Facebook首席科技官：新员工要做好挨批的准备

当Facebook的首席科技官迈克·施洛普弗欢迎新招聘工程人员加入这家社交网络公司时，他首先谈到的是公司在过去几年中遭到的铺天盖地的批评。

施洛普弗对《财富》杂志说：“就像是，‘嗨，欢迎加入Facebook，我先向大家介绍一下有关批评Facebook的一些头条新闻。’我对所有人说，挨批是必然的。”

在这个动荡的时代，在Facebook工作会遇到各种隐私崩溃问题、信息误传问题，以及各类其他不愉快的事情。这些事情都会迫使施洛普弗思考公司科技可能会造成的意外后果，以及工程师保护Facebook这样的平台免受恶人侵害的职责。

施洛普弗在谈到那些试图利用社交网络实现其政治或经济目的的人时表示：“由于平台由乐观主义者和技术人员搭建，因此他们不大会用恶意去揣度别人，而这正是你们必须做的事情，也就是防范这类恶意。”

施罗普弗说自己曾经告诫新招聘员工：“我们的业务规模非常大，因此每个人肩负的责任也是十分重大。人们会仔细审视你所做的每件事情，而我们的工作就是确保通过这类审查。要做到这一点，我们必须付出大量的努力。”

过去几年让施罗普弗发生了改变，今年夏天早些时候接受《纽约时报》采访时激烈的措辞便说明了这一点，而这些措辞也引起了一些人的抵触。Facebook高管对此的反应是：“没事。”

他说：“我正在努力保持自己的乐观情绪，同时也对可能出现危害的事实有了更加深刻的认识。”

人工智能专家施罗普弗最近也在招兵买马，因为除了他自己打造可过滤虚假新闻的人工智能系统的工作之外，Facebook后续还将推出一系列项目。以下是Facebook正在研究的一些新项目：

虚拟现实手部动作跟踪

Facebook即将发布的更新能够让其Oculus Quest虚拟现实头戴设备跟踪手部动作，要实现这一目标则需要大量的人工智能技术作为支撑。尽管Quest设备配备了可识别人手的摄像头，然而人们在虚拟世界中移动人手时，这些摄像头无法从各个角度捕捉人手的动作。为了解决这个问题，Facebook不得不组建一个人工智能模型，以识别手部动作，从而帮助摄像头解决这个问题。

人工智能芯片

当前涌现出的一大批初创企业都在打造用于协助人工智能相关任务的计算机芯片。施罗普弗说：“从硬件层面来讲，现在真是个好时候。”按照他的话说，原因在于“摩尔定律已经失效了。”按照该定义，计算机芯片晶体管数量每两年会翻一番。

如今，越来越多的企业都试图打造新一代的颠覆性半导体。施罗普弗对图形处理芯片（GPU）制造商英伟达表示了赞赏，早在2004年，该公司的芯片便可以在优化视频游戏的同时用于处理其他任务。英伟达的GPU自那之后成为了用于培训神经网络的首要计算机芯片。

老当益壮的CPU

施罗普弗表示，尽管Facebook使用大量的GPU来培训众多的人工智能系统，但公司依然会使用传统的CPU来执行一些重负荷任务。为了培训擅长“点击预测”的庞大神经网络，Facebook使用了“非常庞大的CPU集群”，从而让公司能够更加高效地处理各类信息，即便是处理能力高达32GB的最强GPU也只能甘拜下风。

应对深度伪造的崛起

Facebook最近表示将赞助一项赛事。赛事期间，研究人员将开发用于侦测所谓深度伪造的最佳方式。这些视频的内容与看起来与现实无异，其目的是欺骗或误导观众。作为这个竞赛的一部分，Facebook表示，公司将打造一个深度伪造视频数据库，来帮助研究人员开发辨别真伪的技术。

凑巧的时，谷歌在9月下旬表示，公司也将开展这方面的举措，向人工智能研究人员公布其自有的深度伪造视频海量数据组，来打击深度伪造的崛起。在谈到谷歌深度伪造数据组时，施罗普弗对《财富》杂志说，他希望将数据组进行整合，这样，这场由Facebook赞助、非营利性机构Partnership on AI监管的竞赛可以有更多的伪造视频可供分析。

施罗普弗说，尽管深度伪造这个问题在平台上还不是很严重，但“如果要问我在过去三年中都学会了什么，那就是我不希望打没有准备的仗。”

他说：“我宁愿带着我不需要的工具，随时做好战斗的准备，而不是在某一天，特别是在例如2020年美国大选期间，当深度伪造成为一种问题的时候，我却没有投资相应的技术来应对这个问题。”

译者：冯丰

审校：夏林

When Facebook chief technology officer Mike Schroepfer welcomes new engineering recruits to the social network, the first thing he talks about is the massive criticism facing the company over the past few years.

“It’s like, ‘Hi, welcome to Facebook, let me show you a bunch of news headlines of people criticizing us,” Schroepfer tells Fortune. “And what I tell everyone is the criticism is warranted.”

Working at Facebook during this tumultuous time has included multiple privacy debacles, the spread of misinformation, and various other unpleasantries that have caused Schroepfer to consider the potential unintended consequences of the company's technology—and the role of engineers to safeguard platforms like Facebook from people with bad intentions.

“As optimists and technologists building, it’s hard to assume bad intent, and that’s exactly what you have to do to plan for these things,” Schroepfer says referring to people who are trying to exploit the social network for political or economic gain.

“We’re operating something on massive scale, you bear a huge responsibility—people should scrutinize everything you do, and our job is to prevail in the face of that scrutiny and that requires us to do a lot of hard work,” Schroepfer says he tells his hires.

These past few years have changed Schroepfer, as he revealed in an emotional interview earlier this summer in The New York Times, which prompted some backlash. The Facebook executive's response: “That’s fine.”

“I’m fighting hard to retain my optimism while coupling with a much deeper sense of sort of realism about the harms that can happen,” he says.

Schroepfer, an artificial intelligence expert, has been staffing up recently to because Facebook has on an array of projects in its pipleline, in addition to his work building A.I. systems that can filter fake news. Here are some of the things the company is working on:

Virtual reality hand-tracking

Facebook’s forthcoming update that enables its Oculus Quest virtual reality headsets to track hand movements has required a lot of artificial intelligence to make it possible. Although the Quest device has cameras to help it recognize a person’s hands, it’s not possible for the cameras to capture all of the various angles when someone moves their hands around in a virtual world. To accommodate, Facebook had to build an A.I. model that could recognize hand movements to help the cameras out.

Artificial intelligence chips

Regarding the current boom in startups creating computer chips designed for to aid A.I.-related tasks, Schroepfer says “it’s a really interesting time for hardware.” This is because, he says, “Moore’s Law is dead,” referring to the notion that the number of transistors on a computer chip doubles every two years.

Now there's an increasing amount of players trying to create the next game-changing semiconductor. Schroepfer commends Nvidia, the maker of the graphical processing units (GPUs,) for recognizing as early as 2004 that its chips could be used for other tasks besides making video games look prettier. Nvidia’s GPUs have since become the leading computer chips used to train neural networks.

Good ol' CPUs

Although Facebook uses tons of GPUs to train it many A.I. systems, it also uses conventional CPUs for some heavy-duty tasks, Schroepfer says. To train giant neural networks that are good at “click predicting,” Facebook uses “very large fleets of CPUs” that allow it to crunch information more efficiently than even the beefiest GPUs that top out at 32 GBs.

Combating the rise of deep fakes

Facebook recently said it would fund a competition in which researchers would develop the best ways to detect so-called deepfakes, which are realistic looking videos intended to fool or mislead people. As part of the competition, Facebook said it would create a database of deepfake videos to help researchers develop techniques to detect what’s real and what’s not.

Coincidentally, at late September, Google said it would also contribute to efforts to combat the rise of deepfakes by releasing its own giant dataset of deepfake videos to A.I. researchers. Regarding the Google deepfake datasets, Schroepfer tells Fortune his hope is to merge the datasets so the Facebook-funded competition, overseen by the non-profit Partnership on AI, will have more fake videos to analyze.

While deepfakes are not a clear issue on the platform today, Schroepfer says, “if I’ve learned anything in the last three years, I don’t want to be unprepared for something.”

“I’d rather be ready to go with tools that I don’t need to use, than be in a situation especially with—say, the 2020 U.S. elections—where it becomes an issue and I haven’t invested in the technology to defend against it,” he says.