被问到OpenAI对AI价值链中各类参与者作出的1.4万亿美元“承诺”。这是⼀档质量很⾼的播客,SamAltman给出了深思熟虑、富有反省意味的回答。整期节⽬都值得⼀听,但我将主要聚焦他关于基础设施承诺的评论。以下是他在这⼀点上的原⽂摘录: “…mylearninginthehistoryofthisfieldisoncethesquigglesstartanditliftsoffthex-axisalittlebit,weknowhowtomakethatbetterandbetter.Butthattakeshugeamountsofcomputetodo.Sothatʼsonearea—throwinglotsofAIatdiscoveringnewscience,curingdisease,lotsofotherthings. “……我从这个领域的历史中学到的是,⼀旦曲线开始起伏、稍微脱离x轴,我们就知道如何把它做得越来越好。但这需要投⼊极其庞⼤的算⼒。因此这是⼀个⽅向⸺投⼊⼤量AI来发现新的科学、治愈疾病,以及做许多其他事情。 Akindofrecent,coolexample:webuilttheSoraAndroidappusingCodex.Theydiditinlessthanamonth.Theyusedahugeamount—oneofthenicethingsaboutworkingatOpenAIisyoudonʼtgetanylimitsonCodex.Theyusedahugeamountoftokens,buttheywereabletodowhatwouldnormallyhavetakenalotofpeoplemuchlonger.AndCodexkindofmostlydiditforus.Andyoucanimaginethatgoingmuchfurther,whereentirecompaniescanbuildtheirproductsusinglotsofcompute.微信公众号 404K微信公众号 404K ⼀个相当新、也很酷的例⼦:我们⽤Codex构建了Sora的安卓应⽤,他们⽤时不到⼀个⽉。他们⽤了⾮常多的资源⸺在OpenAI⼯作的⼀个好处是,使⽤Codex没有任何限制。他们消耗了⼤量的tokens,但完成了通常需要很多⼈、花费更⻓时间才能完成的⼯作。⽽Codex基本上帮我们把这件事做完了。你可以想象,这种模式还可以⾛得更远,未来整个公司都能借助海量算⼒来构建⾃⼰的产品。” Peoplehavetalkedalotaboutvideomodelspointingtowardsthesegenerated,real-timegenerateduserinterfacesthatwilltakealotofcompute.Enterprisesthatwanttotransformtheirbusinesswillusealotofcompute.Doctorsthatwanttooffergood,personalizedhealthcarethatareconstantlymeasuringeverysigntheycangetfromeachindividualpatient—youcanimaginethatusingalotofcompute. ⼈们已经讨论了很多关于视频模型所指向的那些⽣成式、实时⽣成的⽤户界⾯,它们将消耗⼤量算⼒。希望实现业务转型的企业会使⽤⼤量算⼒。想要提供优质、个性化医疗服务的医⽣,会不断地从每⼀位患者身上测量他们所能获得的每⼀个指标⸺你可以想象,这同样会使⽤⼤量算⼒。 ItʼshardtoframehowmuchcomputeweʼrealreadyusingtogenerateAIoutputintheworld,butthesearehorriblyroughnumbers,andIthinkitʼsundisciplinedtotalkthisway,butIalwaysfindthesementalthoughtexperimentsalittlebituseful.Soforgivemeforthesloppiness. 要准确描述我们在全球范围内已经⽤来⽣成AI输出的算⼒规模其实很困难,这些数字都是极其粗略的,⽽且我也认为⽤这种⽅式来讨论并不严谨,但我总觉得这类脑中的思想实验多少有点帮助。所以请原谅我的草率。 LetʼssaythatanAIcompanytodaymightbegeneratingsomethingontheorderof10trilliontokensadayoutoffrontiermodels.More,butitʼsnotlikeaquadrilliontokensforanybody,Idonʼtthink.Letʼssaythereʼs8billionpeopleintheworld,andletʼssayonaverage,theaveragenumberoftokensoutputtedbyapersonperdayislike20,000—theseare,Ithink,totallywrong.Butyoucanthenstart—andtobefair,weʼdhavetocomparetheoutputtokensofamodelprovidertoday,notallthetokensconsumed—butyoucanstarttolookatthis,andyoucansay,weʼregonnahavethesemodelsatacompanybeoutputtingmoretokensperdaythanallofhum