Worth Knowing
DeepMind Develops A Matrix-Multiplying AI Agent: DeepMind developed a reinforcement learning-based system capable of identifying more efficient algorithms for performing matrix multiplication problems, an important type of calculation underlying many computational processes. Researchers from the London-based, Alphabet-owned lab outlined their findings in the journal Nature earlier this month. While matrix multiplication is a fairly simple task, finding the most efficient way to multiply matrices is extraordinarily difficult. Because the number of possible solutions makes brute-forcing solutions practically impossible, the DeepMind researchers instead created a single-player game where the goal was to generate a correct matrix multiplication algorithm in the shortest number of moves; they then had an AI agent based on the board game-playing AlphaZero — dubbed AlphaTensor — learn to play it using reinforcement learning. Once trained, AlphaTensor beat decades-old marks for matrix multiplication efficiency, ultimately setting new records for more than 70 matrices. When trained to optimize for hardware-specific efficiency, AlphaTensor generated algorithms that enabled runtime improvements. DeepMind hasn’t said exactly what it plans to do with its findings, but its work may provide clues for researchers to make further breakthroughs.
- More: DeepMind’s Game-Playing AI Has Beaten a 50-Year-Old Record in Computer Science | Measuring Perception in AI Models
NATO Moves Forward with Autonomy and AI Implementation: At NATO’s meeting of defense ministers earlier this month, the treaty organization approved a strategy for implementing autonomous systems and agreed to establish a Data and Artificial Intelligence Review Board. The new developments build on the results of last year’s meeting, at which NATO announced a new AI strategy, a “Data Exploitation Framework” policy and the creation of a €1 billion “NATO Innovation Fund” to support emerging tech innovation. The Autonomy Implementation Plan (of which only a summary is currently available) outlines the alliance’s desired outcomes vis-à-vis autonomous systems — which include the development of a “shared understanding and characterisation of autonomous systems” and the ability to deploy an interoperable “system of systems” — as well as guidelines for the responsible use of such systems, including Lethal Autonomous Weapons Systems. The new Data and Artificial Intelligence Review Board, meanwhile, will develop a “user-friendly Responsible AI certification standard” and serve as a medium for AI developers to collaborate and share best practices.
Government Updates
New Export Controls Take Aim at China’s AI and Chip Development: On October 7, the Commerce Department announced sweeping new export controls on key inputs needed by China for its development of AI and other emerging technologies. The new rules appear to represent a significant evolution in the U.S. government’s views of export controls, the role of semiconductors, and the country’s trade relationship with China. As National Security Advisor Jake Sullivan described last month, the importance of certain emerging technologies has motivated the Biden administration to abandon its strategy of maintaining “relative advantages” over strategic competitors in favor of building “as large of a lead as possible.” And as the new regulations explain, China’s military-civil fusion strategy means that earlier controls were not up to the task of keeping powerful chips out of the hands of the Chinese military. The regulations are too complex to run through in full here (for an excellent in-depth recap, see CSET Senior Fellow Kevin Wolf’s interview with Jordan Schneider) but at a high level, the new measures:
- Aim to limit China’s ability to develop and produce advanced node semiconductors, semiconductor production equipment, advanced computing capabilities and supercomputers. The regulations are not, however, a complete ban on semiconductor exports to China — they only affect chips that meet certain performance and networking standards.
- Attempt to restrict China’s access to the software and hardware needed to design and manufacture high-end semiconductors — defined as “logic chips with non-planar transistor architectures (I.e., FinFET or GAAFET) of 16nm or 14nm, or below,” DRAM chips of 18nm half-pitch or below, and NAND flash chips with more than 128 layers.
- Three new iterations of the foreign direct product rule — a powerful regulation that gives the U.S. government the ability to regulate products that use U.S.-origin hardware or technology, even if the final product was produced abroad — to cover a number of foreign-produced inputs, including those needed by China for advanced computing applications and building supercomputers.
- A licensing requirement on U.S. persons supporting the development or production of specified integrated circuits in China. Reports indicate that these restrictions have already had a significant impact, with work at a number of semiconductor companies grinding to a halt as their American employees were pulled from China-related projects.
- The addition of 31 Chinese firms to the Unverified List (published separately on October 7). While the new regulations do not place any additional Chinese organizations on the Entity List, firms listed on the UVL will be placed on the Entity List if they fail to complete end-use checks by the Bureau of Industry and Security within 60 days.
Biden Administration Releases Its First National Security Strategy: The Biden administration published its first National Security Strategy last week. The document (fact sheet available here) helps add context to the administration’s actions thus far and its plans going forward, including on issues related to emerging technologies. The strategy identifies China as the “only competitor with both the intent to reshape the international order and, increasingly, the economic, diplomatic, military, and technological power to do it” and outlines a three-part strategy for competing with the PRC:
- Investing in U.S. national power through industrial policy like the CHIPS and Science Act and by expanding the talent pool — especially in STEM fields — with workforce development programs and immigration reform.
- Working with allies and partners through organizations like the U.S.-EU Trade and Technology Council and the Quad to, among other things, assure the rules governing critical technologies are “based on shared democratic values.”
- Modernizing and strengthening the U.S. military for “the era of strategic competition with major powers” by investing in advanced technologies such as “trusted artificial intelligence.”
Biden Issues Executive Order on EU-U.S. Data Flows: On October 7, President Biden signed an executive order to implement the commitments the United States made on European data flows as part of the EU-U.S. Data Privacy Framework announced earlier this year. The order (fact sheet available here) establishes safeguards for handling Europeans’ data and creates a “multi-layer mechanism” for qualified individuals to seek redress through the Office of the Director of National Intelligence and an independent Data Protection Review Court housed in the Justice Department. Appointed judges on the court will be experts in data privacy and national security law from outside the U.S. government and will be protected against removal. The order does not mean that EU-U.S. data flows are in the clear yet, however. The European Commission must shepherd the framework through its own multi-step adoption process, after which the framework will likely be at the mercy of the European courts. The EU-U.S. Data Privacy Framework is the third pact governing data flows between the United States and the EU — the previous two having been struck down by the European Court of Justice after legal challenges brought by Austrian lawyer and privacy advocate Max Schrems. NYOB — the non-profit privacy organization headed by Schrems — criticized the order (as did the American Civil Liberties Union) and could bring a third suit.
In Translation
CSET’s translations of significant foreign language documents on AI
CSET’s translations of significant foreign language documents on AI
Chinese Academy of Sciences Five-Year Plan: Outline of the Chinese Academy of Sciences 13th Five-Year Development Plan. This document is the 13th Five-Year Plan — covering the years 2016–2020 — of the Chinese Academy of Sciences, a huge state-run complex of laboratories and research institutions. Notably, this plan names specific breakthroughs for CAS to make in 60 different emerging technologies. The plan also calls on CAS to improve its S&T-related open-source intelligence collection. CSET has not yet observed a comparable CAS plan covering the years from 2021 onward.
If you have a foreign-language document related to security and emerging technologies that you’d like translated into English, CSET may be able to help! Click here for details.
CSET Job Openings
We’re hiring! Please apply or share the roles below with candidates in your network:
- Fellow – Emerging Technology Supply Chains: We are currently seeking candidates to lead and coordinate our Emerging Technology Supply Chains Line of Research, either as a Research Fellow or Senior Fellow (depending on experience). This fellow will shape priorities, lay out an overall research strategy, oversee execution of the research and production of reports, and help hire and manage supporting researchers. Applications due by December 12.
What’s New at CSET
EMERGING TECH OBSERVATORY
- Yesterday, CSET launched the Emerging Technology Observatory, a public platform for data-driven tools, visualizations, and insights into the global emerging technology landscape. Drawing on CSET’s deep data infrastructure and led by a dedicated team of engineers and analysts, the ETO aims to inform critical decisions on emerging tech issues. Launching alongside the ETO are major updates to CSET’s Map of Science and Country Activity Tracker, as well as the Supply Chain Explorer, an interactive guide to the semiconductor supply chain. To learn more, visit eto.tech or read the introductory blog post.
- Foreign Affairs: Big Tech Goes to War: To Help Ukraine, Washington and Silicon Valley Must Work Together by Christine H. Fox and CSET’s Emelia Probasco
- CSET: Data Snapshot: Popularity != Public Implementation: How Organizations Leverage Open Source Code by Christian Schoeberl
- On October 12, Director of Data Science and Research Catherine Aiken joined the National Artificial Intelligence Advisory Committee Field Hearing to discuss Advancing U.S. Leadership in AI Research & Development.
- On October 14, CSET Director Dewey Murdick appeared on the AI Governance and Security Panel at the Summit on AI in Society. He discussed the drivers shaping the AI safety and policy discussions, highlighted the trends shaping the AI landscape, and surveyed a number of the policy goals and ways to achieve these goals.
- On October 19, the CSET webinar Decoupling in Strategic Tech Sectors featured a conversation between CNAS’s Martijn Rasser and CSET’s Tim Hwang and Emily Weinstein about the United States’ past efforts to decouple supply chains in satellite technology and what they can teach us about export control and trade policy for AI-relevant technologies today.
- USA Today: Kim Hjelmgaard cited the 2021 CSET brief China is Fast Outpacing U.S. STEM PhD Growth in a piece about Xi Jinping’s first two terms as general secretary of the Chinese Communist Party.
- The Wire China: Research Fellow Emily Weinstein spoke to Eliot Chen about brain drain from the United States to China.
- South China Morning Post: Weinstein also weighed in on U.S. and European concerns regarding Chinese state access to TikTok data for a recent Coco Feng piece.
- VentureBeat: A Taryn Plumb article about third-party cookies cited Husanjot Chahal, Helen Toner and Ilya Rahkovsky’s 2021 brief, Small Data’s Big AI Potential.
What We’re Reading
Report: State of AI Report 2022, Nathan Benaich and Ian Hogart (October 2022)
Article: Governing Artificial Intelligence in China and the European Union: Comparing Aims and Promoting Ethical Outcomes, Huw Roberts, Josh Cowls, Emmie Hine, Jessica Morley, Vincent Wang, Mariarosaria Taddeo and Luciano Floridi, The Information Society (September 2022)
Upcoming Events
- October 20: United Nations Institute for Disarmament Research, 2022 Innovations Dialogue: AI Disruption, Peace, And Security, featuring Margarita Konaev
- October 21: ScoopNews CyberWeek, Large Language Models and the Future of Disinformation, featuring Josh Goldstein
- November 10: CSET Tech and Security seminar, China’s AI Workforce: Demand for AI Talent in a Global Marketplace, featuring CSET’s Dahlia Peterson and Luke Kolosky, with Oliver Hayman of Amplyfi
What else is going on? Suggest stories, documents to translate & upcoming events here.