Statement of Greg Brockman
Co-Founder and Chief Technology Officer
OpenAI
before the
Subcommittee on Space, Science, and Competitiveness
Committee on Commerce, Science, and Transportation
United States Senate
Hearing on “The Dawn of Artificial Intelligence”
Thank you Chairman Cruz, Ranking Member Peters, distinguished members of the
Subcommittee. Today’s hearing presents an important first opportunity for the members of the
Senate to understand and analyze the potential impacts of artificial intelligence on our Nation
and the world, and to refine thinking on the best ways in which the U.S. Government might
approach AI. I'm honored to have been invited to give this testimony today.
By way of introduction, I'm Greg Brockman, co-founder and Chief Technology Officer of
OpenAI. OpenAI is a non-profit AI research company. Our mission is to build safe, advanced AI
technology and ensure that its benefits are distributed to everyone. OpenAI is chaired by
technology executives Sam Altman and Elon Musk.
The US has led the way in almost all technological breakthroughs of the last hundred years, and
we've reaped enormous economic rewards as a result. Currently, we have a lead, but hardly a
monopoly, in AI. For instance, this year Chinese teams won the top categories in a Stanford
University-led image recognition competition. South Korea has declared a billion dollar AI fund.
Canada produced some technologies enabling the current boom, and recently announced an
investment into key areas of AI.
I'd like to share 3 key points for how we can best succeed in AI and what the U.S. government
might do to advance this agenda. First, we need to compete on applications, but cooperate on
open, basic research. Second, we need to create public measurement and contests. And third, we
need to increase coordination between industry and government on safety, security, and ethics.
I. Competition and Cooperation
AI applications are rapidly broadening from what they were just a few years ago: from helping
farmers decide which fields to seed, to warehouse robots, to medical diagnostics, certain AI-
enabled applications are penetrating and enabling businesses and improving everyday life. These
and other applications will create new companies and new jobs that don't exist today — in much
the same way that the Internet did. But even discovering the full range of applications requires
significant scientific advances. So industry is not just working on applications: companies like
Facebook, Google, and Microsoft are performing basic research as well, trying to create the
essential AI building blocks which can later be assembled into products.
Perhaps surprisingly, the industry labs are publishing everything they discover. Publication
allows them to pool their resources to create faster breakthroughs, and to attract top scientists,
most of whom are motivated more by advancing society and improving the future, than personal
financial gain.
Companies stay competitive by publishing their basic research, but not the details of their
products. The inventor of a technique is usually the first to deploy it, as it has the right in-house
infrastructure and expertise. For example, AI techniques developed by Google's subsidiary
DeepMind to solve Atari video games were applied to increase the efficiency of Google's own
data centers. DeepMind shared their basic techniques by publishing the Atari research papers, but
did not share their applied work on data center efficiency.
Openness enables academia and industry to reinforce each other. Andrew Moore of Carnegie
Mellon University says it's not unusual that between 10 and 20 percent of the staff he hires will
take leaves of absence to work in industry or found a startup. Pieter Abbeel, a researcher at
OpenAI, splits his time between OpenAI and the University of California at Berkeley; likewise,
Stanford Professor Fei-Fei Li is spending time at both Stanford and Google; and many other
companies and organizations work with academics. This ensures that the private sector is able to
master the latest scientific techniques, and that universities are able to understand the problems
relevant for industry.
Openness has concentrated the world's AI research activity around the US (including attracting
many of the Canadian scientists who helped start the current AI boom), and allowed us to define
its culture and values. Foreign firms like China's Baidu have opened US-based research labs and
have also started publishing. As AI becomes increasingly useful, the pool of experts we're
gathering will be invaluable to ensuring that its economic activity also remains centered on the
US.
Recommendations -
We recommend the following, to ensure that our basic AI research community remains the
strongest in the world:
A. Maintain or increase basic research funding for AI:
In 2015 the government's unclassified
investment in AI-related technology was approximately $1.1 billion, according to The National
Artificial Intelligence Research and Development Strategic Plan report from the National
Science and Technology Council [1]. As highlighted by Jason Furman, Chairman of the Council
of Economic Advisers, there's evidence that the socially optimal level of funding for basic
research is two to four times greater than actual spending[2]. Given that it only takes months for
a basic AI advance to result in new companies and products, usually by whoever made the
advance, we support increasing funding for basic research in this domain. If we want these
breakthroughs to be made in the US, we'll need to conduct basic research across a number of
subfields of AI, and encourage the community to share their insights with each other. We'll need
to allow our academics to freely explore ideas that go against consensus, or whose value has high
uncertainty. This is supported by history: companies like Google and Microsoft rely on AI
technologies that originated with a small group of maverick academics.
B. Increase the supply of AI academics:
Industry has an insatiable demand for people with AI
training, which will only increase for the foreseeable future. We need to grow the supply of
people trained in AI techniques; this will let us make more research breakthroughs, give industry
the people it needs to commercialize the basic science, and train the next generation of scientists.
NSF could explore adjusting its policies to allow more competitive salaries for those working on
federal academic grants.
C. Enhance the professional diversity of the AI field:
Today, AI consists mostly of individuals
with degrees in computer science, mathematics, and neuroscience, with a significant gender bias
towards men. As AI increases its societal impact, we need to increase the diversity of
professional views within the AI community. Government can explore making more
interdisciplinary research grants available to incentivize experts in other fields, such as law or
agriculture or philosophy, to work with AI researchers. We also support the White House's
Computer Science for All initiative, and the OSTP's recommendation that government should
create a federal workforce with diverse perspectives on AI.
II. The Need For Public Measurement and Contests
Objective measures of progress help government and the public distinguish real progress from
hype. It's very easy to sensationalize AI research, but we should remember that advanced AI has
seemed just around the corner for decades. Good policy responses and a healthy public debate
hinge on people having access to clear data about which parts of the technology are progressing,
and how quickly. Given that some AI technologies, such as self-driving cars, have the potential
to impact society in a number of significant ways, we support OSTP's recommendation that the
government keep a close watch on the advancement of specific AI technologies, and work with
industry to measure the progression of the technology.
Also, having a measurable goal for AI technologies helps researchers select which problems to
solve. In 2004, DARPA hosted a self-driving car competition along a 150-mile course in the
Mojave Desert — the top competitor made it only seven miles. By 2007, DARPA hosted an
Urban Challenge to test self-driving cars on a complex, urban environment, and six of the eleven
teams completed the course. Today, Uber, Google, Tesla, and others are working on
commercializing self-driving car technology.
Similarly, when Fei-Fei Li and her collaborators at Stanford launched the image recognition
ImageNet competition in 2010, it was designed to be beyond the capabilities of existing systems.
That impossibility gave the world's research community an incentive to develop techniques at the
very edge of possibility. In 2012, academics won first place using a neural network-based
approach, which proved the value of the technique and kickstarted the current AI boom. The
winning ImageNet team formed a startup and were subsequently hired by industry to create new
products. One member, Ilya Sutskever, is one of my co-founders at OpenAI, and the other two
members work at Google. This shows how competitions can provoke research breakthroughs,
and translate into an economic advantage for industry.
We're moving from an era of narrow AI systems to general ones. Narrow AI systems typically do
one thing extremely well, like categorize an image, transcribe a speech, or master a computer
game. General AI systems will contain suites of different capabilities; they will be able to solve
many tasks and improvise new solutions when they run into trouble. They will require new ways
to test and benchmark their performance. Measuring the capabilities of these new multi-purpose
systems will help government track the technology's progress and respond accordingly.
Recommendations -
Government can create objective data about AI progress in the following ways:
A. Modern competitions:
AI systems have often been measured by performance on a static
dataset. Modern systems will act in the real world, and their actions will influence their
surroundings, so static datasets are a poor way to measure performance. We need competitions
which capture more of the complexity of the real world, particularly in developing areas such as
robotics, personal assistants, and language understanding. The government can continue
designing competitions itself, as DARPA did recently with the Cyber Grand Challenge, or
support others who are doing so.
B. Government information gathering:
Government should gather information about the AI
field as a whole. Researchers tend to focus on advancing the state of the art in one area, but the
bigger picture is likely to be crucial for policymakers, and valuable to researchers as well. The
government can invest in careful monitoring of the state of the field, forecasting its progress, and
predicting the onset of significant AI applications.
III. Increase coordination between industry and
government on safety, security, and ethics
The Internet was built with security as an afterthought, rather than a core principle. We're still
paying the cost for that today, with companies such as Target being hacked due to using insecure
communication protocols. With AI, we should consider safety, security, and ethics as early as
possible, and bake these into the technologies we develop.
Academic and industrial participants are starting to coordinate on responsible development of AI.
For example, we recently worked with researchers from Stanford, Berkeley, and Google to lay
out a roadmap for safety research in our paper “Concrete Problems in AI Safety”[3]. Non-profit
groups like the Partnership on AI and OpenAI are forming to ensure that research is done
responsibly and beneficially.
Recommendations -
Industry dialog:
Government can help the AI community by giving feedback about the what
aspects of progress it needs to understand in preparing policy. As the OSTP recommended in its
report, Preparing for the future of Artificial Intelligence[4], the NSTC Subcommittee on Machine
Learning and Artificial Intelligence should meet with industry participants to track the
progression of AI. OpenAI and our peers can use these meetings to understand what we should
monitor in our own work to give government the telemetry needed to calibrate policy responses.
Accenture recently reported that AI has the potential to double economic growth rates by 2035,
which would make it the engine for our future economy. Having the most powerful economy in
the world will eventually require having the most AI-driven one, and the US accordingly must
lead the development and application of AI technologies along the way. The best way to ensure a
good future is to invent it.
Thank you for your time and focus on this critical topic. I am pleased to address any questions.
[1] National Science and Technology Council, Networking and Information Technology
Research and Development Subcommittee. 2016. “The National Artificial Intelligence Research
and Development Strategic Plan” report: https://www.whitehouse.gov/sites/default/files/
whitehouse_files/microsites/ostp/NSTC/national_ai_rd_strategic_plan.pdf
[2] Furman, Jason. 2016. “Is This Time Different? The Opportunities and Challenges of Artificial
Intelligence” report: https://www.whitehouse.gov/sites/default/files/page/files/
20160707_cea_ai_furman.pdf
[3] Amodei, Dario et al. 2016. “Concrete Problems in AI Safety” research paper: https://
arxiv.org/abs/1606.06565
[4] Executive Office of the President, National Science and Technology Council Committee on
Technology. 2016. “Preparing for the future of artificial intelligence” report: https://
www.whitehouse.gov/sites/default/files/whitehouse_files/microsites/ostp/NSTC/
preparing_for_the_future_of_ai.pdf