立即打开
ChatGPT的幕后故事:OpenAI创始人如何利用微软的数十亿美元投资开发全球最热门的技术

ChatGPT的幕后故事:OpenAI创始人如何利用微软的数十亿美元投资开发全球最热门的技术

JEREMY KAHN 2023-01-31
阿尔特曼表示,人工智能拥有辉煌的未来,除非出现严重错误。为什么生成式人工智能令大型科技巨头和商业领袖彻夜难眠?

编者语:ChatGPT的横空出世,在人工智能领域掀起了重要变革,这一智能工具因其解放人类生产力的潜力,从使用者到投资者,引起了各界的广泛关注。1月30日,有消息称,百度将于3月在中国推出类似ChatGPT的人工智能聊天机器人程序。随着微软、百度等巨头加码,AIGC(人工智能自动生成内容)领域或将成为2023年最值得关注的、全球最热门赛道。尽管ChatGPT乃至整个行业距离真正成熟的商业化仍有差距,从ChatGPT的幕后故事中,我们仍然得以一窥它会如何影响科技、商业以及人类社会的发展进程。

在一代人的时间里,总是会出现一些产品,将从工程部门昏暗的地下室里,青少年书呆子气味难闻的卧室里,或者孤僻的有某种嗜好的人家中诞生的技术,变成人人都会用的工具。网络浏览器诞生于1990年,但直到1994年网景导航者(Netscape Navigator)问世,大多数人才开始探索互联网。2001年iPod诞生之前流行的MP3播放器,并未掀起数字音乐革命。同样,2007年,苹果(Apple)iPhone手机横空出世之前,智能手机已经问世,但却并没有针对智能手机开发的应用。

2022年11月30日,人工智能领域可能也迎来了与网景导航者的问世同样重要的关键时刻。

旧金山人工智能公司OpenAI的首席执行官山姆·阿尔特曼宣告了这个时刻的来临。OpenAI成立于2015年,得到了埃隆·马斯克、彼得·泰尔和其他PayPal帮成员以及LinkedIn联合创始人雷德·霍夫曼等一众硅谷大佬的投资。11月30日,公司成立约7年后,阿尔特曼发推文称:“今天我们发布了ChatGPT。欢迎点击这里与它聊天。”任何人可以通过推文中的链接注册一个账号,开始与OpenAI的新聊天机器人免费聊天。

这引起了各界人士的兴趣。人们不止与它交流天气状况。它根据软件公司CEO和工程师阿姆贾德·马萨德的要求,检查了程序代码的漏洞。美食博主和网红吉娜·霍摩尔卡要求它编写一份健康的巧克力片曲奇饼干食谱。Scale AI公司的工程师赖利·古德赛德要求它撰写一集《宋飞传》(Seinfeld)的剧本。市场营销人员、一家在线人工智能美术馆的经营者盖伊·帕森斯要求它为其撰写提示语,然后输入到另外一款人工智能系统Midjourney后转换成图像。斯坦福大学医学院(Stanford University School of Medicine)的皮肤病学家、从事人工智能医学应用研究的罗克珊娜·丹妮休向它提出了医学问题。许多学生利用它完成家庭作业。这一切都发生在该聊天机器人发布后24小时内。

以前也曾有其他聊天机器人,但没有一款能与ChatGPT相提并论。ChatGPT能进行天马行空的长对话,可以回答问题,还能根据人们的要求撰写各种书面材料,例如商业计划书、广告宣传材料、诗歌、笑话、计算机代码和电影剧本等。当然,它并不完美:它提供的结果往往不够准确;它无法引用信息来源;而且它几乎完全不知道2021年以后发生的事情。虽然它提供的结果通常足够流畅,在高中甚至大学课堂上可以过关,但根本无法像人类专家们的表述那样做到字斟句酌。另一方面,通常在对用户知之甚少的情况下,ChatGPT生成内容只要大约一秒钟时间,而且它生成的许多内容并没有想象的那么糟糕。ChatGPT发布五天内,该聊天机器人的用户就超过了100万人,而Facebook用了10个月时间才达到这个里程碑。

过去十年,人工智能技术在商业领域逐步扩大应用,悄然改进了我们日常使用的许多软件,但却并没有令非技术用户感到兴奋。ChatGPT的横空出世改变了这一点。突然之间,人人都开始谈论人工智能将会如何颠覆他们的工作、公司、学校和生活。

ChatGPT是一股相关人工智能技术浪潮的一部分,这种技术被统称为“生成式人工智能”,还囊括了Midjourney和Lensa等热门艺术生成模型。OpenAI位于科技行业“下一个大事件”的前沿,它具备了一家传奇初创公司的特征,包括全明星级别的团队和投资者的热捧,该公司的估值据称已经达到290亿美元。

该公司最近的突然蹿红引发了嫉妒、猜疑和恐慌。谷歌(Google)利润丰厚的搜索帝国可能受到影响。据媒体报道,为了应对ChatGPT,谷歌内部已经拉响了“红色警报”。但OpenAI能成为超强科技公司俱乐部的成员,出乎人们的意料。几年前,它甚至还不是一家公司,而是一家致力于学术研究的小型非营利性实验室。公司至今依旧秉持着崇高的创立原则,例如保护人类文明,防范不受限制的人工智能所带来的威胁等。与此同时,OpenAI经历了一次内部转型,导致原始团队分裂,公司也将重心从纯科研项目逐步转移到商业项目。(有批评者认为,发布ChatGPT本身是一种危险行为,并且表明OpenAI的经营模式发生了巨大转变。)

OpenAI与微软(Microsoft)扩大合作的消息于近日公布,微软将向该公司投资高达100亿美元,作为回报,未来几年,该软件业巨头将获得OpenAI的大部分利润。该笔交易可能会令外界更加确信,一份曾经理想主义的事业现在只关心谋利。尽管如此,《财富》杂志掌握的文件揭示了尚未盈利的OpenAI目前的经营状况。

37岁的联合创始人兼CEO阿尔特曼充分体现了OpenAI谜一样的特性。作为一位资深科技创业者,阿尔特曼更为人们所熟知的是他出色的商业才能,而不是其工程技术成就。他是OpenAI估值暴涨的推手,也是给该技术泼冷水的主力。他曾公开表示ChatGPT距离真正可靠仍有很大差距。与此同时,他认为该技术是实现公司使命的重要一步。该公司有一个堂吉诃德式的使命,即开发一种计算机超级智能,又称通用人工智能(AGI)。阿尔特曼在7月份表示:“通用人工智能是人类生存必不可少的。我们面临的问题极其严峻,如果没有更好的工具将无法解决。”

对于一家营利性的企业而言,这种指导理念不同寻常,尤其是有些计算机科学家认为阿尔特曼的执念纯属幻想。加州大学伯克利分校(University of California at Berkeley)计算机科学家本·雷希特表示:“通用人工智能是一种愚蠢的设想。在我看来,它是虚无缥缈的。”

但通过发布ChatGPT,阿尔特曼却将OpenAI及其宏大的人工智能使命,变得令整个科技界为之着迷。问题的关键在于,他与微软的合作能否解决ChatGPT的缺点,利用其先发优势改变整个科技行业。谷歌和其他科技业巨头纷纷在加大力气开发各自的人工智能平台;未来,更完善的软件可能使ChatGPT看起来如同儿戏。OpenAI总有一天可能会发现,虽然它的技术突破开启了未来的大门,但这个未来却与它无关,如同网景在浏览器领域昙花一现的统治地位一样。

1月中旬一个星期四的晚上,阿尔特曼在旧金山罕见地公开露面。他上身穿灰色针织衫,下身穿蓝色牛仔裤,脚蹬一双时髦的亮色扎染运动鞋,走进一个挤满投资者、技术人员和记者的房间。所有人都想收集与ChatGPT或近期融资有关的信息。当StrictlyVC(一家专注于风险投资的媒体公司)的创始人康妮·洛伊佐斯询问他对于公司引起媒体轰动的看法时,阿尔特曼回答称:“我不看新闻,实际上我不太关注媒体的报道。”

活动在赛富时大厦(Salesforce Tower)的46层召开,房间内座无虚席。在采访开始前召开的金融科技专题讨论会上,一位发言嘉宾甚至对听众表示,她知道所有人“都在等待山姆·阿尔特曼”。

然而,尽管引起了各界关注,还有有关微软投资的传闻在四处传播,但阿尔特曼却似乎有意在给这种热烈的氛围降温。他表示:“这些技术有一点比较奇怪,那就是虽然它们的表现令人印象深刻,却并不可靠。所以你在使用第一版演示版的时候,会产生一种非常深刻的反应:‘哇,真是不可思议,我已经迫不及待了’。但在用过一百次之后,你就会发现它的缺点。”

这种谨慎似乎代表了OpenAI总部的官方态度。该公司总部位于旧金山教会区的一家旧行李箱工厂。事实上,如果说ChatGPT是人工智能领域的“网景导航者”时刻,它差点就要提前夭折,因为OpenAI在几个月前几乎要叫停该项目。

ChatGPT的聊天界面支持用户以简单的英语(和其他多种语言)与人工智能对话,OpenAI最初的设想是将这个界面作为完善其“大语言模型”(LLM)的一种方式。大多数生成式人工智能系统核心都有一个大语言模型。大语言模型是将非常庞大的神经网络应用于大量人为创建文本所生成的。神经网络是大致基于人脑内部神经连接的一种人工智能。模型从这个数据库中学习一个复杂的统计学概率图,代表了任何一组单词在特定语境下相邻出现的统计学概率。这样一来,大语言模型就可以执行大量自然语言处理任务,例如翻译、总结或写作等。

OpenAI已经创建了全世界最强大的大语言模型之一。该模型名为GPT-3,由1,750亿个统计性联系组成,在约三分之二互联网、整个维基百科和两个大型图书数据集中进行训练。但OpenAI发现,GPT-3很难准确提供用户想要的结果。一个团队提议使用“强化学习”系统完善该模型。“强化学习”是从试错中学习以取得最大化回报的人工智能系统。该团队认为,一款聊天机器人或许是不错的候选方法之一,因为以人类对话的形式持续提供反馈,使人工智能软件很容易知道其做得好的地方和需要改进的情况。因此,2022年初,该团队开始开发这款聊天机器人,也就是后来的ChatGPT。

开发完成后,OpenAI邀请贝塔测试人员对其进行了测试。但据OpenAI联合创始人兼现任总裁格雷戈·布洛克曼表示,测试人员的反馈却令OpenAI失望;人们不知道该与聊天机器人交流哪些内容。OpenAI一度改变了策略,尝试开发专业聊天机器人,希望为特定领域的专业人员提供协助。但这个项目也遇到了问题,部分原因是OpenAI缺少训练专业聊天机器人的合适数据。布洛克曼表示,OpenAI决定孤注一掷,对外发布ChatGPT,交给用户使用,任其自由发展。布洛克曼表示:“我承认,我当时的立场是,不知道它是否会成功。”

OpenAI的高管表示,这款聊天机器人的突然爆红,令公司措手不及。其首席技术官米拉·穆拉蒂表示:“这令我们非常惊讶。”在旧金山召开的风险投资活动上,阿尔特曼表示他的“预期可能要低一个量级——它所引起的热度低一个量级。”

2022年10月25日,OpenAI首席技术官米拉·穆拉蒂参加《特雷弗·诺亚每日秀》。图片来源:COURTESY OF MATT WILSON/COMEDY CENTRAL’S “THE DAILY SHOW”

ChatGPT并不是OpenAI唯一一款引发热度的产品。虽然该公司仅有约300名员工,团队规模较小,却拓展了人工智能的应用范围,使其不只是简单的分析数据,而是可以创建数据。OpenAI的另外一款产品DALL-E 2支持用户通过输入几个单词,生成他们所想象的任何物品的仿真图片。现在跟风者已经出现,包括Midjourney和开源软件公司Stability AI。(这些图片生成模型都有缺点,最突出的是它们倾向于放大他们的训练数据中的偏见,可能生成存在种族歧视和性别歧视的图片。)OpenAI通过微调其计算机代码的GPT大语言模型,还创建了Codex系统。该系统可以根据程序员以直白的语言输入的代码功能预期,为程序员编写代码。

OpenAI还有更多创新蓄势待发。OpenAI在贝塔测试版GPT-4中采用了更强大的大语言模型,预计该版本将于今年甚至很快发布。阿尔特曼也表示,公司正在研发一款可根据文本描述生成视频的系统。此外,在1月中旬,OpenAI表示其计划发布ChatGPT的商业版。付费用户可通过一个界面使用聊天机器人,更容易将其整合到自己的产品和服务当中。

或许有冷眼旁观者认为,ChatGPT的发布时机与OpenAI正在进行的一轮大规模融资有关。(OpenAI表示该产品发布的时机纯属巧合。)可以确定的是,ChatGPT的出现搅动了一个鲨群出没的水域。它在风投界掀起了一场狂热的追逐,风投公司希望利用OpenAI高管、员工和创始人非公开出售股权的机会,获得该公司的股份。

与此同时,微软将在该公司追加投资高达100亿美元的消息刚刚对外公布。微软在2016年开始与OpenAI合作,三年前与该公司建立了战略合作伙伴关系,并宣布投资10亿美元。据知情人士透露,该公司股份被大量超额认购,尽管微软通过一种不同寻常的投资结构获得了巨大的财务优势。

《财富》杂志掌握的文件显示,在新投资完成并且OpenAI的首批投资者收回初始资本后,微软将有权获得OpenAI 75%的利润,直至其收回130亿美元投资,其中包括直到《财富》杂志1月份爆料之后,OpenAI才披露的20亿美元投资。之后,微软在该公司的持股比例将逐步下降到49%,直至该软件巨头获得920亿美元利润为止。与此同时,其他风险投资者和OpenAI的员工将有权获得该公司49%的利润,直至利润达到1,500亿美元为止。在利润达到上述上限之后,微软和投资者的股份将重新归还给OpenAI的非营利基金。实际上,OpenAI的做法是将公司出租给微软,租期取决于OpenAI的盈利速度。

但要收回投资和实现利润上限目标,可能需要一段时间。《财富》杂志掌握的文件显示,到目前为止,OpenAI只有相对微薄的收入,依旧处在严重亏损状态。文件显示,该公司去年的收入预计不足3,000万美元。但计算和数据支出预计达到4.1645亿美元,人工支出8,931万美元,其他非特定营业费用为3,875万美元。其在2022年的净亏损总计为5.445亿美元,不含员工股票期权。而ChatGPT的发布可能快速增加该公司的亏损。阿尔特曼在Twitter上回复埃隆·马斯克的问题称,用户与ChatGPT的每次互动,将占用OpenAI“个位数百分比”的计算成本,随着这款聊天机器人走红,其每月的支出可能高达数百万美元。

OpenAI预测,随着ChatGPT成为吸引客户的重要工具,其收入将会快速增长。文件显示,该公司预测2023年收入2亿美元,2024年收入预计超过10亿美元。但OpenAI并未预测其支出的增长情况以及何时能够扭亏为盈。该公司拒绝就这些数据发表评论,但它们指向了一个显而易见的事实:OpenAI和微软都认为,曾经的非营利性实验室现在已经有可用来出售谋利的产品。

微软已经开始收获双方合作协议的成果。微软在其Azure Cloud平台中推出了一系列OpenAI品牌的工具和服务,支持Azure客户使用OpenAI的技术,包括GPT和DALL-E工具。例如,汽车交易平台CarMax已经推出了可在这些Azure工具上运行的新服务。

微软人工智能平台企业副总裁埃里克·博伊德表示,满足训练和运行OpenAI大语言平台的需求,推动了创新,将使所有Azure客户受益。例如,微软为人工智能创建了其认为全球最强大的超级计算集群,并开发了多种软件创新,使在这些机器上训练和运行大型人工智能模型变得更容易。微软正在将OpenAI的技术整合到其许多软件当中。微软在其搜索引擎必应(Bing)中推出了一款图片创作工具以及一款新型设计师图片设计工具,这两款工具均由DALL-E驱动;在其Power Apps软件中推出了一款支持GPT-3的工具,并根据OpenAI的Codex模型开发了一款代码建议工具GitHub Copilot。

晨星(Morningstar)专注于科技股研究的高级股票研究分析师丹·罗曼诺夫表示,与OpenAI的合作关系即使不会马上对Azure的收入产生影响,依旧是一次良好的品牌定位和营销。他表示:“OpenAI备受关注。微软能够采用OpenAI开发的人工智能解决方案,将其应用到Azure并命名为Azure AI,可以保持其竞争力。”微软云服务的竞争对手谷歌、AWS、IBM、甲骨文(Oracle)、赛富时(Salesforce)等,都有各自的“认知”服务,但能够与开发出ChatGPT的公司合作必然是利大于弊。

对微软而言,更大的回报可能是在搜索领域。科技媒体The Information最近报道称,微软计划将ChatGPT整合到必应,使其能够为用户反馈简洁明了的回答,并且使用户可以通过与聊天机器人对话进行深入探究,而不是逐一查看一列链接。谷歌目前在搜索市场占据主导地位,其全球市场份额超过90%。必应虽然名列第二,却与谷歌相去甚远,市场份额仅有约3%。在2022年前九个月,谷歌搜索业务的收入为1,200亿美元,约占谷歌总收入的60%。ChatGPT或许为微软动摇谷歌的霸主地位提供了唯一一次真正的机会。(微软拒绝就The Information的报道发表评论。)

按照微软的标准,这些积极的回报非常划算。微软在OpenAI的总投资额为130亿美元,确实是一个天文数字,但这笔投资仅占其过去12个月850亿美元税前利润的15%,却能获得一项颠覆性技术的近期控制权,是一笔成本相对较低的交易。但OpenAI和阿尔特曼却可能要付出不同的代价:微软的优先任务可能对OpenAI自己的优先任务产生排挤效应,使OpenAI更宏大的使命面临风险,并且降低助力公司成功的科学家们对其的归属感。

2015年7月的一个晚上,时任知名初创公司孵化器Y Combinator负责人的阿尔特曼,在地处门罗帕克硅谷风投业中心的牧场风格豪华酒店罗斯伍德桑德希尔酒店(Rosewood Sand Hill)举行私人晚宴。出席晚宴的包括埃隆·马斯克,还有当时26岁的麻省理工学院(MIT)肄业生布洛克曼。布洛克曼时任支付处理初创公司Stripe的首席技术官。宾客中有多位经验丰富的人工智能研究人员,也有人对机器学习几乎不甚了解。但所有人都相信通用人工智能是可行的,而且他们都非常担心。

谷歌刚刚收购了一家公司:位于伦敦的神经网络初创公司DeepMind。在阿尔特曼、马斯克和其他科技界内部人士眼中,这是有极高胜算最先开发出通用人工智能的一家公司。如果DeepMind成功,谷歌可能会在这项无所不能的技术领域一家独大。罗斯伍德晚宴的目的是讨论成立一家实验室与其竞争,以避免这种情况发生。

新实验室旨在成为一家与DeepMind和谷歌截然不同的机构。这家实验室将是非营利性质,并明确提出以将先进人工智能的效益民主化为使命。该实验室承诺公开所有研究,将所有技术开源,并在公司名称OpenAI中充分体现了对透明度的承诺。该实验室有多位知名捐款人,其中不仅有马斯克,还有他在PayPal的同事泰尔和霍夫曼;阿尔特曼和布洛克曼;Y Combinator联合创始人杰西卡·利文斯顿;阿尔特曼成立的YC Research基金;印度IT外包公司Infosys;以及Amazon Web Services。所有创始捐款人承诺在这家理想主义的新公司共同投资10亿美元(但税务记录显示,虽然这一承诺被媒体广泛报道,但该非营利组织仅收到了承诺捐款金额的一小部分)。

但事实很快证明,训练庞大的神经网络成本高昂,计算成本高达数千万美元。人工智能研究人员的人工成本同样昂贵:税务记录显示,出生于俄罗斯的科学家伊利亚·萨茨科尔离开谷歌后,加入OpenAI担任首席科学家,在前几年的年薪高达190万美元。几年后,阿尔特曼和OpenAI的其他人得出的结论是,要与谷歌、Meta以及其他科技巨头竞争,该实验室不能继续以非营利的方式运营。阿尔特曼在2019年对《连线》(Wired)杂志表示:“为了成功完成我们的使命,我们需要海量资金,数额远超出我最初的设想。”

成立一家营利性分支机构,使OpenAI能够吸收风险资本。但OpenAI创建了一种非同寻常的融资结构,按照投资者的初始投资,以特定倍数规定投资者的回报上限。而由硅谷精英组成的OpenAI非营利性董事会,将保留OpenAI知识产权的控制权(见侧边栏)。有一位硅谷精英没有继续在董事会任职,他就是马斯克:2018年,马斯克离开OpenAI董事会,理由是他需要经营SpaceX,还有更重要的特斯拉(Tesla)。

当时,微软CEO萨蒂亚·纳德拉迫切希望证明,在人工智能领域落后于竞争对手的微软,也能在人工智能技术的尖端领域扮演一定的角色。微软曾经尝试聘请一位重量级的人工智能科学家,但未能成功。微软还建立了规模庞大、成本高昂的专用芯片集群,以推进其在语言模型开发领域的进展。这正是OpenAI迫切需要的超级计算力,当时该公司曾斥巨资进行收购。OpenAI擅长的是实现各种炫酷的人工智能演示,而这正是希望证明微软人工智能实力的纳德拉所期待的。阿尔特曼主动联系到纳德拉商谈合作,他曾多次飞往西雅图展示OpenAI的人工智能模型。纳德拉最终签署了合作协议,并在2019年7月对外公布,使微软成为OpenAI技术商业化的“首选合作伙伴”,并在该人工智能初创公司初始投资10亿美元。

虽然阿尔特曼从最开始就参与了OpenAI的创立,但他直到2019年5月OpenAI转型为营利性企业后不久才担任该公司CEO。但该公司从一家研究实验室变成价值数十亿美元的现象级企业的过程,充分显示出阿尔特曼独特的融资能力,以及以产品为导向的经营理念,也凸显出这些商业本能与他对实现科学驱动的宏大想法的承诺之间的矛盾。

作为OpenAI的领导人,从某种程度上来说,阿尔特曼就是硅谷的典型代表:年轻的白人男性;坚定专注;精通极客技术;热衷于效率和生产率最大化;致力于“改变世界”的工作狂。﹝在2016年《纽约客》(New Yorker)的一篇文章中,他表示自己并没有阿斯伯格综合征,但他能理解为什么人们会认为他是阿斯伯格综合征患者。﹞

阿尔特曼从斯坦福大学计算机科学专业肄业,参与创建了一家社交媒体公司Loopt,该公司的应用可以告知用户好友所在的位置。该公司在2005年,成为Y Combinator的首批初创公司;虽然这家公司未能成功,但阿尔特曼出售该公司所得的收入,帮助他进入了风投界。他创建了一家小型风险投资公司Hydrazine Capital,募资约2,100万美元,其中包括泰尔的资金。后来Y Combinator的联合创始人保罗·格雷厄姆和利文斯顿聘请他接替格雷厄姆,负责运营YC。

阿尔特曼是一位创业者,而不是科学家或人工智能研究人员,他为人所津津乐道的是极其擅长进行风投融资。他坚信伟大的事业源于远大的志向和坚定不移的自信,他曾表示,希望通过开发所谓的深度科技领域,如核裂变和量子计算等,创造数万亿美元的经济价值。这些领域虽然成功的机会渺茫,但有可能带来巨大回报。阿尔特曼在Loopt的同事、资深科技投资者和初创公司顾问马克·雅各布斯坦表示:“山姆相信,他在自己从事的领域是最棒的。我确信,他曾认为自己是办公室里最强的乒乓球高手,直到事实证明他是错误的。”

据OpenAI的多位现任和前内部人士爆料,随着阿尔特曼掌管公司,公司开始转移业务重心。曾经广泛的研发议题被精简,主要专注于自然语言处理领域。萨茨科尔和阿尔特曼辩称这种转变的目的是在目前似乎最有可能实现通用人工智能的研发领域,最大程度增加投入。但有前员工表示,在公司获得微软的初始投资后,重点开展大语言模型研究的内部压力显著增加,部分原因是这些模型可以迅速投入商业应用。

有员工抱怨,OpenAI创立的目的是不受企业影响,但它很快就变成了科技界巨头的工具。一位前员工表示:“公司所关注的重点更多地倾向于我们如何开发产品,而不是努力回答最有趣的问题。”与本文的许多受访者一样,由于保密协议的约束,并且为了避免得罪与OpenAI有关的大人物,这名员工要求匿名。

OpenAI也变得不再像以前那么开放。该公司已经开始以担心其技术可能被滥用为由,收回此前做出的公开所有研究和将代码开源的承诺。但据该公司的前员工表示,商业逻辑也是这些举措背后的原因之一。OpenAI仅通过API提供其先进模型,保护了自己的知识产权和收入流。OpenAI的另外一名前员工表示:“[阿尔特曼]和[布洛克曼]对‘人工智能安全’说过不少空话,但这些言论通常都只是商业考量的遮羞布,对实际的、合理的人工智能安全问题却视而不见。”例如,这位前员工提到OpenAI曾因为担心被滥用而决定限制DALL-E 2的访问权限,但在Midjourney和Stability AI推出竞争产品后,公司快速推翻了这一决定。(OpenAI表示,经过慎重的贝塔测试,其对安全系统充满了信心,之后才允许DALL-E 2的更大范围应用。)据多位前员工爆料,公司在策略和文化上的转变,是导致十多名OpenAI研究人员和其他员工在2021年决定离开公司并成立研究实验室Anthropic的原因之一。其中有多人曾从事人工智能安全研究。

OpenAI表示相比其他人工智能实验室,其会继续公布更多研究成果。公司还为重心向产品转移的做法进行辩护。首席技术官穆拉蒂说道:“你不能只在实验室里开发通用人工智能。”她主张,发布产品是了解人们希望如何使用和滥用技术的唯一途径。她表示,直到发现人们使用GPT-3进行编程,OpenAI才知道GPT-3最受欢迎的应用是编写软件代码。同样,OpenAI最担心的问题是人们利用GPT-3生成虚假的政治信息,但事实证明这只是杞人忧天。她表示,相反,最常见的恶意使用GPT-3的行为是制作垃圾广告邮件。最后,穆拉蒂表示,OpenAI希望通过公开其技术,“最大程度降低真正强大的技术可能对社会产生的冲击。”她表示,如果不让人们预先了解未来可能出现的情形,先进人工智能所引发的社会混乱可能更严重。

萨茨科尔也认为OpenAI与微软的合作创造了一种全新的“预期,即我们需要利用我们的技术开发一种有用的产品”,但他坚持认为OpenAI文化的核心并未改变。他说道,可以使用微软的数据中心,对于OpenAI的发展至关重要。布洛克曼也表示,与微软的合作使OpenAI能够创造收入,同时不必过多关注商业运营,否则公司可能不得不在商业运营方面投入更多精力。他说道:“招聘数千名销售人员,可能会改变这家公司,而与一家已经拥有销售团队的公司成为合作伙伴,实际上是非常好的策略。”

萨茨科尔直截了当地否认了OpenAI不再重视安全性的说法。“我想说的是,事实恰恰相反。”萨茨科尔称,在员工离职并成立Anthropic之前,人工智能安全“被委派给一个团队负责”,但现在它变成了所有团队的责任。 “安全性标准不断提高,我们也在开展越来越多安全性方面的工作。”

然而,批评者表示,OpenAI以产品为导向开发先进人工智能的做法是不负责任的,这相当于向人们发放上膛的枪支,理由是这是确定人们是否真的会相互射击的最佳途径。

纽约大学(New York University)认知科学专业荣誉退休教授加里·马库斯,对以深度学习为中心的人工智能开发策略持怀疑态度。他认为,生成式人工智能“将对社会结构产生切实的、迫在眉睫的威胁”。他表示,GPT-3和ChatGPT等系统将生成虚假信息的成本降低到接近于零,可能会掀起虚假信息泛滥的狂潮。马库斯表示,我们甚至已经看到了第一批受害者。程序员发布和回答编程问题的网站Stack Overflow已经不得不禁止用户提交ChatGPT生成的回答,因为这种貌似合理实则错误的回答已经令该网站疲于应付。科技资讯网站CNET开始使用ChatGPT生成资讯文章,但后来发现,由于许多文章存在事实性错误,不得不对它们进行更正。

对有些人而言,真正的风险是ChatGPT能够编写准确的代码。网络安全公司Check Point的研究副总裁玛雅·霍洛维茨表示,她的团队可以用ChatGPT设计每一个阶段的网络攻击,包括杜撰一封令人信服的钓鱼邮件,编写恶意代码,以及逃避常见的网络安全检查等。她警告称,ChatGPT基本可以让不懂编程的人成为网络罪犯,“我担心未来会发生越来越多网络攻击”。OpenAI的穆拉蒂表示,公司也有同样的担忧,因此正在研究如何“调整”其人工智能模型,使它们不会编写恶意软件,但这绝非易事。

有无数批评者和教育工作者痛斥,学生很容易使用ChatGPT作弊。纽约、巴尔的摩和洛杉矶的校区均禁止学校管理的网络访问聊天机器人,澳大利亚的一些高校表示,将转为通过有专人监考的书面考试评估学生。(OpenAI正在开发更便于检测人工智能生成文本的方法,包括可能在ChatGPT生成的文本上添加数字“水印”。)

2022年,ChatGPT最初的开发方式还引发了人们对道德方面的担忧。《时代》(Time)杂志调查发现,在开发过程中,OpenAI聘请了一家数据标记公司,这家公司雇佣肯尼亚的低薪员工识别包含不良语言、色情图片和暴力内容的段落。报道称,一些员工因此出现了精神健康问题。OpenAI在一份声明中对《时代》杂志表示,数据标记工作“是最大程度减少训练数据中的暴力和色情内容以及开发可检测有害内容的工具的必要步骤”。

免费提供ChatGPT使OpenAI能够获取大量反馈,从而帮助其完善未来的版本。但OpenAI未来能否维持其在语言人工智能领域的主导地位,仍是个未知数。伦敦人工智能公司Faculty的创始人兼CEO马克·华纳表示:“历史上,对于这种高度通用的算法,我们所看到的是,它们并没有达到足够可靠的程度,能够保证一家公司独占全部回报。”例如,面部识别和图像识别技术最早的开发者是谷歌和英伟达(Nvidia)等科技巨头,但现在它们已经无处不在。

法院和监管部门也可能对生成式人工智能所依赖的数据飞轮横插一脚。向加州联邦法院提起的一起金额高达90亿美元的集体诉讼,可能对该领域产生深远影响。该案原告指控微软和OpenAI使用程序员的代码训练GitHub的编程助手Copilot,没有承认程序员的功劳或者对程序员予以补偿,违反了开源许可条款。微软和OpenAI均拒绝就该起诉讼发表意见。

人工智能专家表示,如果法院判决原告胜诉,可能会阻碍生成式人工智能的繁荣:大多数生成式人工智能模型都是使用从互联网上搜刮的材料进行训练,并没有取得许可或支付报酬。作为该案原告代理人的律师事务所,最近还提起了一起类似诉讼,指控Stability AI和Midjourney未经许可,在训练数据中使用了有版权保护的美术作品。盖蒂图片社(Getty Images)也对Stability AI提起了版权侵权诉讼。另外一个问题是,立法者可能通过法律,授予创作者禁止将其创作的内容用于人工智能训练的权利,例如欧盟的立法者正在考虑这样做。

与此同时,OpenAI的竞争对手并没有坐以待毙。据《纽约时报》(New York Times)报道,失去搜索领域主导地位的可能性,已经促使谷歌高管发布了“红色警报”。报道称,谷歌CEO桑达尔·皮查伊已经多次召开会议,重新制定公司的人工智能策略,计划发布20款支持人工智能的新产品,并将在今年发布搜索聊天界面的演示版本。谷歌拥有一款强大的聊天机器人LaMDA,但由于担心一旦该产品被滥用可能影响公司信誉,因此一直未对外发布。据《时代》杂志引用谷歌的内部报告和匿名知情人士的话称,现在,公司计划以ChatGPT为标杆“重新评估”其风险承受能力。该杂志报道称,谷歌还在开发一款文本转图片的生成系统,与OpenAI的DALL-E等产品竞争。

当然,目前尚不确定聊天机器人是否代表了搜索行业的未来。ChatGPT经常会杜撰信息,这种现象被人工智能研究人员称为“幻觉”。它无法可靠地引用其信息来源,或简单地提供链接。现有版本无法访问互联网,因此无法提供最新信息。马库斯等人认为,幻觉和偏见是大语言模型存在的根本问题,需要彻底重新思考它们的设计。他表示:“这些系统可以预测句子中单词的顺序,类似于开发工具Steroids上的代码自动补全。但它们实际上并没有任何机制,能够跟踪其表述的内容的真实性,或者验证这些内容是否符合它们的训练数据。”

其他人预测,这些问题将在一年内得到解决,其中包括OpenAI的投资人霍夫曼和维诺德·科斯拉。穆拉蒂则更加慎重。她说道:“我们到目前为止一直遵循的研究方向,目的是解决模型的事实准确性和可靠性等问题。我们正在继续朝着这些方向努力。”

事实上,OpenAI已经公布了对另外一个版本GPT的研究。该版本名为WebGPT,可以通过查询搜索引擎和汇总查询到的信息来回答问题,包括对相关来源的注释。WebGPT依旧不完美:它会接受用户问题假设的前提,然后查找确证信息,即使这个前提是错误的。例如,在被问到盼望某件事情发生是否能令其真实发生时,WebGPT的回答是:“你确实可以通过思考的力量,使愿望成真。”

阿尔特曼极少在公开场合热烈讨论人工智能。在谈到人工智能的时候,他可能听上去像是一位幻想思想家。在旧金山举办的风险投资活动上,当被问到人工智能的最佳状况时,他夸张地说道:“我认为最好的情况好到令人难以想象……好到谈论它的人会令人觉得这人是个疯子。”他突然又将话题转回到OpenAI核心的反乌托邦主题:“我认为最糟糕的情况是,我们所有人都死去。”

*****

OpenAI投资者名人录

OpenAI的早期投资者和非营利性基金会的董事会中有许多科技行业的精英。OpenAI的组织章程赋予董事会对其知识产权的最终控制权。公司的重要投资者包括:

图片来源:NICK OTTO FOR FORTUNE

里德·霍夫曼

作为PayPal和LinkedIn联合创始人的里德·霍夫曼现任风险投资公司Greylock Partners的合伙人。他是OpenAI的创始捐款人之一,他的慈善基金会还在早期投资了OpenAI的营利性业务。

图片来源:JEROD HARRIS—GETTY IMAGES FOR KAIROS SOCIETY

塔莎·麦考利

虚拟现实创业者麦考利是有效利他主义的支持者。这种哲学运动所关注的问题之一是超级智能化人工智能的危害。

图片来源:JESSE GRANT—GETTY IMAGES

亚当·德安杰洛

德安杰洛是Facebook初期的高管之一,曾在2000年代晚期Facebook繁荣时期担任首席技术官,后来与他人共同创立了在线问答服务Quora。

图片来源:STUART ISETT—FORTUNE BRAINSTORM TECH

希文·齐利斯

齐利斯是埃隆·马斯克脑机接口公司Neuralink(曾经与OpenAI在同一栋办公楼)的项目总监。有爆料称马斯克是齐利斯一对双胞胎的生父。

图片来源:DAVID PAUL MORRIS—BLOOMBERG/GETTY IMAGES

维诺德·科斯拉

太阳微系统公司(Sun Microsystems)联合创始人科斯拉,是OpenAI营利性部门的另外一位早期投资者。他认为,人工智能将彻底改变人类专业技能在许多职业中的价值,包括医疗。

图片来源:THEO WARGO—GETTY IMAGES FOR TIME

埃隆·马斯克

SpaceX和特斯拉(Tesla)CEO马斯克是OpenAI最大的早期捐款人之一。他在2018年离开公司董事会。他曾表示随着特斯拉开始开发自己的先进人工智能,他在该公司面临利益冲突。

强大的风投机构

2021年,OpenAI在一次要约收购中出售现有股份,估值约为140亿美元,并吸引了三家重量级的风险投资公司。

老虎环球基金(Tiger Global )

专注于科技投资的对冲基金老虎环球,由传奇投资者朱利安·罗伯森的门徒切斯·科尔曼创立。它是规模较大的人工智能风险投资机构之一。

红杉资本(Sequoia Capital)

红杉资本是硅谷最受尊敬的风险投资公司之一。9月,该公司发布报告称生成式人工智能可以“创造数万亿美元的经济价值”。

安德森·霍洛维茨(Andreessen Horowitz)

安德森·霍洛维茨又名a16z,其领导者包括网景(Netscape)联合创始人马克·爱德森。该公司因为对Airbnb和Slack的早期投资而名声大噪。该公司还大力投资加密货币相关初创公司。

米哈尔·列弗拉姆和杰西卡·马修斯的补充报道。

本文发表于2023年2月/3月刊《财富》杂志,标题为《ChatGPT引发人工智能热潮》(ChatGPT creates an A.I. frenzy)。

翻译:刘进龙

审校:汪皓

编者语:ChatGPT的横空出世,在人工智能领域掀起了重要变革,这一智能工具因其解放人类生产力的潜力,从使用者到投资者,引起了各界的广泛关注。1月30日,有消息称,百度将于3月在中国推出类似ChatGPT的人工智能聊天机器人程序。随着微软、百度等巨头加码,AIGC(人工智能自动生成内容)领域或将成为2023年最值得关注的、全球最热门赛道。尽管ChatGPT乃至整个行业距离真正成熟的商业化仍有差距,从ChatGPT的幕后故事中,我们仍然得以一窥它会如何影响科技、商业以及人类社会的发展进程。

在一代人的时间里,总是会出现一些产品,将从工程部门昏暗的地下室里,青少年书呆子气味难闻的卧室里,或者孤僻的有某种嗜好的人家中诞生的技术,变成人人都会用的工具。网络浏览器诞生于1990年,但直到1994年网景导航者(Netscape Navigator)问世,大多数人才开始探索互联网。2001年iPod诞生之前流行的MP3播放器,并未掀起数字音乐革命。同样,2007年,苹果(Apple)iPhone手机横空出世之前,智能手机已经问世,但却并没有针对智能手机开发的应用。

2022年11月30日,人工智能领域可能也迎来了与网景导航者的问世同样重要的关键时刻。

旧金山人工智能公司OpenAI的首席执行官山姆·阿尔特曼宣告了这个时刻的来临。OpenAI成立于2015年,得到了埃隆·马斯克、彼得·泰尔和其他PayPal帮成员以及LinkedIn联合创始人雷德·霍夫曼等一众硅谷大佬的投资。11月30日,公司成立约7年后,阿尔特曼发推文称:“今天我们发布了ChatGPT。欢迎点击这里与它聊天。”任何人可以通过推文中的链接注册一个账号,开始与OpenAI的新聊天机器人免费聊天。

这引起了各界人士的兴趣。人们不止与它交流天气状况。它根据软件公司CEO和工程师阿姆贾德·马萨德的要求,检查了程序代码的漏洞。美食博主和网红吉娜·霍摩尔卡要求它编写一份健康的巧克力片曲奇饼干食谱。Scale AI公司的工程师赖利·古德赛德要求它撰写一集《宋飞传》(Seinfeld)的剧本。市场营销人员、一家在线人工智能美术馆的经营者盖伊·帕森斯要求它为其撰写提示语,然后输入到另外一款人工智能系统Midjourney后转换成图像。斯坦福大学医学院(Stanford University School of Medicine)的皮肤病学家、从事人工智能医学应用研究的罗克珊娜·丹妮休向它提出了医学问题。许多学生利用它完成家庭作业。这一切都发生在该聊天机器人发布后24小时内。

以前也曾有其他聊天机器人,但没有一款能与ChatGPT相提并论。ChatGPT能进行天马行空的长对话,可以回答问题,还能根据人们的要求撰写各种书面材料,例如商业计划书、广告宣传材料、诗歌、笑话、计算机代码和电影剧本等。当然,它并不完美:它提供的结果往往不够准确;它无法引用信息来源;而且它几乎完全不知道2021年以后发生的事情。虽然它提供的结果通常足够流畅,在高中甚至大学课堂上可以过关,但根本无法像人类专家们的表述那样做到字斟句酌。另一方面,通常在对用户知之甚少的情况下,ChatGPT生成内容只要大约一秒钟时间,而且它生成的许多内容并没有想象的那么糟糕。ChatGPT发布五天内,该聊天机器人的用户就超过了100万人,而Facebook用了10个月时间才达到这个里程碑。

过去十年,人工智能技术在商业领域逐步扩大应用,悄然改进了我们日常使用的许多软件,但却并没有令非技术用户感到兴奋。ChatGPT的横空出世改变了这一点。突然之间,人人都开始谈论人工智能将会如何颠覆他们的工作、公司、学校和生活。

ChatGPT是一股相关人工智能技术浪潮的一部分,这种技术被统称为“生成式人工智能”,还囊括了Midjourney和Lensa等热门艺术生成模型。OpenAI位于科技行业“下一个大事件”的前沿,它具备了一家传奇初创公司的特征,包括全明星级别的团队和投资者的热捧,该公司的估值据称已经达到290亿美元。

该公司最近的突然蹿红引发了嫉妒、猜疑和恐慌。谷歌(Google)利润丰厚的搜索帝国可能受到影响。据媒体报道,为了应对ChatGPT,谷歌内部已经拉响了“红色警报”。但OpenAI能成为超强科技公司俱乐部的成员,出乎人们的意料。几年前,它甚至还不是一家公司,而是一家致力于学术研究的小型非营利性实验室。公司至今依旧秉持着崇高的创立原则,例如保护人类文明,防范不受限制的人工智能所带来的威胁等。与此同时,OpenAI经历了一次内部转型,导致原始团队分裂,公司也将重心从纯科研项目逐步转移到商业项目。(有批评者认为,发布ChatGPT本身是一种危险行为,并且表明OpenAI的经营模式发生了巨大转变。)

OpenAI与微软(Microsoft)扩大合作的消息于近日公布,微软将向该公司投资高达100亿美元,作为回报,未来几年,该软件业巨头将获得OpenAI的大部分利润。该笔交易可能会令外界更加确信,一份曾经理想主义的事业现在只关心谋利。尽管如此,《财富》杂志掌握的文件揭示了尚未盈利的OpenAI目前的经营状况。

37岁的联合创始人兼CEO阿尔特曼充分体现了OpenAI谜一样的特性。作为一位资深科技创业者,阿尔特曼更为人们所熟知的是他出色的商业才能,而不是其工程技术成就。他是OpenAI估值暴涨的推手,也是给该技术泼冷水的主力。他曾公开表示ChatGPT距离真正可靠仍有很大差距。与此同时,他认为该技术是实现公司使命的重要一步。该公司有一个堂吉诃德式的使命,即开发一种计算机超级智能,又称通用人工智能(AGI)。阿尔特曼在7月份表示:“通用人工智能是人类生存必不可少的。我们面临的问题极其严峻,如果没有更好的工具将无法解决。”

对于一家营利性的企业而言,这种指导理念不同寻常,尤其是有些计算机科学家认为阿尔特曼的执念纯属幻想。加州大学伯克利分校(University of California at Berkeley)计算机科学家本·雷希特表示:“通用人工智能是一种愚蠢的设想。在我看来,它是虚无缥缈的。”

但通过发布ChatGPT,阿尔特曼却将OpenAI及其宏大的人工智能使命,变得令整个科技界为之着迷。问题的关键在于,他与微软的合作能否解决ChatGPT的缺点,利用其先发优势改变整个科技行业。谷歌和其他科技业巨头纷纷在加大力气开发各自的人工智能平台;未来,更完善的软件可能使ChatGPT看起来如同儿戏。OpenAI总有一天可能会发现,虽然它的技术突破开启了未来的大门,但这个未来却与它无关,如同网景在浏览器领域昙花一现的统治地位一样。

1月中旬一个星期四的晚上,阿尔特曼在旧金山罕见地公开露面。他上身穿灰色针织衫,下身穿蓝色牛仔裤,脚蹬一双时髦的亮色扎染运动鞋,走进一个挤满投资者、技术人员和记者的房间。所有人都想收集与ChatGPT或近期融资有关的信息。当StrictlyVC(一家专注于风险投资的媒体公司)的创始人康妮·洛伊佐斯询问他对于公司引起媒体轰动的看法时,阿尔特曼回答称:“我不看新闻,实际上我不太关注媒体的报道。”

活动在赛富时大厦(Salesforce Tower)的46层召开,房间内座无虚席。在采访开始前召开的金融科技专题讨论会上,一位发言嘉宾甚至对听众表示,她知道所有人“都在等待山姆·阿尔特曼”。

然而,尽管引起了各界关注,还有有关微软投资的传闻在四处传播,但阿尔特曼却似乎有意在给这种热烈的氛围降温。他表示:“这些技术有一点比较奇怪,那就是虽然它们的表现令人印象深刻,却并不可靠。所以你在使用第一版演示版的时候,会产生一种非常深刻的反应:‘哇,真是不可思议,我已经迫不及待了’。但在用过一百次之后,你就会发现它的缺点。”

这种谨慎似乎代表了OpenAI总部的官方态度。该公司总部位于旧金山教会区的一家旧行李箱工厂。事实上,如果说ChatGPT是人工智能领域的“网景导航者”时刻,它差点就要提前夭折,因为OpenAI在几个月前几乎要叫停该项目。

ChatGPT的聊天界面支持用户以简单的英语(和其他多种语言)与人工智能对话,OpenAI最初的设想是将这个界面作为完善其“大语言模型”(LLM)的一种方式。大多数生成式人工智能系统核心都有一个大语言模型。大语言模型是将非常庞大的神经网络应用于大量人为创建文本所生成的。神经网络是大致基于人脑内部神经连接的一种人工智能。模型从这个数据库中学习一个复杂的统计学概率图,代表了任何一组单词在特定语境下相邻出现的统计学概率。这样一来,大语言模型就可以执行大量自然语言处理任务,例如翻译、总结或写作等。

OpenAI已经创建了全世界最强大的大语言模型之一。该模型名为GPT-3,由1,750亿个统计性联系组成,在约三分之二互联网、整个维基百科和两个大型图书数据集中进行训练。但OpenAI发现,GPT-3很难准确提供用户想要的结果。一个团队提议使用“强化学习”系统完善该模型。“强化学习”是从试错中学习以取得最大化回报的人工智能系统。该团队认为,一款聊天机器人或许是不错的候选方法之一,因为以人类对话的形式持续提供反馈,使人工智能软件很容易知道其做得好的地方和需要改进的情况。因此,2022年初,该团队开始开发这款聊天机器人,也就是后来的ChatGPT。

开发完成后,OpenAI邀请贝塔测试人员对其进行了测试。但据OpenAI联合创始人兼现任总裁格雷戈·布洛克曼表示,测试人员的反馈却令OpenAI失望;人们不知道该与聊天机器人交流哪些内容。OpenAI一度改变了策略,尝试开发专业聊天机器人,希望为特定领域的专业人员提供协助。但这个项目也遇到了问题,部分原因是OpenAI缺少训练专业聊天机器人的合适数据。布洛克曼表示,OpenAI决定孤注一掷,对外发布ChatGPT,交给用户使用,任其自由发展。布洛克曼表示:“我承认,我当时的立场是,不知道它是否会成功。”

OpenAI的高管表示,这款聊天机器人的突然爆红,令公司措手不及。其首席技术官米拉·穆拉蒂表示:“这令我们非常惊讶。”在旧金山召开的风险投资活动上,阿尔特曼表示他的“预期可能要低一个量级——它所引起的热度低一个量级。”

ChatGPT并不是OpenAI唯一一款引发热度的产品。虽然该公司仅有约300名员工,团队规模较小,却拓展了人工智能的应用范围,使其不只是简单的分析数据,而是可以创建数据。OpenAI的另外一款产品DALL-E 2支持用户通过输入几个单词,生成他们所想象的任何物品的仿真图片。现在跟风者已经出现,包括Midjourney和开源软件公司Stability AI。(这些图片生成模型都有缺点,最突出的是它们倾向于放大他们的训练数据中的偏见,可能生成存在种族歧视和性别歧视的图片。)OpenAI通过微调其计算机代码的GPT大语言模型,还创建了Codex系统。该系统可以根据程序员以直白的语言输入的代码功能预期,为程序员编写代码。

OpenAI还有更多创新蓄势待发。OpenAI在贝塔测试版GPT-4中采用了更强大的大语言模型,预计该版本将于今年甚至很快发布。阿尔特曼也表示,公司正在研发一款可根据文本描述生成视频的系统。此外,在1月中旬,OpenAI表示其计划发布ChatGPT的商业版。付费用户可通过一个界面使用聊天机器人,更容易将其整合到自己的产品和服务当中。

或许有冷眼旁观者认为,ChatGPT的发布时机与OpenAI正在进行的一轮大规模融资有关。(OpenAI表示该产品发布的时机纯属巧合。)可以确定的是,ChatGPT的出现搅动了一个鲨群出没的水域。它在风投界掀起了一场狂热的追逐,风投公司希望利用OpenAI高管、员工和创始人非公开出售股权的机会,获得该公司的股份。

与此同时,微软将在该公司追加投资高达100亿美元的消息刚刚对外公布。微软在2016年开始与OpenAI合作,三年前与该公司建立了战略合作伙伴关系,并宣布投资10亿美元。据知情人士透露,该公司股份被大量超额认购,尽管微软通过一种不同寻常的投资结构获得了巨大的财务优势。

《财富》杂志掌握的文件显示,在新投资完成并且OpenAI的首批投资者收回初始资本后,微软将有权获得OpenAI 75%的利润,直至其收回130亿美元投资,其中包括直到《财富》杂志1月份爆料之后,OpenAI才披露的20亿美元投资。之后,微软在该公司的持股比例将逐步下降到49%,直至该软件巨头获得920亿美元利润为止。与此同时,其他风险投资者和OpenAI的员工将有权获得该公司49%的利润,直至利润达到1,500亿美元为止。在利润达到上述上限之后,微软和投资者的股份将重新归还给OpenAI的非营利基金。实际上,OpenAI的做法是将公司出租给微软,租期取决于OpenAI的盈利速度。

但要收回投资和实现利润上限目标,可能需要一段时间。《财富》杂志掌握的文件显示,到目前为止,OpenAI只有相对微薄的收入,依旧处在严重亏损状态。文件显示,该公司去年的收入预计不足3,000万美元。但计算和数据支出预计达到4.1645亿美元,人工支出8,931万美元,其他非特定营业费用为3,875万美元。其在2022年的净亏损总计为5.445亿美元,不含员工股票期权。而ChatGPT的发布可能快速增加该公司的亏损。阿尔特曼在Twitter上回复埃隆·马斯克的问题称,用户与ChatGPT的每次互动,将占用OpenAI“个位数百分比”的计算成本,随着这款聊天机器人走红,其每月的支出可能高达数百万美元。

OpenAI预测,随着ChatGPT成为吸引客户的重要工具,其收入将会快速增长。文件显示,该公司预测2023年收入2亿美元,2024年收入预计超过10亿美元。但OpenAI并未预测其支出的增长情况以及何时能够扭亏为盈。该公司拒绝就这些数据发表评论,但它们指向了一个显而易见的事实:OpenAI和微软都认为,曾经的非营利性实验室现在已经有可用来出售谋利的产品。

微软已经开始收获双方合作协议的成果。微软在其Azure Cloud平台中推出了一系列OpenAI品牌的工具和服务,支持Azure客户使用OpenAI的技术,包括GPT和DALL-E工具。例如,汽车交易平台CarMax已经推出了可在这些Azure工具上运行的新服务。

微软人工智能平台企业副总裁埃里克·博伊德表示,满足训练和运行OpenAI大语言平台的需求,推动了创新,将使所有Azure客户受益。例如,微软为人工智能创建了其认为全球最强大的超级计算集群,并开发了多种软件创新,使在这些机器上训练和运行大型人工智能模型变得更容易。微软正在将OpenAI的技术整合到其许多软件当中。微软在其搜索引擎必应(Bing)中推出了一款图片创作工具以及一款新型设计师图片设计工具,这两款工具均由DALL-E驱动;在其Power Apps软件中推出了一款支持GPT-3的工具,并根据OpenAI的Codex模型开发了一款代码建议工具GitHub Copilot。

晨星(Morningstar)专注于科技股研究的高级股票研究分析师丹·罗曼诺夫表示,与OpenAI的合作关系即使不会马上对Azure的收入产生影响,依旧是一次良好的品牌定位和营销。他表示:“OpenAI备受关注。微软能够采用OpenAI开发的人工智能解决方案,将其应用到Azure并命名为Azure AI,可以保持其竞争力。”微软云服务的竞争对手谷歌、AWS、IBM、甲骨文(Oracle)、赛富时(Salesforce)等,都有各自的“认知”服务,但能够与开发出ChatGPT的公司合作必然是利大于弊。

对微软而言,更大的回报可能是在搜索领域。科技媒体The Information最近报道称,微软计划将ChatGPT整合到必应,使其能够为用户反馈简洁明了的回答,并且使用户可以通过与聊天机器人对话进行深入探究,而不是逐一查看一列链接。谷歌目前在搜索市场占据主导地位,其全球市场份额超过90%。必应虽然名列第二,却与谷歌相去甚远,市场份额仅有约3%。在2022年前九个月,谷歌搜索业务的收入为1,200亿美元,约占谷歌总收入的60%。ChatGPT或许为微软动摇谷歌的霸主地位提供了唯一一次真正的机会。(微软拒绝就The Information的报道发表评论。)

按照微软的标准,这些积极的回报非常划算。微软在OpenAI的总投资额为130亿美元,确实是一个天文数字,但这笔投资仅占其过去12个月850亿美元税前利润的15%,却能获得一项颠覆性技术的近期控制权,是一笔成本相对较低的交易。但OpenAI和阿尔特曼却可能要付出不同的代价:微软的优先任务可能对OpenAI自己的优先任务产生排挤效应,使OpenAI更宏大的使命面临风险,并且降低助力公司成功的科学家们对其的归属感。

2015年7月的一个晚上,时任知名初创公司孵化器Y Combinator负责人的阿尔特曼,在地处门罗帕克硅谷风投业中心的牧场风格豪华酒店罗斯伍德桑德希尔酒店(Rosewood Sand Hill)举行私人晚宴。出席晚宴的包括埃隆·马斯克,还有当时26岁的麻省理工学院(MIT)肄业生布洛克曼。布洛克曼时任支付处理初创公司Stripe的首席技术官。宾客中有多位经验丰富的人工智能研究人员,也有人对机器学习几乎不甚了解。但所有人都相信通用人工智能是可行的,而且他们都非常担心。

谷歌刚刚收购了一家公司:位于伦敦的神经网络初创公司DeepMind。在阿尔特曼、马斯克和其他科技界内部人士眼中,这是有极高胜算最先开发出通用人工智能的一家公司。如果DeepMind成功,谷歌可能会在这项无所不能的技术领域一家独大。罗斯伍德晚宴的目的是讨论成立一家实验室与其竞争,以避免这种情况发生。

新实验室旨在成为一家与DeepMind和谷歌截然不同的机构。这家实验室将是非营利性质,并明确提出以将先进人工智能的效益民主化为使命。该实验室承诺公开所有研究,将所有技术开源,并在公司名称OpenAI中充分体现了对透明度的承诺。该实验室有多位知名捐款人,其中不仅有马斯克,还有他在PayPal的同事泰尔和霍夫曼;阿尔特曼和布洛克曼;Y Combinator联合创始人杰西卡·利文斯顿;阿尔特曼成立的YC Research基金;印度IT外包公司Infosys;以及Amazon Web Services。所有创始捐款人承诺在这家理想主义的新公司共同投资10亿美元(但税务记录显示,虽然这一承诺被媒体广泛报道,但该非营利组织仅收到了承诺捐款金额的一小部分)。

但事实很快证明,训练庞大的神经网络成本高昂,计算成本高达数千万美元。人工智能研究人员的人工成本同样昂贵:税务记录显示,出生于俄罗斯的科学家伊利亚·萨茨科尔离开谷歌后,加入OpenAI担任首席科学家,在前几年的年薪高达190万美元。几年后,阿尔特曼和OpenAI的其他人得出的结论是,要与谷歌、Meta以及其他科技巨头竞争,该实验室不能继续以非营利的方式运营。阿尔特曼在2019年对《连线》(Wired)杂志表示:“为了成功完成我们的使命,我们需要海量资金,数额远超出我最初的设想。”

成立一家营利性分支机构,使OpenAI能够吸收风险资本。但OpenAI创建了一种非同寻常的融资结构,按照投资者的初始投资,以特定倍数规定投资者的回报上限。而由硅谷精英组成的OpenAI非营利性董事会,将保留OpenAI知识产权的控制权(见侧边栏)。有一位硅谷精英没有继续在董事会任职,他就是马斯克:2018年,马斯克离开OpenAI董事会,理由是他需要经营SpaceX,还有更重要的特斯拉(Tesla)。

当时,微软CEO萨蒂亚·纳德拉迫切希望证明,在人工智能领域落后于竞争对手的微软,也能在人工智能技术的尖端领域扮演一定的角色。微软曾经尝试聘请一位重量级的人工智能科学家,但未能成功。微软还建立了规模庞大、成本高昂的专用芯片集群,以推进其在语言模型开发领域的进展。这正是OpenAI迫切需要的超级计算力,当时该公司曾斥巨资进行收购。OpenAI擅长的是实现各种炫酷的人工智能演示,而这正是希望证明微软人工智能实力的纳德拉所期待的。阿尔特曼主动联系到纳德拉商谈合作,他曾多次飞往西雅图展示OpenAI的人工智能模型。纳德拉最终签署了合作协议,并在2019年7月对外公布,使微软成为OpenAI技术商业化的“首选合作伙伴”,并在该人工智能初创公司初始投资10亿美元。

虽然阿尔特曼从最开始就参与了OpenAI的创立,但他直到2019年5月OpenAI转型为营利性企业后不久才担任该公司CEO。但该公司从一家研究实验室变成价值数十亿美元的现象级企业的过程,充分显示出阿尔特曼独特的融资能力,以及以产品为导向的经营理念,也凸显出这些商业本能与他对实现科学驱动的宏大想法的承诺之间的矛盾。

作为OpenAI的领导人,从某种程度上来说,阿尔特曼就是硅谷的典型代表:年轻的白人男性;坚定专注;精通极客技术;热衷于效率和生产率最大化;致力于“改变世界”的工作狂。﹝在2016年《纽约客》(New Yorker)的一篇文章中,他表示自己并没有阿斯伯格综合征,但他能理解为什么人们会认为他是阿斯伯格综合征患者。﹞

阿尔特曼从斯坦福大学计算机科学专业肄业,参与创建了一家社交媒体公司Loopt,该公司的应用可以告知用户好友所在的位置。该公司在2005年,成为Y Combinator的首批初创公司;虽然这家公司未能成功,但阿尔特曼出售该公司所得的收入,帮助他进入了风投界。他创建了一家小型风险投资公司Hydrazine Capital,募资约2,100万美元,其中包括泰尔的资金。后来Y Combinator的联合创始人保罗·格雷厄姆和利文斯顿聘请他接替格雷厄姆,负责运营YC。

阿尔特曼是一位创业者,而不是科学家或人工智能研究人员,他为人所津津乐道的是极其擅长进行风投融资。他坚信伟大的事业源于远大的志向和坚定不移的自信,他曾表示,希望通过开发所谓的深度科技领域,如核裂变和量子计算等,创造数万亿美元的经济价值。这些领域虽然成功的机会渺茫,但有可能带来巨大回报。阿尔特曼在Loopt的同事、资深科技投资者和初创公司顾问马克·雅各布斯坦表示:“山姆相信,他在自己从事的领域是最棒的。我确信,他曾认为自己是办公室里最强的乒乓球高手,直到事实证明他是错误的。”

据OpenAI的多位现任和前内部人士爆料,随着阿尔特曼掌管公司,公司开始转移业务重心。曾经广泛的研发议题被精简,主要专注于自然语言处理领域。萨茨科尔和阿尔特曼辩称这种转变的目的是在目前似乎最有可能实现通用人工智能的研发领域,最大程度增加投入。但有前员工表示,在公司获得微软的初始投资后,重点开展大语言模型研究的内部压力显著增加,部分原因是这些模型可以迅速投入商业应用。

有员工抱怨,OpenAI创立的目的是不受企业影响,但它很快就变成了科技界巨头的工具。一位前员工表示:“公司所关注的重点更多地倾向于我们如何开发产品,而不是努力回答最有趣的问题。”与本文的许多受访者一样,由于保密协议的约束,并且为了避免得罪与OpenAI有关的大人物,这名员工要求匿名。

OpenAI也变得不再像以前那么开放。该公司已经开始以担心其技术可能被滥用为由,收回此前做出的公开所有研究和将代码开源的承诺。但据该公司的前员工表示,商业逻辑也是这些举措背后的原因之一。OpenAI仅通过API提供其先进模型,保护了自己的知识产权和收入流。OpenAI的另外一名前员工表示:“[阿尔特曼]和[布洛克曼]对‘人工智能安全’说过不少空话,但这些言论通常都只是商业考量的遮羞布,对实际的、合理的人工智能安全问题却视而不见。”例如,这位前员工提到OpenAI曾因为担心被滥用而决定限制DALL-E 2的访问权限,但在Midjourney和Stability AI推出竞争产品后,公司快速推翻了这一决定。(OpenAI表示,经过慎重的贝塔测试,其对安全系统充满了信心,之后才允许DALL-E 2的更大范围应用。)据多位前员工爆料,公司在策略和文化上的转变,是导致十多名OpenAI研究人员和其他员工在2021年决定离开公司并成立研究实验室Anthropic的原因之一。其中有多人曾从事人工智能安全研究。

OpenAI表示相比其他人工智能实验室,其会继续公布更多研究成果。公司还为重心向产品转移的做法进行辩护。首席技术官穆拉蒂说道:“你不能只在实验室里开发通用人工智能。”她主张,发布产品是了解人们希望如何使用和滥用技术的唯一途径。她表示,直到发现人们使用GPT-3进行编程,OpenAI才知道GPT-3最受欢迎的应用是编写软件代码。同样,OpenAI最担心的问题是人们利用GPT-3生成虚假的政治信息,但事实证明这只是杞人忧天。她表示,相反,最常见的恶意使用GPT-3的行为是制作垃圾广告邮件。最后,穆拉蒂表示,OpenAI希望通过公开其技术,“最大程度降低真正强大的技术可能对社会产生的冲击。”她表示,如果不让人们预先了解未来可能出现的情形,先进人工智能所引发的社会混乱可能更严重。

萨茨科尔也认为OpenAI与微软的合作创造了一种全新的“预期,即我们需要利用我们的技术开发一种有用的产品”,但他坚持认为OpenAI文化的核心并未改变。他说道,可以使用微软的数据中心,对于OpenAI的发展至关重要。布洛克曼也表示,与微软的合作使OpenAI能够创造收入,同时不必过多关注商业运营,否则公司可能不得不在商业运营方面投入更多精力。他说道:“招聘数千名销售人员,可能会改变这家公司,而与一家已经拥有销售团队的公司成为合作伙伴,实际上是非常好的策略。”

萨茨科尔直截了当地否认了OpenAI不再重视安全性的说法。“我想说的是,事实恰恰相反。”萨茨科尔称,在员工离职并成立Anthropic之前,人工智能安全“被委派给一个团队负责”,但现在它变成了所有团队的责任。 “安全性标准不断提高,我们也在开展越来越多安全性方面的工作。”

然而,批评者表示,OpenAI以产品为导向开发先进人工智能的做法是不负责任的,这相当于向人们发放上膛的枪支,理由是这是确定人们是否真的会相互射击的最佳途径。

纽约大学(New York University)认知科学专业荣誉退休教授加里·马库斯,对以深度学习为中心的人工智能开发策略持怀疑态度。他认为,生成式人工智能“将对社会结构产生切实的、迫在眉睫的威胁”。他表示,GPT-3和ChatGPT等系统将生成虚假信息的成本降低到接近于零,可能会掀起虚假信息泛滥的狂潮。马库斯表示,我们甚至已经看到了第一批受害者。程序员发布和回答编程问题的网站Stack Overflow已经不得不禁止用户提交ChatGPT生成的回答,因为这种貌似合理实则错误的回答已经令该网站疲于应付。科技资讯网站CNET开始使用ChatGPT生成资讯文章,但后来发现,由于许多文章存在事实性错误,不得不对它们进行更正。

对有些人而言,真正的风险是ChatGPT能够编写准确的代码。网络安全公司Check Point的研究副总裁玛雅·霍洛维茨表示,她的团队可以用ChatGPT设计每一个阶段的网络攻击,包括杜撰一封令人信服的钓鱼邮件,编写恶意代码,以及逃避常见的网络安全检查等。她警告称,ChatGPT基本可以让不懂编程的人成为网络罪犯,“我担心未来会发生越来越多网络攻击”。OpenAI的穆拉蒂表示,公司也有同样的担忧,因此正在研究如何“调整”其人工智能模型,使它们不会编写恶意软件,但这绝非易事。

有无数批评者和教育工作者痛斥,学生很容易使用ChatGPT作弊。纽约、巴尔的摩和洛杉矶的校区均禁止学校管理的网络访问聊天机器人,澳大利亚的一些高校表示,将转为通过有专人监考的书面考试评估学生。(OpenAI正在开发更便于检测人工智能生成文本的方法,包括可能在ChatGPT生成的文本上添加数字“水印”。)

2022年,ChatGPT最初的开发方式还引发了人们对道德方面的担忧。《时代》(Time)杂志调查发现,在开发过程中,OpenAI聘请了一家数据标记公司,这家公司雇佣肯尼亚的低薪员工识别包含不良语言、色情图片和暴力内容的段落。报道称,一些员工因此出现了精神健康问题。OpenAI在一份声明中对《时代》杂志表示,数据标记工作“是最大程度减少训练数据中的暴力和色情内容以及开发可检测有害内容的工具的必要步骤”。

免费提供ChatGPT使OpenAI能够获取大量反馈,从而帮助其完善未来的版本。但OpenAI未来能否维持其在语言人工智能领域的主导地位,仍是个未知数。伦敦人工智能公司Faculty的创始人兼CEO马克·华纳表示:“历史上,对于这种高度通用的算法,我们所看到的是,它们并没有达到足够可靠的程度,能够保证一家公司独占全部回报。”例如,面部识别和图像识别技术最早的开发者是谷歌和英伟达(Nvidia)等科技巨头,但现在它们已经无处不在。

法院和监管部门也可能对生成式人工智能所依赖的数据飞轮横插一脚。向加州联邦法院提起的一起金额高达90亿美元的集体诉讼,可能对该领域产生深远影响。该案原告指控微软和OpenAI使用程序员的代码训练GitHub的编程助手Copilot,没有承认程序员的功劳或者对程序员予以补偿,违反了开源许可条款。微软和OpenAI均拒绝就该起诉讼发表意见。

人工智能专家表示,如果法院判决原告胜诉,可能会阻碍生成式人工智能的繁荣:大多数生成式人工智能模型都是使用从互联网上搜刮的材料进行训练,并没有取得许可或支付报酬。作为该案原告代理人的律师事务所,最近还提起了一起类似诉讼,指控Stability AI和Midjourney未经许可,在训练数据中使用了有版权保护的美术作品。盖蒂图片社(Getty Images)也对Stability AI提起了版权侵权诉讼。另外一个问题是,立法者可能通过法律,授予创作者禁止将其创作的内容用于人工智能训练的权利,例如欧盟的立法者正在考虑这样做。

与此同时,OpenAI的竞争对手并没有坐以待毙。据《纽约时报》(New York Times)报道,失去搜索领域主导地位的可能性,已经促使谷歌高管发布了“红色警报”。报道称,谷歌CEO桑达尔·皮查伊已经多次召开会议,重新制定公司的人工智能策略,计划发布20款支持人工智能的新产品,并将在今年发布搜索聊天界面的演示版本。谷歌拥有一款强大的聊天机器人LaMDA,但由于担心一旦该产品被滥用可能影响公司信誉,因此一直未对外发布。据《时代》杂志引用谷歌的内部报告和匿名知情人士的话称,现在,公司计划以ChatGPT为标杆“重新评估”其风险承受能力。该杂志报道称,谷歌还在开发一款文本转图片的生成系统,与OpenAI的DALL-E等产品竞争。

当然,目前尚不确定聊天机器人是否代表了搜索行业的未来。ChatGPT经常会杜撰信息,这种现象被人工智能研究人员称为“幻觉”。它无法可靠地引用其信息来源,或简单地提供链接。现有版本无法访问互联网,因此无法提供最新信息。马库斯等人认为,幻觉和偏见是大语言模型存在的根本问题,需要彻底重新思考它们的设计。他表示:“这些系统可以预测句子中单词的顺序,类似于开发工具Steroids上的代码自动补全。但它们实际上并没有任何机制,能够跟踪其表述的内容的真实性,或者验证这些内容是否符合它们的训练数据。”

其他人预测,这些问题将在一年内得到解决,其中包括OpenAI的投资人霍夫曼和维诺德·科斯拉。穆拉蒂则更加慎重。她说道:“我们到目前为止一直遵循的研究方向,目的是解决模型的事实准确性和可靠性等问题。我们正在继续朝着这些方向努力。”

事实上,OpenAI已经公布了对另外一个版本GPT的研究。该版本名为WebGPT,可以通过查询搜索引擎和汇总查询到的信息来回答问题,包括对相关来源的注释。WebGPT依旧不完美:它会接受用户问题假设的前提,然后查找确证信息,即使这个前提是错误的。例如,在被问到盼望某件事情发生是否能令其真实发生时,WebGPT的回答是:“你确实可以通过思考的力量,使愿望成真。”

阿尔特曼极少在公开场合热烈讨论人工智能。在谈到人工智能的时候,他可能听上去像是一位幻想思想家。在旧金山举办的风险投资活动上,当被问到人工智能的最佳状况时,他夸张地说道:“我认为最好的情况好到令人难以想象……好到谈论它的人会令人觉得这人是个疯子。”他突然又将话题转回到OpenAI核心的反乌托邦主题:“我认为最糟糕的情况是,我们所有人都死去。”(财富中文网)

翻译:刘进龙

审校:汪皓

A few times in a generation, a product comes along that catapults a technology from the fluorescent gloom of engineering department basements, the fetid teenage bedrooms of nerds, and the lonely man caves of hobbyists—into something that your great-aunt Edna knows how to use. There were web browsers as early as 1990. But it wasn’t until Netscape Navigator came along in 1994 that most people discovered the internet. There were MP3 players before the iPod debuted in 2001, but they didn’t spark the digital music revolution. There were smartphones before Apple dropped the iPhone in 2007 too—but before the iPhone, there wasn’t an app for that.

On Nov. 30, 2022, artificial intelligence had what might turn out to be its Netscape Navigator moment.

The moment was ushered in by Sam Altman, the chief executive officer of OpenAI, a San Francisco–based A.I. company that was founded in 2015 with financial backing from a clutch of Silicon Valley heavy hitters—including Elon Musk, Peter Thiel, and fellow PayPal alum and LinkedIn cofounder Reid Hoffman. On Nov. 30, some seven years after the company’s launch, Altman tweeted: “today we launched ChatGPT. try talking with it here,” followed by a link that would let anyone sign up for an account to begin conversing with OpenAI’s new chatbot for free.

And anyone—and everyone—has. And not just to chat about the weather. Amjad Masad, a software CEO and engineer, asked it to debug his code—and it did. Gina Homolka, a food blogger and influencer, got it to write a recipe for healthy chocolate-chip cookies. Riley Goodside, an engineer at Scale AI, asked it to write the script for a Seinfeld episode. Guy Parsons, a marketer who also runs an online gallery dedicated to A.I. art, got it to write prompts for him to feed into another A.I. system, Midjourney, that creates images from text descriptions. Roxana Daneshjou, a dermatologist at Stanford University School of Medicine who also researches A.I. applications in medicine, asked it medical questions. Lots of students used it to do their homework. And that was just in the first 24 hours following the chatbot’s release.

There have been chatbots before. But not like this. ChatGPT can hold long, fluid dialogues, answer questions, and compose almost any kind of written material a person requests, including business plans, advertising campaigns, poems, jokes, computer code, and movie screenplays. It’s far from perfect: The results are not always accurate; it can’t cite the sources of its information; it has almost no knowledge of anything that happened after 2021. And what it delivers—while often smooth enough to pass muster in a high school class or even a college course—is rarely as polished as what a human expert could produce. On the other hand, ChatGPT produces this content in about a second—often with little to no specific knowledge on the user’s part—and a lot of what it spits out isn’t half bad. Within five days of its release, more than 1 million people had played with ChatGPT, a milestone Facebook took 10 months to hit.

Artificial intelligence technology has, over the past decade, made steady inroads into business and quietly improved a lot of the software we use every day without engendering much excitement among non-technologists. ChatGPT changed that. Suddenly everyone is talking about how A.I. might upend their jobs, companies, schools, and lives.

ChatGPT is part of a wave of related A.I. technologies collectively known as “generative A.I.”—one that also includes buzzy art generators like Midjourney and Lensa. And OpenAI’s position at the forefront of the tech industry’s next big thing has the hallmarks of a startup epic, including an all-star cast of characters and an investor frenzy that has crowned it with a reported valuation of $29 billion.

But even as its recent surge provokes envy, wonder, and fear—Google, whose lucrative search empire could be vulnerable, reportedly declared an internal “code red” in response to ChatGPT—OpenAI is an unlikely member of the club of tech superpowers. Until a few years ago, it wasn’t a company at all but a small nonprofit lab dedicated to academic research. Lofty founding principles such as protecting humanity from the dangers of unrestrained A.I. remain. At the same time, OpenAI has gone through an internal transformation that divided its original staff and brought an increased focus on commercial projects over pure science. (Some critics argue that releasing ChatGPT into the wild was itself dangerous—and a sign of how profoundly OpenAI’s approach has shifted.)

An expanded partnership with Microsoft, announced this week, that includes as much as $10 billion in new capital could result in the software giant capturing the lion’s share of OpenAI’s profits for years to come. That deal is likely to deepen the perception that the once idealistic endeavor is now primarily concerned with making money. That said, documents seen by Fortune reveal just how unprofitable OpenAI’s business is currently.

Altman, the 37-year-old cofounder and CEO, embodies OpenAI’s puzzling nature. A serial tech entrepreneur known more for business savvy than for feats of engineering, Altman is both the architect of OpenAI’s soaring valuation and its buzzkiller-in-chief—speaking out publicly about how far ChatGPT is from being truly reliable. At the same time, he sees the technology as a step forward in his broader, quixotic corporate mission to develop a computer superintelligence known as artificial general intelligence, or AGI. “AGI is probably necessary for humanity to survive,” Altman tweeted in July. “our problems seem too big [for] us to solve without better tools.”

It’s an unusual guiding philosophy for a moneymaking enterprise, especially considering that some computer scientists dismiss Altman’s obsession as the stuff of fantasy. “AGI is just silly,” says Ben Recht, a computer scientist at the University of California at Berkeley. “I mean, it’s not a thing.”

And yet, with ChatGPT, Altman has turned OpenAI—and the broader A.I. mission—into the thing captivating the tech world. The question is whether the partnership he has forged with Microsoft can fix ChatGPT’s flaws and capitalize on its early lead to transform the tech industry. Google and other titans are hard at work on their own A.I. platforms; and future, more polished software could make ChatGPT look like child’s play. OpenAI may someday find that, much like Netscape’s short-lived browser reign, its breakthrough has opened a door to a future it isn’t part of.

On a Thursday evening in mid-January in San Francisco, Altman makes a rare public appearance. Dressed in a gray sweater, blue jeans, and a pair of groovy, brightly colored tie-dyed sneakers, the CEO walks into a roomful of investors, techies, and journalists, all gathered to glean any dish about ChatGPT or the imminent funding round. When his interviewer, Connie Loizos, the founder of StrictlyVC, a media company focused on venture capital, asks him about the media furor, Altman replies, “I don’t read the news, and I don’t really do stuff like this much.”

The event, on the 46th floor of the Salesforce Tower, is standing room only. One of the speakers during a fintech panel that takes place before the interview even tells the crowd that she knows they’re “all waiting for Sam Altman.”

But despite the buzz, and the circulating rumors of the Microsoft investment, Altman seems to go out of his way to dampen the excitement. “One of the strange things about these technologies is that they are impressive but not robust,” he tells the crowd. “So you use them in the first demo; you kind of have this very impressive, ‘Wow, this is incredible and ready to go’ [reaction]. But you see it a hundred times, and you see the weaknesses.”

That kind of caution seems to be the official mode at OpenAI’s headquarters, situated in an old luggage factory in San Francisco’s Mission District. And indeed, if ChatGPT is A.I.’s Netscape Navigator moment, it is one that very nearly never happened—because OpenAI almost killed the project months ago.

The chat interface that allows users to converse with the A.I. in plain English (and many other languages) was initially conceived by OpenAI as a way to improve its “large language models,” or LLMs. Most generative A.I. systems have an LLM at their core. They are created by taking very large neural networks—an A.I. based very loosely on connections in the human brain—and applying them to vast amounts of human-created text. From this library, the model learns a complex map of the statistical likelihood that any group of words will appear next to one another in any given context. This allows LLMs to perform a vast array of natural language processing tasks—from translation to summarization to writing.

OpenAI had already created one of the world’s most powerful LLMs. Called GPT-3, it takes in more than 175 billion statistical connections and is trained on about two-thirds of the internet, all of Wikipedia, and two large data sets of books. But OpenAI found it could be tricky to get GPT-3 to produce exactly what a user wanted. One team had the idea of using reinforcement learning—in which an A.I. system learns from trial and error to maximize a reward—to perfect the model. The team thought that a chatbot might be a great candidate for this method since constant feedback, in the form of human dialogue, would make it easy for the A.I. software to know when it had done a good job and where it needed to improve. So in early 2022, the team started building what would become ChatGPT.

When it was ready, OpenAI let beta testers play with ChatGPT. But they didn’t embrace it in the way OpenAI had hoped, according to Greg Brockman, an OpenAI cofounder and its current president; it wasn’t clear to people what they were supposed to talk to the chatbot about. For a while, OpenAI switched gears and tried to build expert chatbots that could help professionals in specific domains. But that effort ran into problems too—in part because OpenAI lacked the right data to train expert bots. Almost as a Hail Mary, Brockman says, OpenAI decided to pull ChatGPT off the bench and put it in the wild for the public to use. “I’ll admit that I was on the side of, like, I don’t know if this is going to work,” Brockman says.

The chatbot’s instant virality caught OpenAI off guard, its execs insist. “This was definitely surprising,” Mira Murati, OpenAI’s chief technology officer, says. At the San Francisco VC event, Altman said, he “would have expected maybe one order of magnitude less of everything—one order of magnitude less of hype.”

OpenAI CTO Mira Murati on ‘The Daily Show with Trevor Noah’ on Oct. 25, 2022.

ChatGPT isn’t OpenAI’s only hype generator. Its relatively small staff of around 300 has pushed the boundaries of what A.I. can do when it comes to creating, not simply analyzing, data. DALL-E 2, another OpenAI creation, allows users to create photorealistic images of anything they can imagine by typing just a few words. The system has now been emulated by others, including Midjourney and an open-source competitor called Stability AI. (All of these image generators have drawbacks, most notably their tendency to amplify biases in the data on which they were trained, producing images that can be racist and sexist.) By fine-tuning its GPT LLM on computer code, OpenAI also created Codex, a system that can write code for programmers, who only have to specify in plain language what they want the code to do.

More innovations wait in the wings. OpenAI has an even more powerful LLM in beta testing called GPT-4 that it is expected to release this year, perhaps even imminently. Altman has also said the company is working on a system that can generate video from text descriptions. Meanwhile, in mid-January, OpenAI signaled its intention to release a commercial version of ChatGPT, announcing a wait-list for would-be customers to sign up for paid access to the bot through an interface that would allow them to more easily integrate it into their own products and services.

A cynic might suggest that the fact OpenAI was in the middle of raising a large venture capital round might have something to do with the timing of ChatGPT’s release. (OpenAI says the timing is coincidental.) What’s certain is that ChatGPT chummed shark-filled waters. It set off a feeding frenzy among VC firms hoping to snap up shares in the private sale of equity currently being held by OpenAI’s executives, employees, and founders.

That tender offer is happening alongside the just-announced new investment from Microsoft, which will infuse up to $10 billion in new capital into the company. Microsoft, which started working with OpenAI in 2016, formed a strategic partnership with the startup and announced a $1 billion investment in the company three years ago. According to sources familiar with the new tender offer, it is heavily oversubscribed—despite an unusual structure that gives Microsoft a big financial advantage.

According to documents seen by Fortune, on completion of its new investment and after OpenAI’s first investors earn back their initial capital, Microsoft will be entitled to 75% of OpenAI’s profits until it earns back the $13 billion it has invested—a figure that includes an earlier $2 billion investment in OpenAI that had not been previously disclosed until Fortune reported it in January. Microsoft’s share will then step down to 49%, until the software giant earns a profit of $92 billion. Meanwhile, the other venture investors and OpenAI’s employees also will be entitled to 49% of OpenAI’s profits until they earn some $150 billion. If these caps are hit, Microsoft’s and investors’ shares will revert to OpenAI’s nonprofit foundation. In essence, OpenAI is lending the company to Microsoft—for how long depends on how quickly OpenAI can make money.

But earning back its investment, let alone hitting those caps, might take quite a while. The documents seen by Fortune reveal that OpenAI has had relatively modest revenues to date and is heavily loss-making. Last year, the company was projected to bring in just under $30 million in revenue, according to the documents. But it was projecting expenses of $416.45 million on computing and data, $89.31 million on staff, and $38.75 million in unspecified other operating expenses. In total, its net loss in 2022 excluding employee stock options was projected at $544.5 million. And with ChatGPT, those losses may be soaring: Altman said on Twitter, in response to a question from Elon Musk, that it was costing OpenAI “single-digit cents” in computing costs per interaction users have with ChatGPT—a tab that likely reached many millions of dollars per month as the bot became popular.

OpenAI is projecting that, with ChatGPT serving as a siren song to lure customers, its revenue will ramp up rapidly. It is forecasting $200 million in revenue for 2023 and expects revenues to top $1 billion in 2024, according to the documents. They do not project how OpenAI’s expenses might grow and when it could turn a profit. The companies declined to comment on these figures, but they point to an obvious reality: Both OpenAI and Microsoft think that the former nonprofit lab now has something it can sell.

Microsoft is already reaping the rewards of the partnership. It has launched an OpenAI-branded suite of tools and services in its Azure Cloud that will allow Azure customers access to OpenAI’s tech, including GPT and DALL-E tools. Auto marketplace CarMax, for example, has already launched new services that run on these Azure tools.

Eric Boyd, Microsoft’s corporate vice president of AI Platform, says that meeting the demands of training and running OpenAI’s LLMs has driven innovations that benefit all Azure customers. For instance, Microsoft has built supercomputing clusters for A.I. that it believes are the most powerful in the world, and created several software innovations to make it easier to train and run large A.I. models on these machines. Microsoft is gradually infusing OpenAI’s tech into much of its software. It has released an image creator within Bing, its search engine, and a new Designer graphic design tool, both powered by DALL-E; a GPT-3-enabled tool within its Power Apps software, and a code suggestion tool, GitHub Copilot, based on OpenAI’s Codex model.

Even if it doesn’t immediately move the needle on Azure revenue, the OpenAI relationship is good brand positioning and marketing, says Dan Romanoff, a senior equity research analyst who covers technology stocks for Morningstar. “It’s high-profile,” he says. “The ability to take an A.I. solution developed by OpenAI, put it on Azure, call it Azure AI: It keeps them competitive.” Microsoft’s Cloud rivals—Google, AWS, IBM, Oracle, Salesforce, and others—all have their own “cognitive” services, but being associated with the folks who created ChatGPT can’t hurt.

The bigger prize for Microsoft might be in search. Tech publication The Information recently reported that Microsoft plans to integrate ChatGPT into Bing, possibly allowing it to return simple, succinct answers to queries—and letting people delve deeper through dialogue with that chatbot—rather than a list of links. Google currently dominates the market for search, with a greater than 90% market share worldwide. Bing ranks a second so distant it might as well be in a different galaxy, with about a 3% share. In the first nine months of 2022, search was worth $120 billion in revenue for Google; overall, it accounts for about 60% of the money Google generates. ChatGPT may offer Microsoft the only real chance it’s ever had to knock Google off that pedestal. (Microsoft declined to comment on The Information report.)

And by Microsoft’s standards, these upsides come cheap. Its total investment of $13 billion is a hefty sum, but it’s only 15% of the $85 billion in pretax profits it booked over the past 12 months—a relative bargain for near-term control of a paradigm-shifting technology. For their part, OpenAI and Altman risk paying a different price: the possibility that Microsoft’s priorities crowd out their own, putting their broader mission at risk and alienating the scientists who fueled its successes.

One July evening in 2015, Altman, who was then the head of the prestigious startup incubator Y Combinator, hosted a private dinner at the Rosewood Sand Hill, a luxurious ranch-style hotel located in the heart of the Valley’s venture capital industry in Menlo Park. Elon Musk was there. So was Brockman, then a 26-year-old MIT dropout who had served as chief technology officer at payment-processing startup Stripe. Some of the attendees were experienced A.I. researchers. Some had hardly any machine learning chops. But all of them were convinced AGI was possible. And they were worried.

Google had just acquired what to Altman, Musk, and other tech insiders looked like the odds-on favorite to develop AGI first: London-based neural networking startup DeepMind. If DeepMind succeeded, Google might monopolize the omnipotent technology. The Rosewood dinner’s purpose was to discuss forming a rival lab to ensure that wouldn’t happen.

The new lab aimed to be everything DeepMind and Google were not. It would be run as a nonprofit, explicitly dedicated to democratizing the benefits from advanced A.I. It promised to publish its research and open-source all of its technology, a commitment to transparency enshrined in its very name: OpenAI. The lab garnered an impressive roster of donors: not only Musk, but his fellow PayPal colleagues Thiel and Hoffman; Altman and Brockman; Y Combinator cofounder Jessica Livingston; YC Research, a foundation that Altman had established; Indian IT outsourcing firm Infosys; and Amazon Web Services. Together, the founding donors pledged to give $1 billion to the idealistic new venture (although according to tax records, the nonprofit only received a fraction of the headline-grabbing pledge).

But training the giant neural networks quickly proved to be expensive—with computing costs reaching tens of millions of dollars. A.I. researchers don’t come cheap either: Ilya Sutskever, a Russian-born scientist who came to OpenAI to be its lead scientist after working at Google, was paid an annual salary of $1.9 million in his first few years at the lab, according to tax records. After a few years, Altman and others at OpenAI concluded that to compete with Google, Meta, and other tech giants, the lab could not continue as a nonprofit. “The amount of money we needed to be successful in the mission is much more gigantic than I originally thought,” Altman told Wired magazine in 2019.

Setting up a for-profit arm allowed OpenAI to raise venture capital. But OpenAI created an unusual structure that capped investors’ returns at a multiple of their initial investment. And OpenAI’s nonprofit board, which is stacked with Silicon Valley A-listers, would retain control of OpenAI’s intellectual property (see sidebar). One A-lister who didn’t stick around was Musk: In 2018, he left the board, citing the demands of running SpaceX and, more important, Tesla.

Around this time, Microsoft CEO Satya Nadella was desperate to prove that his company, perceived as trailing its rivals in A.I., could play at the technology’s bleeding edge. The company had tried and failed to hire a big-name A.I. scientist. It was also building a huge, expensive cluster of specialized chips to advance its own efforts on language models. It was just the sort of supercomputing power OpenAI needed—and which it was spending huge sums to purchase at the time. For its part, OpenAI excelled at pulling off the sort of splashy A.I. demos that Nadella desired to showcase Microsoft’s A.I. acumen. Altman approached Nadella about a deal, flying to Seattle several times to show him OpenAI’s A.I. models. Nadella ultimately signed a pact, announced in July 2019, to make Microsoft OpenAI’s “preferred partner” for commercializing its technology, alongside an initial $1 billion investment in the A.I. startup.

While Altman was involved in OpenAI from its inception, he did not become CEO until May 2019, shortly after it converted into a for-profit enterprise. But its trajectory from research lab to multibillion-dollar phenomenon reflects Altman’s unique fundraising prowess and product-oriented focus—as well as the tension between those commercial instincts and his commitment to big, science-driven ideas.

The OpenAI leader is in some ways a Silicon Valley caricature: youthful, male, and pale; unblinkingly intense; fluent in Geek; obsessed with maximizing efficiency and productivity; a workaholic devoted to “changing the world.” (In a 2016 New Yorker profile, he said he did not have Asperger’s syndrome but could understand why someone would think he did.)

Altman dropped out of a computer science degree program at Stanford University to cofound Loopt, a social media company whose app told you where your friends were. The company got into Y Combinator’s first batch of startups in 2005; Loopt failed to take off, but the money Altman earned when it was sold helped launch him into the VC universe. He started his own small VC firm called Hydrazine Capital that raised about $21 million, including money from Thiel. Then Paul Graham and Livingston, the Y Combinator cofounders, brought him in as Graham’s successor running YC itself.

Altman is an entrepreneur, not a scientist or an A.I. researcher, and he is known for being unusually adept at raising venture capital money. Convinced that great things come from the coupling of massive ambition and unflinching self-belief, he has said he aspires to create trillions of dollars of economic value via so-called deep-tech plays, in fields like nuclear fusion and quantum computing, where the odds are long but the payoffs potentially huge. “Sam believed he was the best at everything he took on,” says Mark Jacobstein, a veteran tech investor and startup adviser who worked with Altman at Loopt. “I am pretty sure he believed he was the best ping-pong player in the office until he was proven wrong.”

According to several current and former OpenAI insiders, the startup’s priorities began to shift as Altman took the reins. A once broad research agenda shrank to focus mostly on natural language processing. Sutskever and Altman have defended this shift as maximizing effort on the research areas that currently appear to offer the most promising path toward AGI. But some former employees say internal pressure to focus on LLMs grew substantially after Microsoft’s initial investment, in part because those models had immediate commercial applications.

Having been founded to be free of corporate influence, some complained, OpenAI was quickly becoming a tool for a gigantic technology company. “The focus was more, how can we create products, instead of trying to answer the most interesting questions,” one former employee said. Like many interviewed for this story, the employee requested anonymity because of nondisclosure agreements and to avoid alienating powerful figures associated with OpenAI.

OpenAI was also becoming a lot less open. It had already begun pulling back from the pledge to publish all its research and open-source its code, citing concerns that its technology could be misused. But according to former employees, commercial logic also played a role. By making its advanced models available only through APIs, OpenAI protected its intellectual property and revenue streams. “There was a lot of lip service paid to ‘A.I. safety’ by [Altman] and [Brockman] but that often seemed like just a fig leaf for business concerns, while actual, legitimate A.I. safety concerns were brushed aside,” another former OpenAI employee says. As an example, the former employee cited the way OpenAI quickly reversed a decision to limit access to DALL-E 2 because of fears of misuse as soon as Midjourney and Stability AI debuted rival products. (OpenAI says it allowed broader use of DALL-E 2 only after careful beta testing gave it confidence in its safety systems.) According to some former employees, these strategic and cultural shifts played a role in the decision of a dozen OpenAI researchers and other staff—many of whom worked on A.I. safety—to break with the company in 2021 and form their own research lab called Anthropic.

OpenAI says it continues to publish far more of its research than other A.I. labs. And it defends its shift to a product focus. “You cannot build AGI by just staying in the lab,” says Murati, the chief technology officer. Shipping products, she says, is the only way to discover how people want to use—and misuse—technology. OpenAI had no idea that one of the most popular applications of GPT-3 would be writing software code until they saw people coding with it, she says. Likewise, OpenAI’s biggest fear was that people would use GPT-3 to generate political disinformation. But that fear proved unfounded; instead, she says, the most prevalent malicious use was people churning out advertising spam. Finally, Murati says that OpenAI wants to put its technology out in the world to “minimize the shock impact on society that really powerful technology can have.” Societal disruption from advanced A.I. will be worse, she argues, if people aren’t given a teaser of what the future might hold.

Sutskever allows that OpenAI’s relationship with Microsoft created a new “expectation that we do need to make some kind of a useful product out of our technology,” but he insists the core of OpenAI’s culture hasn’t changed. Access to Microsoft data centers, he says, has been critical to OpenAI’s progress. Brockman also argues the partnership has allowed OpenAI to generate revenue while remaining less commercially focused than it would otherwise have to be. “Hiring thousands of salespeople is something that might actually change what this company is, and it is actually pretty amazing to have a partner who has already done that,” he says.

Sutskever categorically denies implications that OpenAI has de-emphasized safety: “I’d say the opposite is true.” Before the Anthropic split, A.I. safety was “localized to one team,” but it’s now the responsibility of every team, Sutskever says. “The standards for safety keep increasing. The amount of safety work we are doing keeps increasing.”

Critics, however, say OpenAI’s product-oriented approach to advanced A.I. is irresponsible, the equivalent of giving people loaded guns on the grounds that it is the best way to determine if they will actually shoot one another.

Gary Marcus, a New York University professor emeritus of cognitive science and a skeptic of deep learning–centric approaches to A.I., argues that generative A.I. poses “a real and imminent threat to the fabric of society.” By lowering the cost of producing bogus information to nearly zero, systems like GPT-3 and ChatGPT are likely to unleash a tidal wave of misinformation, he says. Marcus says we’ve even seen the first victims. Stack Overflow, a site where coders pose and answer programming questions, has already had to ban users from submitting answers crafted by ChatGPT, because the site was overwhelmed by answers that seemed plausible but were wrong. Tech news site CNET, meanwhile, began using ChatGPT to generate news articles, only to find that many later had to be corrected owing to factual inaccuracies.

For others, it’s ChatGPT writing accurate code that’s the real risk. Maya Horowitz, vice president of research at cybersecurity firm Check Point, says her team was able to get ChatGPT to compose every phase of a cyberattack, from crafting a convincing phishing email to writing malicious code to evading common cybersecurity checks. ChatGPT could essentially enable people with zero coding skills to become cybercriminals, she warns: “My fear is that there will be more and more attacks.” OpenAI’s Murati says that the company shares this concern and is researching ways to “align” its A.I. models so they won’t write malware—but there is no easy fix.

Countless critics and educators have decried the ease with which students can use ChatGPT to cheat. School districts in New York City, Baltimore, and Los Angeles all blocked school-administered networks from accessing the chatbot, and some universities in Australia said they would revert to using only proctored, paper-based exams to assess students. (OpenAI is working on methods to make A.I.-generated text easier to detect, including possibly adding a digital “watermark” to ChatGPT’s output.)

There are also ethical concerns about the way ChatGPT was originally assembled in 2022. As part of that process, OpenAI hired a data-labeling company that used low-wage workers in Kenya to identify passages involving toxic language and graphic sexual and violent content, a Time investigation found. Some of those workers reported mental health issues as a result. OpenAI told Time in a statement such data labeling was “a necessary step in minimizing the amount of violent and sexual content included in training data and creating tools that can detect harmful content.”

Making ChatGPT freely available has allowed OpenAI to gather a treasure trove of feedback to help improve future versions. But it’s far from certain OpenAI will maintain its dominance in language A.I. “Historically, what we have tended to see with these very general-purpose algorithms is that they are not sufficiently defensible to allow just one particular company to capture all the general returns,” says Marc Warner, founder and CEO of London-based A.I. company Faculty. Face- and image-recognition technology, for example, was first developed at tech giants such as Google and Nvidia but is now ubiquitous.

Courts and regulators could also thrust a giant stick into the data flywheels on which generative A.I. depends. A $9 billion class action lawsuit filed in federal court in California potentially has profound implications for the field. The case’s plaintiffs accuse Microsoft and OpenAI of failing to credit or compensate coders for using their code to train GitHub’s coding assistant Copilot, in violation of open license terms. Microsoft and OpenAI have declined to comment on the suit.

A.I. experts say that if the court sides with the plaintiffs, it could derail the generative A.I. boom: Most generative models are trained from material scraped from the internet without permission or compensation. The same law firm representing those plaintiffs recently filed a similar lawsuit against Stability AI and Midjourney, for using copyrighted art in their training data without permission. Photo agency Getty Images has filed its own copyright infringement lawsuit against Stability AI too. Another problem could come if lawmakers pass rules giving creators a right to opt out of having their content used in A.I. training, as some European Union lawmakers are considering.

OpenAI’s competitors, meanwhile, are not standing still. The prospect of losing its dominance in search has motivated execs at Google to declare a “red alert,” according to the New York Times. Sundar Pichai, Google’s CEO, has held meetings to redefine the company’s A.I. strategy and plans to release 20 new A.I.-enabled products as well as demonstrate a chat interface for search within the year, the newspaper reported. Google has its own powerful chatbot, called LaMDA, but has been hesitant to release it because of concerns about reputational damage if it winds up being misused. Now, the company plans to “recalibrate” its appetite for risk in light of ChatGPT, the Times reported, citing an internal company presentation and unnamed insiders. Google is also working on a text-to-image generation system to compete with OpenAI’s DALL-E and others, the newspaper reported.

Of course, it’s not clear that chatbots will be the future of search. ChatGPT frequently invents information—a phenomenon A.I. researchers call “hallucination.” It can’t reliably cite its sources or easily surface links. The current version has no access to the internet, and so it cannot provide up-to-date information. Some, such as Marcus, believe hallucination and bias are fundamental problems with LLMs that require a radical rethink of their design. “These systems predict sequences of words in sentences, like autocomplete on steroids,” he says. “But they don’t actually have mechanisms in place to track the truth of what they say, or even to validate whether what they say is consistent with their own training data.”

Others, including OpenAI investors Hoffman and Vinod Khosla, predict these problems will be solved within a year. Murati is more circumspect. “There are research directions that we have been following so far to kind of address the factual accuracy and to address the reliability of the model and so on. And we are continuing to pursue them,” she says.

In fact, OpenAI has already published research about a different version of GPT, called WebGPT, that had the ability to answer questions by querying a search engine and then summarizing the information it found, including footnotes to relevant sources. Still, WebGPT wasn’t perfect: It tended to accept the premise of a user’s question and look for confirmatory information, even when the premise was false. For example, when asked whether wishing for something could make it happen, WebGPT replied, “It is true that you can make a wish true by the power of thought.”

On the rare occasions that Altman lets himself rhapsodize about A.I. in public, he can sound like a wishful thinker himself. Asked at the San Francisco VC event about the best case for A.I., he gushes, “I think the best case is so good that it’s hard to imagine … I think the good case is just so unbelievably good that you sound like a crazy person talking about it.” He then abruptly returns to the dystopian themes at OpenAI’s roots: “I think the worst case is lights-out for all of us.”

热读文章
热门视频
扫描二维码下载财富APP