Facebook to Build Its Own Silicon for Real-Time Content Filtering

Facebook to Build Its Own Silicon for Real-Time Content Filtering

Content filtering is a major problem for all social media sites. Faced with a veritable avalanche of material being uploaded every second, firms like Facebook, Instagram, and YouTube have struggled to handle a suite of problems, ranging from mundane copyright detection and flagging to issues like revenge porn. Human beings are far better than computers at content classification and determining whether images and video violate a site’s terms of service. But human scanning doesn’t scale well, and often causes significant levels of psychological distress for the employees in question. There’s a huge potential market for products that can handle this work more effectively — and FB, apparently, is tossing its own hat into the silicon design ring.

At the Viva Technology conference, Facebook’s chief AI scientist, Yann LeCun, said the company wants to be able to intervene in a case where someone uploads video or livestreams a murder or suicide, but that current solutions now require either a huge amount of compute power, a great deal of electricity, or both. Tom’s Hardware reports LeCun as saying:

There’s a huge drive to design chips that are more energy-efficient for that. A large number of companies are working on this, including Facebook…You’ve seen that trend from hardware companies like Intel, Samsung, Nvidia. But now you start seeing people lower in the pipeline of usage having their own needs and working on their own hardware.

On the one hand, it makes perfect sense that Facebook and other social media companies would be working on this kind of capability. YouTube has been hammered in the past year over inappropriate content essentially created by bots and dumped into channels intended for children. Facebook’s solution to revenge porn involves preemptively uploading porn of yourself to the company, and trusting them to use it as a blueprint for removing illegal uploads if any ever occur. Meanwhile, the tricks used to bypass scanning techniques and filters have become ever-more sophisticated.

Creating hardware-based solutions that can handle this workload in real-time also represents a potential weapon. In the West, where freedom of speech is generally well-protected, this is less of an issue. But it’s downright chilling when combined with aspects of China’s Great Firewall or the country’s recent shift to deploy a social credit system. And, of course, systems that can quickly flag content for being for or against a given company’s terms of service could be repurposed for other reasons as well — including monitoring what people post and flagging content that could be seen as “problematic” for reasons beyond depicting suicide, murder, or similarly illegal activity.

China’s social credit scoring system isn’t just a means to monitor citizens — it applies to companies as well, with penalties assessed for firms that fail to pay taxes on time or sell substandard products. Foreign Policy has a detailed write-up of how the scheme has been implemented to date in China, and the degree to which it’s already being baked into decisions about who qualifies for social services in that country. At this point, it’s not even dystopian to suggest a plausible future in which machine intelligence and real-time content monitoring are used to create systemic profiles of how people use social media to control access to various societal benefits. That future is literally happening. And projects like Facebook’s, while well-intentioned, are likely to embed such initiatives even further into society.

Facebook to Build Its Own Silicon for Real-Time Content Filtering

The other major shake-up implied by work like this will have less of an impact on society, though it presages an enormous shift in the computing industry. For decades, overall compute performance has been driven forward by generalists. First, CPUs accelerated workload performance across the entire industry. Then, companies like Nvidia and AMD built programmable GPUs, which can be thought of as offering a “generalist” programmable core for some specific types of problems (this is one reason why the acronym GPGPU or General Purpose Graphics Processing Unit is sometimes used to refer to modern programmable GPUs). In both cases, the argument being made is that companies like Facebook, Google, and Microsoft can maximize the value of their R&D spending by relying on the compute capabilities being built by specialized firms like Intel, AMD, and Nvidia.

Google’s second-generation TPU.
Google’s second-generation TPU.

If the major players like Facebook and Google continue to design their own specialized hardware, it could collectively challenge the dominance of major hardware players that have effectively defined computing for decades. If you were to make a list of the companies that have shaped the environments we compute in, it’s not a particularly large list, and the major entrances and exits tend to also happen around major shifts in computing paradigms.

Firms that dominate one era don’t always succeed in later time periods, and it’s not unusual for a company that plays a part in multiple distinct time periods to play a role in very different ways. IBM, for example, has shifted from a hardware-centric business to a software and solutions provider across the last few decades. Intel, AMD, Nvidia, and a host of other companies are betting that they can shift to a data center-centric business by powering these new compute paradigms with general purpose hardware. If the self-designed trend keeps up, it could threaten these revenue streams long-term.