The Mad Rush to publish AI Research

By 2017, it became clear that big tech companies were deeply interested in AI. Companies like Apple, Alphabet, Amazon, Meta, IBM and Microsoft amped up their investments in AI, specifically in the area of machine learning. In the same year, Pitchbook reported that companies around the world had spent USD 21.3 billion on AI-related mergers and acquisitions, an amount believed to be 26 times more than its value in 2015. Jeff Wilke, former chief executive of Amazon worldwide consumer and close ally to company CEO Jeff Bezos then stated, “If you’re a tech company and you’re not building AI as a core competence, then you’re setting yourself up for an invention from the outside.” 

History

Between 2000 and 2016, companies like IBM and Microsoft were already investing heavily in AI research. Google and Facebook were only moderately involved in AI research and hiring researchers depending on how profitable the project they were working on was. Apple and Amazon were on the other end of the spectrum and emptied their purse only when the research work led to a product. Things, however, changed drastically within a year. 

Source: The Economist

By mid-2018, Google made a definitive pivot. The company announced that its Google Research division would be rebranded as Google AI just ahead of its Google I/O conference. The move showed that the tech giant was ready to aggressively pursue AI. Google changed how the Google AI homepage looked—the blog would now show published research papers, new AI tools and other developments in the field. A month prior to that, Jeff Dean, one of the earliest programmers to join Google was appointed as head of Google AI. Within Google AI, was ‘Google Brain’, the team behind ‘Tensorflow’ that also collaborated with the company’s AV unit, ‘Waymo’. By 2016, Google already stood on top for the most number of AI-related patents. 

This indicated a change in trend in AI-related research work. Companies that used to publish little or no AI research started putting out hordes of papers. Consequently, the secrecy around the work being done dissolved. Big tech now looked like it was in a mad rush to showcase seemingly random research papers in AI/ML. 

In the early 2000s, businesses were not considered as traditional contributors to scientific literature. Big corporations like Microsoft have remained an exception to this rule. Microsoft Research was founded early in September 1992. During this time, the R&D for most other places like IBM Research, Bell Labs, GE Research (General Electric research) and Xerox’s research division, which eventually became the Palo Alto Research Center or PARC, had essentially become defunct. 

Nathan Myhrvold, who was leading advanced technology in the company at the time, formed the proposal for Microsoft Research. Myhrvold argued for corporate research emphasising that despite the massive failure that Xerox faced, research was in fact beneficial. He claimed, “The fact is that research does work at a lot of companies. When research fails it almost never is because of an intrinsic problem in research itself (i.e., the inability to think of something new). Instead, the research usually falls prey to problems that can be traced to general management issues—having the right goals, transitioning technology to company benefit etc.” 

Microsoft Research ended up creating a blueprint for how R&D centres could sustain themselves. Myhrvold created a draft plan for the research centre over the next five years that proved accurate for Microsoft. 

Funding

Aside from the general excitement for algorithms, a significant reason behind corporations supporting research was that for the majority of time, they were the only ones who could afford it. Several research papers are often published simply to create a positive feedback loop for the companies. Considering the pace at which technology is moving, teams choose not to hold onto research as there is a possibility that the technology might perish quickly. Companies are wary that the economic value extracted from the research might drop at any point in time which causes them to rush. 

In cases where the fundamental research is promising but the timeline to extract economic gains from it is long term, smaller tech companies normally back out leaving the field only for tech companies with deep pockets. Companies like Google are aware that since capital is not a barrier for them, they can afford to keep pushing the timeline and eventually pour in notable monetary gains to train their AI models. This gives big tech companies the first mover advantage. For instance, OpenAI’s 175-billion parameter model GPT-3 cost around USD 12 million to train. ‘OpenAI’, a well-funded AI startup supported by Microsoft, has often stated that AGI can only be achieved if reinforcement learning is combined with a massive computing power, which can prove to be very expensive. 

Hiring talent

Since the AI boom, companies have joined in the gold rush. While Google AI is perceived to be in the lead, the race has only just begun. Brands that are able to hire more AI talent will have a competitive edge over their rivals. In AI, talent scarcity has become the point of contention as data and computation power has become almost evenly available to companies. Academic stars like Meta’s Yann LeCun and Google’s Geoffrey Hinton are sometimes a major attraction for researchers. However, most researchers are mainly interested in publishing their own work and the more research a company publishes, the more content that functions as a marketing tool to continue attracting more researchers from academia. 

AI/ML conferences and scientific journals award research papers that put forth path breaking innovations. Scientists and researchers who are already working in tech corporations are instructed to work in a specific area. Such directives narrow down the scope of research for many researchers and streamline the topics for their research. Consequently, there is a steady stream of papers that is produced. 

Additionally, the research doesn’t merely build prestige for the company but also for the researchers. In an Economist interview, Meta VP and Chief AI scientist Yann LeCun explained how the AI talent war between companies has compelled them to become more open about ongoing work. LeCun further explained that instructing researchers to remain tight-lipped about their research did not work in the long run as a significant number of researchers considered it a career killing move. 

Patents

Publishing a research paper usually gives the company behind it sufficient time to file a patent. According to U.S. patent laws if a competitor attempts to file a patent for a new tool or concept in ML, there will be no transfer of ownership owing to the published research. Typically, if companies view the research as economically profitable, they tend to file a patent along with the paper or prior to its publication.