[0:00:00] The Committee on Science, Space, and Technology will come to order. Without objection, the Chair is authorized to declare recesses of the Committee at any time. Good morning and welcome to today's hearing entitled Artificial Intelligence, With Great Power Comes Great Responsibility. I now recognize myself for five minutes for an opening statement. First, I would like to note that one of our witnesses, Dr.
[0:00:32] Jaime Carbonell from Carnegie Mellon University is unable to be here today due to a medical emergency. We wish him well and a speedy recovery. Without objection, we'll ensure his written testimony is made part of the hearing record. One of the reasons I've been looking forward to today's hearing is to get a better sense from our witnesses about the nuances of the term artificial intelligence and implications for our society in a future where AI is ubiquitous.
[0:01:02] Of course, one might say AI is already pervasive. Since the term was first coined in the 1950s, we have made huge advances in the field of artificial narrow intelligence, which has been applied to many familiar everyday items such as the technology underlying Siri and Alexa. Called ANI for short, such systems are designed to conduct specific and usually limited tasks. For example, a machine that excels at playing poker
[0:01:30] Wouldn't be able to parallel park a car. Conversely, AGI or artificial general intelligence refers to intelligent behavior across a range of cognitive tasks. If you enjoy science fiction movies, this definition may conjure up scenes from any number of classics such as Blade Runner, The Matrix, or The Terminator.
[0:01:51] For many individuals, the term AGI invokes images of robots or machines with human intelligence. As it turns out, we are decades away from realizing such AGI systems. Nevertheless, discussions about AGI in a future in which AGI is commonplace lead to some interesting questions worthy of analysis. For example, Elon Musk has been quoted as saying that AI, quote,
[0:02:18] It's a fundamental risk to the existence of human civilization and poses vastly more risk than North Korea. Does that mean that AGI may evolve to a point one day when we will lose control over machines of our own creation? As far-fetched as that sounds, minds and scientists are certainly discussing such questions.
[0:02:40] For the short term, however, my constituents are concerned about less existential issues that usually accompany new and evolving technologies, topics such as cybersecurity, protecting our privacy, and impacts to our nation's economy and to jobs. I am an original co-sponsor of a bill introduced earlier this year titled the AI Jobs Act of 2018 to help our workplace prepare for the ways AI will shape the economy of the future.
[0:03:09] I will also introduce legislation today to reauthorize the National Institute of Standards and Technology, which includes language directing NIST to support development of artificial intelligence and data science. There is immense potential for AGI to help humans.
[0:03:26] Thank you, Chairwoman Comstock.
[0:03:57] Thank you, Chairman Weber, for holding this hearing to understand the current state of artificial intelligence technology. Because of the rapid development of computational power, the capacity of AI to perform new and more complicated tasks is quickly advancing. Depending on who you ask, AI is the stuff of dreams or nightmares. I believe it is definitely the former, and I strongly fear that it could also be the latter. The science fiction fantasy world's
[0:04:27] Depicted on Hollywood's big and small screens alike, capture imaginations about what the world might be like if humans and highly intelligent robots shared the earth. Today's hearing is an opportunity to begin to understand the real issues in AI and to begin to move forward with informed science-based policymaking. This is a hearing that we may remember years from now, hopefully as a bright beginning of a new era.
[0:04:57] Current AI technologies touch a broad scope of industries and sectors, including manufacturing, transportation, energy, healthcare, and many others. As we will hear from the witnesses today, artificial intelligence can be classified as artificial general intelligence or artificial narrow intelligence. From my understanding, it is applications of the latter, such as machine learning, that are underlying technologies that support some of the services and devices widely used by Americans today.
[0:05:29] These include virtual assistants such as Siri and Alexa, translation services such as Google Translate, and autonomous vehicle technologies. As the capabilities of AI improve, it will undoubtedly become a more essential part of our lives and our economy. While technology developers and industry look forward to making great strides in AI, I want to make sure my colleagues and I in Congress are asking the tough questions
[0:05:58] And carefully considering the most crucial roles that the federal government may have in shaping the future of AI. Federal investments in AI research are long standing. And we must consider the appropriate balance and scope of federal involvement as we begin to better understand the various roles AI will play in our society. We are not starting from scratch in thinking about the appropriate role of the federal government in this arena. In 2016,
[0:06:28] The White House issued the National Artificial Intelligence Research and Development Strategic Plan that outlines seven priorities for federally funded AI research. These included making long-term investments in AI, developing effective methods for human-AI collaboration, and addressing the ethical, legal, and societal implications of AI. Additional issues to address are safety and security, public data sets, standards,
[0:06:57] And workforce needs. Earlier this year, the Government Accountability Office issued a technology assessment report led by one of our witnesses, Dr. Pearsons, titled Artificial Intelligence, Emerging Opportunities, Challenges, and Implications. While noting significant potential for AI to improve many industries, including finance, transportation, and cybersecurity, the report also noted areas where research is still needed, including how to optimally regulate AI,
[0:07:29] How to ensure the availability and use of high quality data, understanding AI's effects on employment and education, and the development of computational ethics to guide the decisions made by software. These are all critical issues, but more and more I hear concern and widely varying predictions about AI's impact on jobs. AI has potential to make some job functions safer and more efficient, but it also may replace others.
[0:08:01] We need to ask, what are the long-term projections for the job market as AI grows? In this context, we also need to ask, how well do our AI capabilities compare to those of other countries? What education, skills, and retraining will the workforce of the future need? These are very important questions as we think about ensuring a skilled workforce for the future that will help solidify U.S. leadership in AI as other countries vie for dominance in the field.
[0:08:31] If AI threatens some careers, it likely creates many others. We need to consider what Congress should do to shape this impact and make sure Americans are ready for it and make sure the benefits of AI are distributed widely. One other obvious issue of major concern when it comes to AI is ethics. There are many places where this becomes relevant. Currently, we need to grapple with issues regarding the data that are being used to educate machines.
[0:09:01] Biased data will lead to biased results from seemingly objective machines. A little further down the line are many difficult questions being raised in science fiction about a world of humans and intelligent robots. These are questions we'll likely be called on to deal with in Congress and we need to be ready. I want to thank all of our witnesses for being here today and I look forward to your testimony. I'll yield back.
[0:09:29] Thank you, Mr. Lipinski, and I now recognize the chairman of the Energy Subcommittee, the gentleman from Texas, Mr. Weber, for his opening statement. Madam Chair, can I defer to the chairman of the full committee for his statement? Yes, you may. Thank you. Thank you, Madam Chair. Thank you, Mr. Chairman. Didn't know you were going to do that. Madam Chair, often unknown to us, advances in artificial intelligence, or AI, touch many aspects of our lives.
[0:09:58] In the area of cybersecurity, AI reduces our reaction times to security threats. In the field of agriculture, AI monitors soil moisture and targets crop watering. And in the transportation lane, AI steers self-driving cars and manages intelligent traffic systems. Multiple technical disciplines, including quantum computing science, converge to form AI.
[0:10:22] Tomorrow, the science committee will mark up the National Quantum Initiative Act, which establishes a federal program to accelerate quantum research and development. This is a bipartisan bill that ranking member Eddie Bernice Johnson and I and others will introduce today. My hope is that every member of the committee will sponsor it, or at least a majority.
[0:10:43] Transforming our current quantum research into real-world applications will create scientific and technological discoveries, especially in the field of artificial intelligence. These discoveries will stimulate economic growth and improve our global competitiveness, important considerations in light of China's advances in artificial intelligence and quantum computing.
[0:11:04] By some accounts, China is investing $7 billion in AI through 2030 and $10 billion in quantum research. The European Union has also issued a preliminary plan outlining a $24 billion public-private investment in AI between 2018 and 2020. And Russian President Putin has noted that, quote, the leader in AI will rule the world, end quote.
[0:11:33] No doubt that's appealing to him. Yet the Department of Defense's unclassified investment in AI was only $600 million in 2016, while federal spending on quantum totals only about $250 million a year. The committee will mark up a second piece of legislation to reauthorize the National Institute of Standards and Technology. The bill directs NIST...
[0:11:57] To continue supporting the development of artificial intelligence and data science, including the development of machine learning and other artificial intelligence applications. It is simply vital to our nation's future that we accelerate our quantum computing and artificial intelligence efforts. Thank you, Madam Chair, and you're back. Thank you. And I now recognize, let's see where we are, the ranking member of the Energy Subcommittee, the gentleman from Texas, Mr. Vesey, for an opening statement.
[0:12:28] I want to thank you Chairwoman Comstock and Chairman Weber for holding this hearing today and thank you for all of the witnesses for providing expertise on this topic. I'm looking forward to hearing what everyone has to say today. America of course is a country of innovation and in the digital world of today more and more industries are relying on advanced technologies and connectivity to overcome new challenges. Artificial intelligence and big data are impacting every facet of production and commerce.
[0:12:58] AI has the ability to mimic cognitive functions such as problem solving and learning, making it a critical resource as we encounter never before seen problems. Those in the energy sector have already seen improvements in productivity and efficiency and can expect to see even more advancement in the coming years. AI can be used to process and analyze data in previously unexplored ways. Technologies such as sensor equipped
[0:13:25] Aircraft engines, locomotive, gas, and wind turbines are now able to track production efficiency and wear and tear on vital machinery. AI could also significantly improve our ability to detect failures before they occur and prevent disasters saving money, time, and lives. And through the use of analytics, sensors, and operational data, AI can be used to manage, maintain, and optimize systems ranging from energy storage components to power plants.
[0:13:54] As digital technologies revolutionize the energy sector, we must ensure safe and responsible execution of these processes. AI systems can learn and adapt through continuous modeling of interaction and data feedback. Production must be put in place to guarantee the integrity of these mechanisms as they evaluate mass quantities of machine and user data. With Americans' right to privacy under threat,
[0:14:21] Security of these connected systems is of the utmost importance. Nevertheless, I'm excited to learn about the valuable benefits that AI may be able to provide for our economy and our well-being alike, with a Gartner research study reporting that AI will generate 2.3 million jobs by 2020. That's a lot of jobs. The growth AI will bring not only to the energy sector, but to health care, transportation, education, and so many others will help
[0:14:50] I look forward to seeing what light our witnesses can shed on these topics and what we can do in Congress to help enable the development and deployment of these promising technologies. Madam Chairwoman, I yield back the balance of my time. Thank you, and I now recognize Mr. Weber for his opening statement. Thank you, Madam Chair. Today we will hear from a panel of experts on next-generation artificial intelligence, AI, as we've all heard it described.
[0:15:19] And while some have raised concerns about the negative consequences of AI, this technology has the potential to solve fundamental science problems and improve everyday life. In fact, it's likely that everyone in this room benefits from artificial intelligence. For example, users of voice assistance, online purchase prediction, fraud detection that the gentleman from Texas mentioned,
[0:15:47] And music recommendation services are already utilizing aspects of this technology in their day-to-day life. In the past few years, the use of AI technology has rapidly expanded due to the increase in the volume of data worldwide and to the proliferation of advanced computing hardware that allows for the powerful parallel processing of this data.
[0:16:11] The field of AI has broadened to include other advanced computing disciplines such as machine learning we've heard about, neural networks, deep learning computer vision, and natural language processing, just to name a few. These learning techniques are key to the development of AI technologies and can be used to explore complex relationships and produce previously unseen results on unprecedented time scales. The Department of Energy, DOE,
[0:16:41] is the nation's largest federal supporter of basic research in the physical sciences with expertise in big data science, high performance computing, advanced algorithms, and data analytics, and is uniquely positioned to enable fundamental research in AI and machine learning. DOE's Office of Science Advanced Scientific Computing Research Program, or OSCQR as we call it, program develops
[0:17:11] Next generation supercomputing systems that can achieve the computational power needed for this type of critical research. This includes the department's newest and most powerful supercomputer called Summit, which just yesterday, just yesterday, was ranked as the fastest computing system in the entire world. AI also has broad applications in the deal mission space.
[0:17:36] In material science, AI helps researchers speed the experimental process and discover new compounds faster than ever before. In high energy physics, AI finds patterns in atomic and particle collisions previously unseen by scientists. In fusion energy research, AI modeling predicts plasma behavior that will assist in building tokamak reactors, making the best of our investments in space.
[0:18:05] Even in fossil fuel energy production, AI systems will optimize efficiency and predict needed maintenance at power generating facilities. AI technology has the potential to improve computational science methods for any big data problem, any big data problem. And with the next generation of supercomputers, the exascale computing systems that DOE is expected to fill by 2021,
[0:18:34] American researchers utilizing AI technology will be able to track even the bigger challenges. We cannot afford to fall behind in this compelling area of research. And big investments in AI by China and Europe already threaten US dominance in this field. With the immense potential for AI technology to answer fundamental scientific challenges, it's quite clear we should prioritize this research.
[0:19:03] We should maintain, I will add, American competitive edge and American exceptionalism. This will help us to do that. I want to thank our accomplished panel of witnesses for their testimony today, and I look forward to hearing what role Congress can play and should play in advancing this critical area of discovery science. Madam Chair, I yield back. Thank you, and I will now introduce today's witnesses.
[0:19:28] Our first witness today is Dr. Tim Persons, Chief Scientist at the U.S. Government Accountability Office. He also serves as the Director for GAO's Center for Science, Technology and Engineering. Dr. Persons received a Bachelor of Science in Physics from James Madison University and a Master of Science in Nuclear Physics from Emory University. He also earned a Master of Science in Computer Science and PhD in Biomedical Engineering.
[0:19:56] Both from Wake Forest University. Next, we have Mr. Greg Brockman, our second witness, who is co-founder and chief technology officer of OpenAI, a nonprofit artificial intelligence research company. Mr. Brockman is an investor in over 30 startups and a board member of the Stellar Digital Currency System. He was previously the CTO of Stripe, a payments startup now valued at over $9 billion.
[0:20:26] And he studied mathematics at Harvard and computer science at MIT. And our final witness is Dr. Fei-Fei Li, chairperson of the board and co-founder of AI for All. In addition, Dr. Li is a professor in the computer science department at Stanford and the director of the Stanford Artificial Intelligence Lab. In 2017, Dr. Li also joined Google Cloud as chief scientist of AI and machine learning.
[0:20:55] Dr. Lee received her bachelor of arts degree in physics from Princeton and her PhD in electrical engineering from the California Institute of Technology. I now recognize Dr. Persons for five minutes to present his testimony. Good morning. Thank you, Chairwoman Constock, Chairman Weber, ranking members Lipinski and Veazey, and members of the subcommittee.
[0:21:15] I'm pleased to be here today to discuss GAO's technology assessment on artificial intelligence. To ensure the US remains a leader in AI innovation, special attention will be needed for our education and training systems, regulatory structures, frameworks for privacy and civil liberties, and our understanding of risk management in general. AI holds substantial promise for improving human life, increasing the nation's economic competitiveness,
[0:21:44] And solving some of society's most pressing challenges. Yet, as a disruptive technology, AI poses risks that could have far-reaching effects on, for example, the future labor force, economic inclusion, and privacy and civil liberties, among others. Today, I'll summarize three key insights arising from our recent work. First, the distinction between narrow versus general AI. Second, the expected impact of AI on jobs.
[0:22:11] Competitiveness and workforce training. And third, the role the federal government can play in research, standards development, new regulatory approaches, and education. Regarding narrow versus general AI. Narrow AI refers to applications that are task specific, such as tax preparation software, voice and face recognition systems, and algorithms that classify the content of images. General AI refers to a system exhibiting intelligence on par with
[0:22:41] Thank you.
[0:23:07] Regarding jobs competition in the workforce, there is considerable uncertainty about the extent to which jobs will be displaced by AI and how much any losses will be offset by job creation. In the near term, displacement to certain jobs such as call center or retail workers may be particularly vulnerable to automation. However, in the long term, demand for skills that are complementary to AI is expected to increase.
[0:23:35] To better understand the impact of AI on employment moving forward, several experts underscored the need for new data and methods to enable greater insight into this issue. Regarding the role of the federal government, it will continue its crucial role in research and data sharing, contributions to standards development, regulatory approaches, and education. One important research area the federal government could support is enhancing the explainability of AI.
[0:24:05] Which could help establish trust in the behavior of AI systems. The federal government could also incentivize data sharing, including federal data that are subject to limitations for how they can be used. As well as creating frameworks for sharing data to improve the safety and security of AI systems. Such efforts may include supporting standards for explainability, data labeling and safety, including risk assessment and benchmarking of AI.
[0:24:31] It's always risk versus risk. Related to this, new regulatory approaches are needed, including the development of regulatory sandboxes for testing AI products, services, and business models, especially in industries like transportation, financial services, and healthcare.
[0:24:50] GAO's recent report on FinTech found, for example, that regulators use sandboxes to gain insight into key questions, issues, and unexpected risks that may arise out of the emerging technologies. New rules governing intellectual property and data privacy may also be needed to manage the deployment of AI. Finally, education and training will need to be reimagined so workers have the skills needed to work with and alongside emerging AI technologies.
[0:25:19] For the U.S. to remain competitive globally and effectively manage AI systems, its workers will need a deeper understanding of probability and statistics across most, if not all, academic disciplines. That is, not just the physical, engineering, and biological sciences, as well as competency in ethics, algorithmic auditability, and risk management.
[0:25:40] In conclusion, the emergence of what some have called the fourth industrial revolution and AI's key role in driving it will require new frameworks for business models and value propositions for the public and private sectors alike. Even if AI technologies were to cease advancing today, no part of society or the economy would be directly or indirectly untouched by its transformative effects.
[0:26:03] I thank committee leaderships, committees, thanks to the members here for your holding a hearing on this very important topic today for such a time as this. Madam Chairwoman, Mr. Chairman, ranking members, this concludes my prepared remarks. I would be happy to respond to any questions that you or other members of the subcommittees have at this time. Thank you, and I now recognize Mr. Brockman for five minutes.
[0:26:28] Chairwoman Comstock, Chairman Weber, Ranking Member Lipinski, Ranking Member Vesey, members of both subcommittees, thank you for having me today to deliver testimony. I'm Greg Brockman, co-founder of OpenAI, a San Francisco-based nonprofit with a mission to ensure that artificial general intelligence, which we define as highly autonomous systems that outperform humans at most economically valuable work, benefits all of humanity. Now, I'm here to tell you about the generality of modern AI.
[0:26:55] Why AGI might actually be in reach sooner than commonly expected, and what action policymakers can take today. So first, what's OpenAI? We're a research company with one of the world's most advanced AI research and development teams. Yesterday, we announced major progress towards a milestone that we, Alphabet subsidiary DeepMind, and Facebook have separately been trying to reach, which is solving complex strategy games, which start to capture many aspects of the real world that were just not seen in board games like chess or Go.
[0:27:21] We built a system called OpenAI 5, which learned to devise long-term plans and navigate scenarios far too complex to be programmed in by a human in order to solve a massively popular competitive game called Dota 2. Now in the past, AI-like technology was written by humans in order to solve one specific problem at a time. It was not capable of adapting to solve new problems. Today's AI, it's all based on one core technique, which is the artificial neural network.
[0:27:45] A single simple idea that as it's run on faster computers is proving to match a surprising amount of human capability. And this was in fact something that was shown in part by my fellow witnesses, Dr. Lee's work in image recognition. And artificial neural networks can be trained to perform speech recognition or computer vision. It just depends on the data that they're shown. Now further along the spectrum of generality is AGI.
[0:28:06] Rather than being developed for any one use case, AGI would be developed for a wide range of important tasks. And AGI would also be useful for non-commercial applications, including thinking through complex international disputes or city planning. Now, people have been talking about AGI for decades. And so how should we think about the timeline? Well, all AI systems, they're built on three foundations. That's data, computational power, and algorithms. Next generation AI systems are already starting to rely less on conventional data sets where a human has provided the right answer.
[0:28:36] For example, one of our recent neural networks learned by reading 7,000 books. We also recently released a study showing that the amount of computation powering the largest AI training runs has been doubling every three and a half months since 2012. That's a total increase of 300,000 times. And we expect this to continue for the next five years using only today's proven hardware technologies and not assuming any breakthroughs like quantum or optical. Now to put that into perspective,
[0:29:05] That's like if your phone battery, which today lasts for a day, started to last for 800 years, and then five years later started to last for 100 million years. It's this torrent of compute, this tsunami of compute. We've never seen anything like this. And so the open question is, will this massive increase in computational power combined with near-term improvements in algorithmic understanding be enough to develop AGI? We don't know the answer to this question today, but given the rapid progress that we are seeing, we can't confidently rule it out.
[0:29:36] And so now, what should we be thinking about today? What can policymakers be doing today? And so the first thing to recognize is the core danger of AGI is that it has fundamentally the potential to cause rapid change, whether that's through machines pursuing goals that are misspecified by their operator, whether it's through malicious humans subverting deployed systems, or whether it's an economy that grows in an out-of-control way for its own sake rather than in order to improve human lives. Now, we spent two years' worth of policy research
[0:30:04] To create the OpenAI charter, which in fact is a document I have right here in front of me, this contains three sections defining our views on safe and responsible AGI development. So that's one is leaving time for safety, and in particular, refusing a race to the bottom on safety in order to reach AGI first. The second is to ensure that people at large, rather than any one small group, receive the benefits of this transformative technology. And the third is working together as a community in order to solve safety and policy challenges.
[0:30:33] Now, our primary recommendation to policymakers is to start measuring progress in this field. We need to understand how fast the field is moving, what capabilities are likely to arrive when, in order to successfully plan for AGI challenges. That moves towards forecasts rather than intuition. Measurement's also a place where international coordination's actually viable, and this is important if we want to spread safety and ethical standards globally. So thank you for your time, and I look forward to questions. Thank you, and we now recognize Dr. Lee.
[0:31:02] Thank you for the invitation, Congresswoman and Congressman. My name is Fei-Fei Li. I'm here today as the co-founder and chairperson of AI for All, a national nonprofit organization focusing on bringing hands-on experience in AI research to high school students that have been traditionally underrepresented in the STEM fields, such as girls, people of color, and members of low-income communities. Our program began at Stanford.
[0:31:32] AI for all are expanded across North America to six university campuses.
[0:31:54] It has an impact on people. It's a powerful tool we're only just beginning to understand. And that's a profound responsibility. I'm here today because the time has come to have an informed public conversation about that responsibility. With proper guidance, AI will make life better. But without it,
[0:32:16] It stands to widen the wealth divide even further, make technology even more exclusive, and reinforce biases we've spent generations trying to overcome. This will be an ethical
[0:32:30] Philosophical and humanistic challenge and it will require a diverse community of contributors. It's an approach I call human-centered AI and it's made up of three pillars that I believe will help ensure AI plays a positive role in the world.
[0:32:50] The first is that the next generation of AI technology must reflect more of the qualities that make us human, such as deeper understanding of the context we rely on to make sense of the world. Progress on this front will make AI much better at understanding our needs, but will require a deeper relationship between AI and fields like neuroscience, cognitive science, and the behavior sciences. The second is that
[0:33:18] Emphasis on enhancing and augmenting human skills, not replacing them. Machines are unlikely to replace nurses and doctors, for example, but machine learning assistive diagnosis will help their job tremendously. Similar opportunities to intelligently augment human capabilities abound, from healthcare to education, from manufacturing to agriculture.
[0:33:45] Finally, AI must be guided by a concern for its impact. We must address challenges of machine biases, security, privacy, as well as at the society level. Now is the time to prepare for the effect of AI on laws, ethics and even culture. To put these ideas in practice, governments, academia and industry will have to work together.
[0:34:12] This will require better understanding of AI in all three branches of government. AI is simply too important to be owned by private interests alone, and publicly funded research and education can provide a more transparent foundation for its development. Next, academia has a unique opportunity to elevate our understanding and development of this technology. Universities are
[0:34:39] Perfect environment for studying its effect on our world, as well as supporting cross-disciplinary next generation AI research. Finally, businesses must develop a better balance between their responsibility to shareholders and their obligations to their users. Commercial AI products have the potential to change the world rapidly, and the time has come to complement this ambition with ethical, socially conscious policies.
[0:35:09] Human-centered AI means keeping humans at the heart of this technology's development. Unfortunately, lack of diverse representation remains a crisis in AI. Women hold a fraction of high-tech positions and even fewer at the executive level. And this is even worse for people of color. We have good reasons to worry about bias in our algorithms.
[0:35:36] A lack of diversity among the people developing these algorithms will be among its primary causes. One of my favorite quotes comes from technology ethicist Shannon Valer, who says that there is no independent machine values. Machine values are human values. However autonomous our technology becomes, its impact on the world will always be our responsibility.
[0:36:03] With a human-centered approach, we can make sure it's an impact we'll be proud of. Thank you. Thank you. And I now recognize myself for five minutes of questions. Dr. Lee, there is a generally accepted potential for AI-enabled teaching to a minimum, provide a backup for traditional classroom education. So as a...
[0:36:34] AI technology advances. It seems reasonable to assume that your traditional education, vocational training, homeschooling, and even college coursework will need to change and adapt. Could you maybe comment about how education in general and for specific groups and individuals might be transformed by AI and how we can make that positive and really sort of have more of a democratization of education, particularly
[0:37:03] Higher education and stamina in science.
[0:37:05] Thank you for the question. Of course, I feel passionate about education. So I want to address this question from two dimensions. One is how could we improve the education of AI and STEM in general to more students and general community? Second is what can AI as a technology do to help education itself? On the first dimension, as our work in AI for All, we really believe that it's
[0:37:35] It's simultaneously a crisis and an important opportunity that we involve more people in the development of AI technology. AI represents... Humanity has never created a technology so similar or trying to resemble who we are. And we need AI to...
[0:37:57] Technologists and leaders of tomorrow to represent this technology. So personally, I think we need to democratize AI's education to reach out to more students of color, girls, women of traditionally underrepresented minority. At AI for All for the past four years, we've...
[0:38:19] Already created an alumni population of more than 100 students, and through their own community and outreach effort, we have been touching lives of more than 1,400 youth, ranging from middle schoolers to high schoolers, in disseminating this AI knowledge. And we need more of that in higher education. The second dimension that I want to answer your question is AI as a technology itself
[0:38:46] can help improve education itself. In the machine learning community, I'm sure, Greg, you also agree with me that
[0:38:54] There's an increasing recognition of the opportunity for lifelong learning using technology as an assistive technology. I have colleagues at Stanford who focuses on research in reinforcement learning and education, how to bring more technological assistance into the teaching and tutorialization of education itself. And I think this
[0:39:21] Could become a huge tool, as I was saying, to augment human teachers and human educators so that our knowledge can reach to more students and wider community. Excellent, excellent. And for other witnesses, could you maybe comment on how academic institutions and industry could work with government on AI?
[0:39:51] For OpenAI's recommendation, it's really about starting with measurement, right? To really start to understand what's happening in the field. I think it's really about, for example, the study that we did showing the 300,000 times increase. We need more of that. We need to understand where things are going, where we are. I think the government is uniquely positioned to set some of the goalposts as well. We've been pretty encouraged by seeing some of the work that is happening at GAO and also DIUX has had some success with this. So we think it's really about...
[0:40:19] Start in a low touch way for the dialogue to start happening, because I think right now the dialogue is not happening to the extent that it should. Thank you for the question. I do think that, as the committees have pointed out, this is a
[0:40:33] Whole of society issue. It's going to be government in partnership with the private sector, with academia to look at things. And so I think there is room for thought about how to learn by doing, creating internships and ways to try and solve real world problems so that you have a mix of the classroom experience as well as making, building. You'll fail a lot, of course, with these things, but learning in a safe environment and then being able to grow expertise.
[0:41:03] Thank you. And Dr. Lee, did you have anything you wanted to add to that also? Okay. Well, thank you, and I now will recognize Mr. Lipinski for five minutes. Thank you. This is a fascinating topic, and I want to try to move through some things quickly, but get some good answers here. It seems to me that
[0:41:33] Mr. Brockman, you have a different view of AGI, the possibility of AGI and how quickly it can come than the GAO report. Is there a reason for this? Is there something you think that GAO is missing? And if Dr. Pearsons could respond to that. So I don't know if I can comment directly on the report, just not being familiar enough with all the details in there, but I can certainly comment on
[0:42:04] Our perspective on AGI and its possibility. And a lot of it really comes down to rather than, you know, I think that there's been a lot of more emotion or intuition-based argument. And to your opening remarks, you know, I think that science-based reasoning in order to project what's happening in this field is extremely important. And that's something that we've spent quite a lot of effort on since starting OpenAI almost three years ago. And so looking at the barriers to progress as compute data algorithms.
[0:42:31] Data is something that's changing very rapidly in terms of what data we can use. The computation, the power there is growing at a rate that we've just never seen over the course of this decade. We're going to be talking, I think, about 10 orders of magnitude. And that's something where, if you were to compare that to the typical growth of compute, something like Moore's law, over the period where we saw 300,000x increase in the past six years, we would have only seen 12x.
[0:42:55] All right, that's a huge gap and that this is somewhere where we're sort of being projected into the future a lot faster than people realize. Now it doesn't mean that it's going to happen soon. It means that we can't rule it out. It means that for the next five years, as long as this hardware growth is happening, we're in a fog and it's hard to make confident projections. And so my position is that we can't rule it out. We know that this is, you know, we're talking about a technological.
[0:43:18] Revolution on the scale of the agricultural revolution, something that could be so beneficial to everyone in this world. And if we aren't careful in terms of thinking ahead and trying to be prepared, we could be caught unaware. Thank you. Dr. Pearsons, do you have a response on that?
[0:43:36] Sure. I think, and with all respect for our Silicon Valley innovators who are upstarts and challenge the SASCO, I think it's great that we have the system. The key thing that we're seeing is the convergence of these technologies that was mentioned by my panelists of the exponential power in computing, the ubiquitous nature of data, the sophistication of algorithms are all coming in. But that said, many folks in the community are
[0:44:04] Mildly skeptical about the rate at which general AI may come in this area because for several reasons. One is just the way that we think about the problem now, the super complexity that is manifest in addressing the various challenges or looking at larger data sets and looking at all the facets of them. It's much easier to say than to do. And again, I think a lot of the, as you pointed out, the driving force here is
[0:44:31] The concern about general AI and taking over the world kind of thing. And it's just much harder to mimic human intelligence, especially in an environment where intelligence isn't even really defined or understood. And I think, as Dr. Lee pointed out, a lot of this is really about augmentation. It's something else we heard from our experts. It wasn't a replacement of humans. It was a how can we become better humans, more functional humans, in doing these things.
[0:44:58] A lot of it just gets down to the- Let me extend it a short time, sir. Thank you. I just want to throw out quickly, there have been very different, vastly different opinions about the replacement of jobs and the disappearance of jobs and what the impact's going to be. Mr. Brocken, what do you think the impact will be?
[0:45:26] With new technologies in the short term, we always overestimate the degree to which they can make rapid change. But I think in the long term that they do. I think technology is changing that we've seen with things like the internet, that there's been a lot of job displacement, both creation and destruction. And I think AI will be no different. I think the question of exactly which jobs and when, I think we don't have enough information yet. And I think that that's where measurement really starts to come in. So we view it as an open question and a very important one.
[0:45:56] And, sir, I'm going to just say as a bottom line, nobody really knows the impact on this. And, of course, our experts were saying to know more, we might need to be able to encourage, let's, for example, our Bureau of Labor Statistics data type agency out of the federal government to help provide more data or different data or things to help try and answer the question of what is the impact as this technology continues to unfurl.
[0:46:22] That said, there's also a history of, it goes back and attributed to Ned Ludd in the era of British industrialization and the concern of destroying the machines for the concern about loss of jobs. And yet, in many times throughout history, it's happened in an array of technologies where net jobs actually increased. It just, they were more sophisticated jobs. They were toward higher value creation and more productivity. So there is hope for this technology as well.
[0:46:52] Chair, when we allow, I want to hear from Dr. Lee.
[0:46:55] I just want to say that technology inevitably throughout human civilization has an impact to change the landscape of jobs. But it's really, really critical, like my fellow panelists said, that we need to invest in the research of how to assess this change. It's not a simple picture of replacement, especially when this technology has a much greater potential and power to augment it. I just spent days in the hospital ICU with my mother.
[0:47:25] And with my own healthcare and AI research, you recognize that a nurse in a single shift is doing hundreds of different tasks in our ICU unit where they're fighting for life and death for our patients. And these are not a simple question of replacing jobs, but creating better technology to assist them and to make their jobs better and make the lives better for everyone. And that's what I hope we focus on using.
[0:47:54] Thank you. Thank you, Dr. Lee. That's a wonderful example of really vividly explaining to us how that can be used. Because certainly, as we're an aging population in this country, that's a challenge we're all facing. And so the quality of life and improvement in each of those employees and nurses being able to do a better job, thank you for outlining that. I now recognize Mr. Weber.
[0:48:23] Thank you, Madam Chair. Dr. Lee, is your mom okay? We hope that she is, pray and hope that she is okay. Thank you. I'm here. That means she's better. Okay. Otherwise, we were going to be missing two witnesses. Good. She's watching me right now. Well, good. Hi, Mom. She's doing a great job. You're doing excellent. She's a proud mom, and that's some good medicine in and of itself right there. Thank you. So we're glad for that. Dr. Brockman, you in your statement say that
[0:48:51] Your mission was to actually make sure that artificial intelligence benefited people and was better for the most economically valuable work. Do you remember that? It's in your written statement. That's right. So our definition of what AGI will be, whether created by us or anyone else, but just the milestone is a system that can outperform humans at...
[0:49:17] Let me read it to you real quick. I'm Greg Baughman, co-founder and nonprofit development organization mission is to ensure that artificial general intelligence, by which we mean highly autonomous systems, that outperform humans at, quote, most economically valuable work, end quote, benefits all humanity. How would you define most economically valuable work? So I think that, again, and first of all, the question of,
[0:49:47] You know, AGI is something that the whole field has been working towards for, you know, really since the beginning of the field 50 years ago. And so the question of how to define it, I think, is something that is not entirely agreed upon, that our definition is this. And when we think of it, we think of, you think about things like starting companies or very high intellectual work like that. And, you know, also to things like going and cleaning up.
[0:50:14] Disaster sites are things that humans would be unable to do very well today. Okay. Well, I noticed that in your disagreement that Congressman Lipinski referred to with the report, and you call them Silicon Valley upstarts. At least you didn't call them young upstarts. So that's an advantage. Thank you for doing that. But you're literally looking at a new industry that even though the shift, bless you, even though the shift is going to be changing,
[0:50:44] You're actually creating jobs for another industry, and going back to Dr. Lee's example with her mom in the IU, talking about how much the nurses do. How do you train for those jobs if it's moving as fast as you think it is? Yeah, and so one thing I think is also very important is that I don't think that we have much ability to change the timeline to this technology. I think that there are a lot of stakeholders, there are a lot of different pieces of the ecosystem, and that what we do is we step back and we look at the trends and we say what's going to be possible when. And so I think that...
[0:51:14] The question of how to train, again, that's going to be something we're not the only ones that are going to have to help answer that question. But I think that the place to start, it really comes back to measurement. If we don't know what's coming, if we can't project well, then we're going to be taken by surprise. And so I think that there are going to be lots of jobs and already have been created jobs that are surprising in terms of you think about with autonomous vehicles that we need to label all this data. We need to make sure that these systems are doing what we expect and that all of that.
[0:51:42] There's going to be humans that are going to help make these systems. But we would all agree, I hope, and on this question for all three panelists, all three witnesses, that the jobs they're going to create are well worth the transformation into all that technology. Dr. Persons, would you agree with that?
[0:52:00] I would agree with that. Let me give you a quick example, if I may. Speaking with a former secretary of transportation recently, just a simple example of toll booth collectors. We have now a system where you need the easy pass, you drive through, and you have less of a workforce there that did that.
[0:52:18] Could have had an impact at that time for a short period on the number or loss of jobs for towbooth collectors, and yet it freed them up and enabled them to perhaps do other things that were needed in large problems. Mr. Brockman, you were shaking your head. You would agree with that statement. Absolutely. I think the purpose of technology and improvement is to improve people's lives. So Dr. Lee, I see you shaking your head too. Yeah, absolutely. In addition to the example, Mr.
[0:52:43] I think deeply about the jobs that are currently dangerous and harmful for humans, from fighting fires to search and rescue to natural disaster recovery. Not only we shouldn't put humans in harm's way if we can avoid it, but also we don't have enough help in these situations, and this is where technology should be of tremendous help.
[0:53:10] Very quickly, I'm out of time. Just yes or no. If we lose dominance in AI, that puts us in a really bad spot in worldwide comparatives. Would you agree? Yes. Yes. Thank you. Madam Chair, you're back. Thank you. Good question. Now I recognize Mr. Vesey for five minutes. Thank you, Madam Chair. We have heard about already from your testimony.
[0:53:37] Some of the advantages of AI and how it can help humankind, how it can help advance us as a nation and a country. But as you know, there are people also that have concerns about AI. There's been a lot of sort of doomsday-like comparisons about AI and what the future of AI can actually mean. To what extent do you think this scenario, this sort of, you know,
[0:54:04] And this is for anybody on the panel to answer. So I think thinking about...
[0:54:30] Artificial general intelligence today is a little bit like thinking about the internet in maybe the late 50s. If someone was to describe to you what the internet was going to be, how it would affect the world, and the fact that all these weird things were going to start happening. You're going to have this thing called Uber, which you're going to be able to... You'd be very confused. It'd be very hard to understand what that would look like. And the fact that, oh, we'd forget to put security in there. And then we'd be paying for that for 30 years worth of trying to fix things. And now imagine that that whole story, which played out over really the course of the past 60, almost 70 years now,
[0:54:59] What's going to play out in a much more compressed time scale? And so that's the perspective that I have when it comes to artificial general intelligence is the fact that it can cause this rapid change. And it's already hard for us to cope with the changes the technology brings. And so the question of, is it going to be malicious actors? Is it going to be that the technology itself just wasn't built in a safe way? Or is it just that the deployment that who owns it and the values that it's given aren't something that we're all very happy with? All of those, I think, are real risks. And again, that's something that we want to start thinking about today.
[0:55:30] Thank you, sir. So I agree with that. I think the key thing is being clear-eyed about what the risks actually are and not necessarily being driven by the entertaining and yet it's science fiction type narrative sometimes on these things, projecting or going to extremes and assuming.
[0:55:47] Far more than where we actually are in the technology. So there are risks. It's understanding the risks as they are. And they're always contextual risks. Risks in automated vehicles are going to be different than risks in this technology in financial services, let's say. So it's really working, again, symbiotically with the community of practice and identifying what are the things there, what are the opportunities, and there's going to be opportunities, and then what undesirable things do we want to focus on and then optimize.
[0:56:16] From there on how to deal with them. Thank you. Mr. Brockman, in your testimony, you had referenced a report outlined in some malicious actors in this area. Could you sort of elaborate on some of your findings in these areas? That's right. So OpenAI was a collaborator on this research report, projecting not necessarily today what people are doing, but looking forward, what are some of the malicious activities that people could use AI for.
[0:56:45] So that report, let's see, I think maybe the most important things here, you start thinking about things around information, privacy, the question of how do we actually ensure that these systems do what the operator intends despite potential hacking. You think about autonomous systems that are taking action on behalf of humans that are subverted and whether, again, it's, you know, that this report focuses on.
[0:57:17] On active action, you think about autonomous vehicles. And if a human hacker can go and take control of a fleet of those, some of the bad things that could happen. And so I think that this report should really be viewed as we need to be thinking about these things today before these are a problem, because a lot of these systems are going to be deployed in a large scale way. And if we're able to subvert them, then we did.
[0:57:39] All of the problems that we've seen to date are going to start having a very different flavor where it's not just privacy anymore. It's also systems that are deployed in the real world that are actually able to affect our own well-being. Thank you, Madam Chair. I yield back. Thank you. And I now recognize, let's see, Mr. Rohrabacher. Thank you very much, Madam Chairman. This, as in all advances in technology, can be seen as a...
[0:58:10] The great hope for making things better or the new idea that there might be new dangers involved and or the new technologies will help certain peoples but be very damaging to others. And I think that where that fear would be most recognizable is in terms of employment and how in a free society people earn a living.
[0:58:38] And are we talking about here about the development of technology that will help get the tedious and remedial or the lower skilled jobs that can be done by machine? Or are we talking about
[0:59:05] The loss of employment by machines that are designed to really perform better than human beings perform in high-level jobs. What are we talking about here?
[0:59:19] I'm still going to use healthcare as an example because I'm familiar with that area of research. If you look at recent studies by McKinsey and other institutions on employment and AI, there is a recognition that
[0:59:38] We need to talk a little more nuance than just an entire job, but the tasks under each job. The technology has a potential to change the nature of different tasks. Again, for example, take nurse, a job of a nurse as an example.
[0:59:58] No matter how rapidly we develop the technology in the most optimistic assessment, it's very hard to imagine that entire profession of nursing would be replaced. Yet, within the nursing jobs, there are many opportunities that certain tasks can be assisted by AI technology. For example, a simple one that costs a lot of time and effort in nursing jobs is
[1:00:24] Charting. Our nurses in our, again, ICU rooms, our patient rooms, spends a lot of time typing and charting into a system, into a computer, while that's time away from patients and other more critical care. So these are the kind of tasks under a bigger job description that we can hope to use technology to assist and augment. So are we talking about robots here or a box that thinks?
[1:00:54] So AI technology is a technology of many different aspects. It's not just robot. In this particular case, for example, natural language understanding and speech recognition, possibly in the form of a voice assistant, would help.
[1:01:16] But maybe delivering of simple tools on the factory floor will be in the form of a small, simple delivery robot. So there are different forms of machines. There are many dangerous jobs that I could see that we'd prefer not having human life put at risk in order to accomplish the goal.
[1:01:42] For example, at nuclear power plants, it would be a wondrous thing to have a robotic response to something that could cause great damage to the overall community, but would kill somebody if they actually went in to try to solve a problem. I understand that, and also possibly with communicable diseases where people need to be treated, but you're putting
[1:02:10] People at great risk for doing that. However, with that said, when people are seeking profit in a free and open society, I would hate to think that we're putting out of work people with lower skills. And we need the dignity of work and of earning your own way. Once we know now, when you take that away, it really has a major impact, negative impact.
[1:02:39] So I want to thank you all for giving us a better understanding of what we're facing on this. And let's hope that we can develop this technology in a way that helps the widest variety of people and not just perhaps a small group that will keep their jobs and keep the money. So thank you very much.
[1:03:04] Thank you, and I now recognize Ms. Bonamici for five minutes. Thank you so much. Thank you to our witnesses. First, I want to note that our nation has some of the best scientists and researchers and engineers in the world, but without stronger investments in research and development, especially long-term foundational research, we risk falling behind, especially in this important area.
[1:03:27] I hope the research continues to acknowledge the socioeconomic aspects as well of integrating AI technologies. In my home state at the University of Oregon, we have the Urbanism Next Center. They're doing some great work bringing together interdisciplinary perspectives, including planning and architecture and engineering and urban design and public administration.
[1:03:50] Public, private and academic sectors to discuss how leveraging technology will shape the future of our communities. The research has been talking about emerging technologies like autonomous vehicles and the implications for equity, health, the economy and the environment.
[1:04:07] Dr. Pearson, can you discuss the value of establishing this type of partnership between industry, academia, and the private sector to help especially identify and address some of the consequences, intended and unintended, of AI as it becomes more prevalent? And I do have a couple more questions. Sure, I'll answer quickly. The short answer is yes. Our experts and what we're seeing is the value in public-private partnerships because, again, it would be a mistake to look at this technology and say,
[1:04:36] Sort of isolated stove pipes, and it would need to be an integrated approach to things. The federal government having its various roles, but key, like you were mentioning at University of Oregon, key academic and research questions. There's many, many things to research and questions to answer. And then of course industry, which has an incredible amount of innovation and thinking and power to drive things forward.
[1:04:57] Terrific. Thank you. Dr. Lee, I have a couple questions. You discussed the labor disruption, and I know that's brought up a couple of times, and the need for retraining. We really have sort of a dual skills gap issue here because we want to make sure there are enough people who have the education needed for the AI industries, but we also are talking about the workers, like you mentioned.
[1:05:18] The workers in toll booths who will be displaced. But with the rapid development of technologies and the changes in this field, what knowledge and skills are the most important for a workforce capable of addressing the opportunities and the barriers to the development? I serve on the Educational Workforce Committee, and this is a really important issue, is how do we educate people to be prepared for such rapid changes?
[1:05:44] AI is fundamentally a scientific and engineering discipline. As an educator, I really believe in more investment in STEM education.
[1:05:56] Early age, we look at, in our experience at AI for All, when we invited these high school students in the age of 14, 15, 16 to participate in AI research, their capabilities and potentials just amazes me. We have high school students who have worked
[1:06:17] In my lab and winning best paper award at this country's best AI academic conferences. And so I believe passionately that STEM education is critical for the future for preparing AI. Thank you. And as everyone on this committee knows, I always talk about STEAM because I'm a big believer in educating both halves of the brain and students who have arts education tend to be more creative and innovative. I also, Dr. Lee, in your
[1:06:46] In your testimony, you talk about how AI engineers need to work with neuroscientists and cognitive scientists to help AI systems develop a more human feel. Now, I know Dr. Carbinol is not here today, but I noted in his testimony, he wrote, AI is the ability to create machines who perform tasks normally associated with human intelligence. I'm sure that was an intentional choice to humanize the machine, but I wanted to ask you, Dr. Lee, about...
[1:07:15] I'm not here to explain, but I have no doubt that was intentional. In your testimony, you talk about the laws that codify ethics. How is this going to be done? Can you go into more depth about how would these laws be done? Who would determine what is ethical? Would it be a combination of industry, government determining standards?
[1:07:39] Set the stage for an ethical development of AI. Yeah, so thank you for the question. I think for a technology as impactful as AI is to human society, it's critical that we have ethical guidelines. And different institutions from government to academia to industry will have to participate in this dialogue together and also by themselves. Are they already doing that, though? You said that they'll have to, but where is...
[1:08:09] Is somebody convening all of this to make sure that there are? So there are efforts. I'm sure Greg can add to that. The industries in Silicon Valley, we're seeing companies starting to roll out AI ethical principles and responsible AI practices. In academia, we see that ethicists and social scientists coming together with technologists holding seminars, symposiums, classes to discuss the ethical impact of AI.
[1:08:38] And hopefully government will participate in this and support and invest in this kind of efforts. Thank you. I see my time has expired. Thank you, Madam Chair. I yield back. Oh, Mr. Chairman, thank you. I thank the gentlelady. The gentlelady from Arizona is recognized for five minutes. Thank you, Mr. Chair. I want to thank the testifiers today. Very interesting subject.
[1:09:07] Something that kind of spurs the imagination about science fiction shows and those type of things. I do have a question on what countries are the major players in AI and where does the U.S. rank in competition with them? And that's to any panelist or all panelists. So today I think that the U.S. actually ranks possibly top of the list.
[1:09:37] I think that there are lots of other countries that are investing very heavily. China's investing heavily. Lots of countries in Europe are investing heavily. DeepMind is a subsidiary of a US company, but located in London. And I think that it's very clear that AI is going to be something with global impacts. And I think the more that we can understand what's happening everywhere and figure out how we can coordinate on safety and ethics in particular, the better it's going to go.
[1:10:01] Yes, thank you for the question. I think wherever there's large amounts of computing, large amounts of data, and a strong desire to innovate and continue to develop again in this sort of fourth industrial revolution that we're moving on, then it drives toward certainly China and then our allies and colleagues in Western Europe and developed worlds. Thank you. And is there, did you want to answer?
[1:10:33] If I could just add that the most important thing to continue to lead in the field, it's really about the talent. And right now, we're doing a great job of bringing all the talent in. At OpenAI, we have a very wide mix of national backgrounds and origins. I think as long as we can keep that up, that we'll be in very good shape. Thank you. And Mr. Chair, I have one more question. And that is, what steps, I think this has been asked in different ways before, but
[1:10:59] What steps are we guarding against espionage from, let's say, you said China is involved in this, and that's basically my question. Espionage, hacking, those type of things. What guidelines are currently taking place and who's preventing this? Is it the private companies themselves? Is government involved? Thank you.
[1:11:26] So one thing that's very atypical about this field is because it really grew out of a very small number of academic labs that the overarching ethos in the field is actually to publish. And so all of the core research and development is actually being shared pretty widely. And so I think that as we're starting to build these more powerful systems, and this is one of the parts of our charter, that we need to start thinking about safety and keeping
[1:11:53] Think about things that should not be shared. And so I think that this is a new muscle that's being built. It's right now kind of up to each company. And I think that that's something that we're all starting to develop. But I think having a dialogue around what's OK to share and what things are kind of too powerful and should be kept private, that's just the dialogue that's starting now.
[1:12:13] And certainly IP, intellectual property protection, is a critical issue. I think of one former director of the National Security Agency mentioned that at the time it was an unprecedented theft of U.S. intellectual property at that time just because it's the blessing and curse of the Internet. The blessing is it's open and the curse is it's open. And so AI is going to, I think, be in that.
[1:12:38] That category, in terms of what's being done in terms of cybersecurity, it is something our experts pointed out and said it is an issue. As this committee well knows, it's easier said than done. And who has jurisdictions in the U.S. federalist system about, particularly the private company and protection of that, the role of the federal government versus the company itself, in an era where, as I think Mr. Brockman's pointed out, it's sort of the big data era where data are the new oil, yet we want to be open at the same time so that we can innovate.
[1:13:07] So managing that dialectical tension is going to be a critical issue, and there is no easy answer. Thank you. Mr. Chair, I yield back. Chair recognizes Ms. Esty for five minutes.
[1:13:23] Thank you, Mr. Chair, and I want to thank the witnesses for this extremely informative and important conversation that we're having here today. I hail from the state of Connecticut, where we see a lot of innovation at UConn, at Yale, at lots of spin-offs on the sort of narrow AI question. But I think for us, really,
[1:13:44] The issue is more about that general AI. And Mr. Brockman, your discussion of the advances, which makes Moore's law look puny in comparison, is really where I want to take this conversation, about Dr. Lee, your discussion, which I think is incredibly important, about diversity. We saw what happened to Lehman Brothers by not being diverse. I am extremely concerned about what the implications are for teaching a, as it were,
[1:14:12] If it's garbage in, it's going to be garbage out. If it's a very narrow set of parameters and thought patterns and life experiences that go into AI, we will get very narrow results out. So first, I want to just get your thoughts on that. And the second is on this broader ethical question. We've looked for many years. I remember back when I was a young lawyer working on
[1:14:35] Bioethical issues. The Hastings Center got created to begin to look at these issues. This committee's been grappling with CRISPR and the implications of CRISPR. I think about this being very similar, that AI has many similar implications for ethical input. So if you can opine on both of those questions and recognize we've got to, you know.
[1:14:56] Two minutes, three minutes left, about both the ethical, whether we need centers to really bring in ethicists as well as technologists, and then the importance of diversity on the technology side so that we get the full range of human experience represented as we're exciting, our exciting new entry into this fourth revolution. Thank you.
[1:15:19] Yes, in fact, thank you for asking that question. Just now when someone is using the term doomsday scenario, to me, I think if we wake up 20 years from now, whatever years, and we see the lack of diversity in our technology and leaders and practitioners, that would be my doomsday scenario. So it's so important and critical to have diversity for the following three reasons, like you mentioned. One is sheer diversity.
[1:15:49] There are jobs that we're talking about. This is a technology that could have potential to create jobs and improve quality of life, and we need all talents to participate in that. Second is innovation and creativity, like you mentioned in Connecticut and other places. We need that kind of broad talent to add in the force of AI development. And the third is really justice and moral.
[1:16:16] And moral values that if we do not have this wide representation of humanity representing the development of this technology, we could have face recognition algorithms that are more accurate in recognizing white male faces. And we could have dangers of biased algorithms making unfair long application decisions. There are many potential
[1:16:46] Pitfalls of a technology that's biased and not diverse enough. Which brings us to this conversation dialogue of ethics and ethical AI. You're right, previous disciplines like nuclear physics, like biology, have shown us the importance of this.
[1:17:10] I don't know if there's a single recipe, but I think the need for centers, institutions, boards, and government committees are all potential ways to create an openness, this dialogue. And again, we're starting to see that, but I think you're totally right. It's critical issues. So I agree completely with my fellow witness.
[1:17:37] Crucial to success here. So actually, so we have a program called OpenAI Scholars where we brought in a number of people from underrepresented backgrounds into the field and provide mentorship and that they're working on projects and spinning up. One thing that we found that I think is very encouraging is actually very easy to take people who do not have any AI or machine learning background and to make them into.
[1:17:59] Extremely productive first class researchers and engineers very quickly. And that's one benefit of this technology being so new and nascent is that in some ways that we're all discovering as we go along too. So becoming an expert, there just isn't that high of a bar. So I think that everyone putting effort in, the places where the expertise is, I think it's on them to make sure that they're also bringing in the rest of the world.
[1:18:23] On the ethical front, that's really core to my organization. That's the reason we exist. We do think that, for example, when it comes to the benefits of who owns this technology, where do the dollars go, we think it belongs to everyone. And so one of the reasons that I'm here is because I think that this shouldn't be a decision that's made just in Silicon Valley. I don't think that the question of the ethics and how this is going to...
[1:18:43] Work should be in the hands solely of people like me. I think that it's really important to have a dialogue. And again, that's something where I hope that that will be one of the outcomes of this hearing. Thank you very much. The gentleman now recognizes Mr. McNerney. Well, I thank the chair for holding this and the ranking member. And I think the witnesses have really very interesting testimony and diverse in its own right. One of the things that I think that's important here with this committee is how
[1:19:16] How does the government react to AI? Do we need to create a specific agency? Does that agency report to Congress or the administration? Those sorts of things I think are very important. Dr. Brockman, you said I think one of the most important things was that we need a measure of AI progress. Do you have a model or some description of what that would look like?
[1:19:41] Yes, I do. Thank you for the question. And so first of all, I don't think that we need to create new agencies for this. I think that existing agencies are well set up for this. And I was actually very encouraged to hear that people are talking about giving NIST a remit to think about these problems. Again, GAO.
[1:19:58] And DIUx are already starting to work on this. For example, DIUx had a satellite imagery data set hosted a public competition. The kind of thing that we think would be great for government to do as well is to have standardized environments where academics and private sector can test robotic approaches, setting up competitions towards specific problems that various agencies and departments want to be solved. All of those, I think, can be done.
[1:20:23] Without any new agency, and I think that that's something that you can both get benefits directly to the relevant agencies, also understand the field, and also start to build ties between private sector and public sector. I'm one of the founders of the Grid Innovation Caucus. What are the most likely areas we'll see positive benefits to the grid to electric grid stability and resiliency? Who would be the best answer? Mr. Pearsons? Sure. Thank you for the question.
[1:20:53] I think one of the ways GAO has done a
[1:20:57] A good deal of work on this issue, but it's just protection of the electrical grid in the cybersecurity dimension. So it's one of our scenarios or profiles that we did in this report, what our experts and what folks were saying. And again, based on the leadership of this committee and the importance of cyber is that it's a without which nothing. AI is going to be a part of cyber moving forward. And so protection of the grid in the cyber dimension is there.
[1:21:26] Also, I think, as the chairman mentioned earlier, the word optimization. So how we optimize things and how algorithms might be able to compute and find optimums faster and better than humans is an opportunity for grid management and production. Thank you. So AI is also going to be used as a cyber weapon against infrastructures or potentially used as a weapon. Is that right?
[1:21:53] There are concerns now when you look at a broad definition of AI and you look at bots now that are attacking networks and doing distributed, what are DDOS or distributed denial of service attacks and things like that, that exists now.
[1:22:07] Unfortunately, in the black hat assumption, you're going to assume that as AI becomes more sophisticated in the white hat sense, so too, unfortunately, the black hat side of things, the bad guys, are going to also become more sophisticated. And so that's going to be the cat and mouse game, I think, moving forward. Another question for you, Dr. Pearsons. In your testimony, you mentioned that there's considerable uncertainty in the jobs impact of AI. Yes. What would you do to improve that?
[1:22:37] Situation.
[1:22:38] Our experts were encouraging specific data collected on this. Again, we have important federal agencies like BLS, Bureau of Labor Statistics, that work on these issues, what's going on in the labor market, for example. And it may just be an update to what we collect, what questions we ask as a government, how we provide that data, which is, of course, very important to our understanding of unemployment metrics and so on.
[1:23:08] There are economists that have thoughts about this. We had some input on that. There's no easy answer at this time, but the idea that there is an existing agency doing that sort of thing is there. The key question is how could we ask more or better questions on this particular issue on artificial systems. Thank you. Dr. Lee, you gave three conditions for progress in AI being positive.
[1:23:35] Do you see any acceptance or general wide acceptance of those conditions? How can we spread the word of that so that the industry is aware of them and the government is aware of them and that they follow those sorts of guidelines? Thank you for asking. Yeah, I would love to spread the word.
[1:23:57] I do see the emergence of efforts in all three conditions. The first one is about more interdisciplinary approach to AI and ranging from universities to industry, we see the recognition of neuroscience, cognitive science to cross-pollinate with AI research. I want to add
[1:24:21] We are all very excited by this technology. But as a scientist, I'm very humbled by how nascent the science is. It's only a science of 60 years old compared to traditionally classic science that's making human lives better every day, physics, chemistry, biology. There is a long, long way to go for AI to realize its full potential to help people. So that recognition really is important, and we need to get .
[1:24:50] You get more research and cross disciplinary research into that. Second is the augmenting human and again a lot of
[1:24:58] Academic research as well as industry startup efforts are looking at assistive technology from disability to helping humans. And the third is what many of us focus on today is the social impact from studying it to having a dialogue to working together through different industry and government agencies. So all three are the elements of human-centered AI approach and I see that happening more and more.
[1:25:28] Thank you. Neil Beck. Chairman, I recognize the gentleman from New York. Nope. Chairman, I recognize the gentleman that's not from New York, Mr. Palmer. Thank you, Mr. Chairman. I'd like to know if AI can help people who are geography challenged. Gentlemen, time has expired. I request that that question and response be removed from the record.
[1:25:58] I do have some questions. In my district, we have the National Computer Forensics Institute, which deals with cyber crime. And what I'm wondering about is with the emergence and evolution of AI, what are we putting in place because of the potential that that creates for committing crime and for solving crime?
[1:26:26] Dr. Persons, do you have any thoughts on that? Well, certainly in one of the areas we did, thank you for the question, one of the areas we did look at in general was just criminal justice. So just the risks that are there in terms of the social risks, making sure the scales are balanced exactly as they ought to be, that justice is blind and so on was the focus of that. However, in terms of the criminal forensics, AI could be a tool that helps us out.
[1:26:55] In a retrospective sense, what happened. But again, it's an augmentation. It's helping the forensic analysts who would know what things look like. And the algorithm would need to, in the machine learning sense of things, would need to learn what the risks might be going forward so that you perhaps could identify things more proactively and perhaps in near or at real time. So that's the opportunity for this. Again, AI as a tool in cyber was a key message we heard moving forward.
[1:27:26] Any thoughts on that? So today, we're already starting to see some of the security problems with the methods that we're creating. For example, there's a new class of attack called adversarial examples, where researchers are able to craft a physical patch that you could print out and put on any object. They'll make a computer vision system think that it's whatever object you want it to be. So you could put that on a stop sign and confuse a self-driving car, for example. So these sorts of...
[1:27:54] Ways of subverting these powerful systems is something we're going to have to solve and gonna have to work on just like we've been working on computer security for more conventional systems. And I think that the way to think about if you could successfully build and deploy an AGI, what that would look like in many ways, it's kind of like the internet in terms of being very deeply integrated in people's lives, but also having this increasing amount of autonomy and representation and taking action on people's behalves. And so you'll have kind of this question of how do you make sure
[1:28:22] First of all, that's something that could be great for security if these systems are well built and have safety in their core and are very hard to subvert. But also if it's possible for people to hack them or to cause them to do things that are not aligned with the value of the operator, then I think that you can start having very large scale disruption. It also concerns me in the context of it was announced a couple of weeks ago that
[1:28:49] The United States plans to form a space core. We know that China has been very aggressive in militarizing space. If you have any thoughts on that discussion of how artificial intelligence will be used in regard to space, communication systems that are highly vulnerable already, I think that there's some additional vulnerability that would be created. Any thoughts on that?
[1:29:20] And any one of the three of the panelists. Yes, sir. So in terms of the risk in space, obviously one of the key concerns for AI is weaponization, which I think is part of that, and so much less the space domain or any other one. And so I know our Defense Department has key leadership thinking on this and working strategically on how do we operate in an environment where we have to assume there's going to be the adversary might not operate in the ethical.
[1:29:51] A framework that we do and to defeat that. But there's no simple answer at this time other than our Defense Department is thinking about it and working on it. And he's not here, obviously, to testify. But in Dr. Carbonell's testimony, he made a statement that we need to produce more AI researchers, especially more US citizen or permanent resident AI researchers. And I think that kind of plays into
[1:30:20] That issue of how do we deal with AI in space, that's one of the reasons why I've been pushing for a college program like an ROTC program to recruit people into the space core in these areas. Start identifying students when they're maybe even in junior high and scholarship them through college to get them into these positions. Any thoughts on that?
[1:30:51] I'll just answer quickly and just say I think, as Dr. Lee has, I think, elegantly pointed out before, this is really an interdisciplinary thing. I think there's going to be a need for sort of the STEM, STEAM specialists that particularly focus on this, but I think any particular vocation is going to be impacted in one way or the other. Just like you could imagine rewinding a couple decades or a few decades, I'll date myself, but when the advent of the personal computer or the PC coming in and how that affected now we walk into any vocation and someone's using a PC or something like that.
[1:31:20] And it's not unusual, but at the time you had to learn how to augment yourself or your tasks with that. And I think that will be the case. If I may, Mr. Chairman, just to add this final thought is that we've had to deal with some major hacks, federal government systems that are hacked. And what we're faced with, we're competing with the private sector for the best and brightest in terms of cybersecurity. We're going to find ourselves in the same situation.
[1:31:50] with AI experts, the truly skilled people. And that's why I'm suggesting that we may need to start thinking about how do we recruit these people and get them as employees of the federal government. And that was my thoughts on setting up an ROTC-type program where we recruit people in with scholarship, whether it's for cybersecurity or for AI.
[1:32:17] And with a four or five year commitment to work for the federal government. Because there's going to be tremendous competition. And the federal government has a very difficult time competing for those type people. So with that, Mr. Chairman, I yield back. Now the chair recognizes the gentleman from New York. It's okay, we're patient.
[1:32:39] I thank our respective chairs and ranking members for today's very informative hearing. And welcome and thanks to our witnesses. I'm proud to represent New York's 20th congressional district, where our universities are leading the way in artificial intelligence research and education initiatives. SUNY Polytechnic Institute is currently the home of groundbreaking research developing neuromorphic circuits, which could be used for deep learning, such as pattern recognition.
[1:33:07] But are also useful for AI or machine learning.
[1:33:10] In addition, the Institute has established an ongoing research program on restive memory devices. Rensselaer Polytechnic Institute, RPI, is pushing the boundaries of artificial intelligence in a few different areas. In the healthcare front, RPI is focusing on improving people's lives and patient outcomes by collaborating with Albany Medical Center to improve the performance of their emergency department by using AI and analytics to reduce the recurrence of costly ER visits by patients.
[1:33:39] And RPI researchers are also collaborating with IBM to use the Watson computing platform to help people with prediabetes avoid developing the disease. In our fight to combat climate change and protect our environment, researchers at RPI and Earth and Environmental Science are working with computer science and machine learning researchers to apply cutting edge AI to climate issues.
[1:34:04] In the education space, RPI is exploring new ways to use AI to improve teaching, as well as new approaches to teaching AI and data science to every student at RPI. With all that being said, there are tremendous universities across our country that are excelling in AI research and education. And what are some of the keys to helping AI institutions like them to excel? What do we need to do? What would be the most important?
[1:34:32] That's to any one of our panelists. So thank you for asking that question. I think just like we recognize AI really is such a widespread technology, that I think one thing to recognize is that
[1:34:51] It is still so critical to support basic science research and education in our universities. This technology is far from being done. Of course, the industry is making tremendous investment and effort into AI, but it's a nascent science. It's a nascent technology. We have many unanswered questions, including the socially relevant AI, including AI for good.
[1:35:18] Including AI for education, health care, and many other areas. So one of the biggest things I see would be investment into the basic science research into our universities and encouraging more students thinking in.
[1:35:36] Interdisciplinary terms, taking courses. They can be STEM students, STEAM students. AI is not just for engineers and scientists. It could be for students with policy making mind, for students with law interests, and so on. So I hope to see universities participating in this in a tremendous way, like many great schools in New York State. Thank you. Dr. Persons or Mr. Brockman, either of you?
[1:36:05] So I agree with Dr. Lee, but I'll also point out that I think it is also becoming increasingly hard to truly compete as an academic institution.
[1:36:15] Because if you look at what's happening, industry right now is actually doing fundamental research. It's very different from most scientific fields. And that the salary disparity between what you can get at one of these industrial labs versus what you can get in academia, it's very, very large. And there's a second piece, which is in order to do the research, you need access to massive computational resources. And for example, the work that we just did with this .
[1:36:41] Game Breakthrough, that required basically a giant cluster of something around 10,000 machines. And that's something where in an academic setting, it's not clear how you can get access to those resources. And so I think for the playing field to still be accessible, I think that there needs to be some story for how people
[1:36:59] in academic institutions can get access to that. And I think that the question of where is the best research going to be done and where are the best people going to be, I think that's something that it's playing out right now I think in industry's favor, but it's not necessarily set in stone.
[1:37:14] Thank you. Dr. Persons? Yes, sir. Thank you for the question. And I would just add to my fellow panelists the fact that our experts had said that real-world testbeds are important to this. You don't know what you don't know. So not only in addition to adding access to data, but being able to test and do things. One thing for sure, and I learned, in fact, from OpenAI, that a lot of the times these things come out with surprising results. And so that's the whole reason of creating safe environments to try things out.
[1:37:43] We do risk those technologies, and that's something that I think is going to be important to enable that base of research to have an avenue to perhaps move up the technology maturity scale, possibly into the market, and certainly, hopefully, to solve critical, complex real-world problems. Thank you. Very informative. Mr. Chair, I yield back. The chairman now recognizes Jim from Illinois.
[1:38:08] Thank you, Mr. Chairman, and thank you for coming to testify today. I've been interested in artificial intelligence for quite a long time. Back in the 1990s, working in particle physics, we were using neural network classifiers to have a look at trying to classify particle physics interactions.
[1:38:25] When I couldn't stand it during the government shutdown not so long ago, I went and downloaded TensorFlow and worked through part of the tutorial on it. And the algorithms are not so different than what we were using back in the 1990s, but the computing power difference is breathtaking. And I very much resonated with your comments on the huge increase in dedicated computer power for deep learning and similar.
[1:38:53] That is likely to be transformative, given the recent. And so we have to think through that, because even with no new brilliant ideas on algorithms, there's going to be a huge leap forward. So thank you for that. That's a key observation here. In Congress, I'm the co-chair of the New Democrats Coalition on Future of Work Task Force, where we have been trying to think through what this means for the workplace of the future.
[1:39:21] And so I'd like to, Mr. Chairman, I'd like to submit for the record a white paper entitled Closing the Skills and Opportunity Gaps Without Objection. That objection is ordered. Thank you. And I will be asking for the record if you could have a look at this and see what sort of coverage you think this document has for the near-term policy responses. Because this is coming at us, I think, faster than a lot of people in politics really understand.
[1:39:51] And also, I will be asking for the record, I guess you may not have to respond right now, where are the best sources of information on how quickly this will be coming at us? There are conferences here and there, but you attend and your friends attend a lot of them. I'd be interested in where you think you really come together to get the techno experts, the economic experts, the labor economists, people like that, all in the same room. I think it's...
[1:40:19] It's something we should be putting more effort into.
[1:40:21] On another tech, I've been very involved in Congress in trying to resurrect something called the Office of Technology Assessment. What the GAO did here is very good, which is to bring, we had a conference of the experts, and you brought in a good set of experts, and a year later now we are getting a report on this, and you need more bandwidth in Congress than that, just on all technological issues, but this is a perfect example, a year old.
[1:40:49] A group of experts on AI, those are opinions that are sort of dated a little bit, even a year in the past. And so the Office of Technology Assessment for decades provided immediate
[1:41:04] High bandwidth advice to Congress on all sorts of technological issues. And so we are coming closer and closer every year in getting it refunded after it was defunded in the 1990s. And so I think, well, to ask you a question here, is there anyone on the panel who thinks that Congress has enough technological capacity as it currently stands to deal with issues like this?
[1:41:33] I can answer that. Yeah. No, it's a huge problem. And it's been aggravated by the fact that people have decided in their wisdom to cut back on the size and salaries available for congressional staff. One of my previous...
[1:41:52] Members who talked about here talked about the difficulty the federal government will have in getting real professionals, top-of-the-line professionals in here. And, you know, we're saying members of Congress are willing to do anything but give them the salaries that will be necessary to actually compete for those jobs.
[1:42:16] I would advocate everyone have a look at the Reference 5, which is your malicious use of AI, your Reference 5 in your testimony, which I stayed up way too late last night reading that, and it is real. Along the same lines, members of Congress have access to the classified version of a National Academies of Science study on the implication of autonomous drones. And this is something that I think...
[1:42:45] It has to be understood by the military. We're about to mark up a military authorization bill, an appropriations bill, that is spending way too much money fighting the last war and not enough fighting the wars of the future. And then finally, Dr. Lee.
[1:43:07] In the educational aspects of this, one thing I struggle for, I guess if you look through the bios of people who are the heroes of artificial intelligence, they tend to come from physics, math, places like that. And in theoretical physics or mathematics, a huge fraction of the progress comes from a tiny fraction of people. It's just a historical truth. And I was wondering, is AI like that? Are there a small number of heroes that really do most of the work and everyone else sort of fills in the thing?
[1:43:37] So like I said, Dr. Foster, AI is a very nascent field. So even though it is collecting a lot of enthusiasm worldwide societally, as a science, it's still very young. And as a young science, it starts from a few people. I was also trained as a physics major, and I think about early days of
[1:44:06] Newtonian physics and that was a smallish group of people as well. I mean it would be too much to compare directly but what I really do want to say is that we might be in the early even pre-Newtonian days of AI. We are still developing this.
[1:44:28] The number of people are still small. Having said that, there are many, many people who have contributed to AI. Their names might not have made it to the news, to the blogs, to the tweets, but these are the names that as students and experts of this field, we...
[1:44:45] We remember them. And I want to say many of them are members of the underrepresented minority group. There are many women in the first generation of AI experts. So we need to hear more from them. Two or three clicks down in the references cited by your testimony, and you look at the papers there and the author lists, it's pretty clear that our dominance in AI is due to immigrants.
[1:45:10] And Dr. Lee, I suspect you might not have come to this country under the conditions that are now being proposed by our president. And so I won't ask you to answer that. But it's important when we talk about what it is that makes this country dominant in things like AI. It is immigrants. And I'll just leave it at that. And I guess my time is up.
[1:45:32] Thank you, gentlemen. I thank the witnesses for their testimony and the members for their questions. The record will remain open for two weeks for additional written comments and written questions from members. Hearing is adjourned.