October 1, 2020 | technology | No Comments
London-based venture capital firm Air Street Capital today published the State of AI Report 2020, its third annual survey canvassing research, talent, industrial, and political trends in the field of AI. Coauthored by University College London visiting professor Ian Hogarth and AI investor Nathan Benaich, the report aims to highlight technological breakthroughs and areas of commercial application for AI as well as the regulation of AI, its economic implications, and emerging geopolitical issues.
Among other findings, this year’s report implies AI remains mostly closed source, harming accountability and reproducibility, while corporate-driven academic “brain drain” appears to be impacting entrepreneurship. Self-driving cars are in the Precambrian stages. And political leaders are beginning to question whether acquisitions of AI startups should be scrutinized or outright blocked.
According to Air Street Capital’s report, only 15% of AI research papers publish their code, and there’s been little improvement on the metric since mid-2016. Based on data from the website Papers With Code, which highlights trending research and the code to implement it, code availability has actually decreased from above 20% in December 2019. “For the biggest tech companies, their code is usually intertwined with proprietary scaling infrastructure that can’t be released,” the coauthors of the report concluded. “This points to the centralization of AI talent and compute as a huge problem.”
This development perhaps isn’t surprising, given that professors are departing universities for corporations at an accelerating rate. The report points out that Google, DeepMind, Amazon, and Microsoft hired 52 tenured and tenure-track professors from U.S. colleges between 2004 and 2018 and that Carnegie Mellon, the University of Washington, and Berkeley lost 38 professors during the same period. To put that in perspective, no AI professor left in 2004, whereas in 2018, 41 AI professors resigned.
The report’s coauthors believe this has a traceable impact on entrepreneurship. Citing research conducted by the University of Rochester and CKGSB Business School China, graduates are 4% less likely to start an AI company for 4-6 years after the departure of tenured AI professors, which could trickle down to immigration. The Paulson Institute-affiliated think tank MacroPolo found that the majority of the top AI researchers working in the U.S. (69%) weren’t trained in America.
On the other hand, there’s evidence to suggest international AI Ph.D. candidates studying in the U.S. are countering the brain drain despite immigration challenges, with data from the Center for Security and Emerging Technology showing that 92% of graduates work in the U.S. post-graduation and 80% intend to stay if they can. In other encouraging news, universities continue to expand AI course enrollment. Stanford now teaches 10 times the AI students per year as during 1999-2004 and twice as many as 2012-2014; over 1,200 Stanford students took classes in natural language processing between 2016 and 2017.
Beyond AI and machine learning research broadly, the Air Street Capital report takes a critical look at progress (or lack thereof) in the over $7 billion autonomous vehicles industry. While over half of U.S. states have enacted legislation related to autonomous vehicles, the safety of those vehicles remains difficult to measure because the market tends to focus on miles per disengagement, a flawed metric. Baidu recently made the dubious claim that it saw an 8,697% year-over-year improvement in disengagements from 2018 to 2019, beating out Waymo, Cruise, Zoox, Nissan, Nvidia, and Mercedes Benz.
Overcoming the industry’s hurdles might require new technological approaches, the report’s coauthors suggest, citing the failure of Starsky Robotics. The now-shuttered Starsky, which was the first company to drive an autonomous unmanned truck on a public highway, openly cited the challenges of scaling supervised machine learning. Use of machine learning in self-driving is mostly limited to perception, with large parts of the technology stack hand-engineered.
Commercialization and politics
Despite the pandemic, academic brain drain, and setbacks in autonomous vehicles, the report asserts that the AI industry remains strong. According to Pitchbook data retrieved by the coauthors on August 13, the year 2020 is likely to hit over $25 billion in deal volume across over 350 deals. Since 2018, rounds over $100 million have consistently accounted for about 10% of all funding.
But the report also warns that investments will likely slow down as countries enact legislation intended to keep companies within their borders. In June 2020, Germany passed a law to allow the government to review or block investments or takeovers by non-European Union-based robotics, AI, and semiconductor companies. That same month, the U.K. expanded its powers to intervene in mergers on public interest grounds, enabling it to review mergers and acquisitions involving AI companies where the target company has revenues exceeding £1 million (~$1.27 million).
The report’s coauthors predict that in response, a wave of Chinese and European defense-focused AI startups will collectively raise over $100 million within the next 12 months. “Despite fears of an AI winter, [there’s] innovation and progress behind the scenes,” Benaich said in a statement. “Many of these AI developments are powered by the massive computing infrastructure that is increasingly in the hands of big tech companies. We need to think carefully about what this means for the future of the field across both industry and academia.”