Microsoft and Nvidia’s massive new language model, TSMC plans to build fab in Japan, and the CIA launches tech mission center

Sign up
CSET Survey

As CSET continues to grow, we want to learn more about our readers and how we can better serve their interests. If you’re interested in helping, you can do so by filling out a short, anonymous survey. Thank you for your support of CSET and policy.ai!

Worth Knowing

Microsoft and Nvidia Debut Massive Language Model: Last week, Microsoft and Nvidia announced a new natural language generation model, the latest to claim the title of “largest natural language generation model,” following in the footsteps of OpenAI’s GPT-3 and Microsoft’s Turing-NLG. The monolithic transformer language model, dubbed the Megatron-Turing Natural Language Generation (MT-NLG) model, contains 530 billion parameters, more than three times as many as GPT-3. According to Microsoft, the model’s size helps it achieve “unmatched accuracy” in tasks such as reading comprehension, completion prediction and commonsense reasoning. The computing power required to train the model was similarly massive — training took place on Nvidia’s Selene supercomputer, powered by 4480 A100 GPUs across 560 DGX A100 servers. The training set used to train the model consisted of 15 different datasets, with 11 of those coming from the massive open source dataset The Pile. While the dataset gave MT-NLG plenty of material on which to train, Microsoft and Nvidia pointed out that the model likely picked up on biases present in the training data — a persistent problem for natural language processing systems.
TSMC to Build Chip Foundry in Japan: Taiwanese semiconductor manufacturer TSMC announced it would build its first chip plant in Japan, the latest in the chip giant’s expansionary efforts. While Japan is still a key supplier of semiconductor materials and manufacturing equipment, its status as the world’s leading chip producer is a distant memory — since its peak in the 1980s, Japan has fallen behind Taiwan and South Korea in total chip production and produces none of today’s high-end, sub-10 nm chips. While the new TSMC plant won’t fix the latter problem — it will reportedly produce chips using the less-advanced 22 and 28 nm process nodes — observers say that increased domestic production is likely to help Japanese companies that have been hit hard by the ongoing chip shortage, such as Toyota, Honda and Sony (which may invest in and co-run the plant, though those reports have not been confirmed). The plant, which is expected to begin production in 2024, is another sign of TSMC’s shift to building manufacturing capacity outside of Taiwan. While most of its fabs remain there, over the past year the company has moved some of its production off the island and closer to its key customers — it recently broke ground on a $12 billion Arizona plant, expanded production at its Nanjing fab, and has reportedly considered building its first factory in Europe.
Former U.S.A.F. Software Head Says China Has Major AI Lead, Sparking Debate: In an interview with The Financial Times, the former chief software officer of the U.S. Air Force, Nicolas Chaillan, criticized the U.S. military’s R&D efforts and said the United States would “have no competing fighting chance in China in 15 to 20 years,” sparking a heated discussion about the U.S. military’s priorities and competitiveness. Chaillan — who resigned last month after nearly three years of government service, posting a resignation letter on LinkedIn that hit many of the same notes — told the FT that Pentagon bureaucracy and overregulation stifled innovation and kept costs too high. He also criticized companies like Google for their resistance to working with DOD and said concerns about ethics had delayed U.S. AI development. While Chaillan’s critiques earned a good deal of attention, they also generated significant pushback: the Army’s chief information officer, Raj Iyer, said Chaillan’s argument was “absolutely not true”; Tim Culpan wrote in Bloomberg that Beijing’s AI lead is overstated, with most of its advantages concentrated in the narrow (if concerning) field of population surveillance; and JAIC’s inaugural director, Lt. Gen. Jack Shanahan, offered a one-word response.  
Government Updates

White House Officials Call for “Bill of Rights for an AI-Powered World”: Earlier this month, Biden administration officials from the Office of Science and Technology Policy announced they would be developing a “bill of rights” for AI. In an opinion piece published in Wired, two senior OSTP officials, Director Eric Lander and Deputy Director for Science and Society Alondra Nelson, argued that in the face of powerful technologies that can lead to discrimination, invasions of privacy and manipulation, the United States should “clarify the rights and freedoms we expect data-driven technologies to respect.” As a first step in the process, OSTP issued a request for information about biometric technologies, including facial recognition, used to identify people or assess their emotional or mental states. While the Wired piece was relatively light on details about the final form and implementation of a “bill of rights,” Lander and Nelson wrote that possible enforcement mechanisms could include restrictions on federal and contractor use of non-compliant technologies and “laws and regulations to fill gaps.” Comments on the RFI will close on January 15, 2022.  

Army Exercise Uses AI to Select Bombing Targets: As part of its “Scarlet Dragon” exercise, the Army’s XVIII Airborne Corps used AI to select bombing targets from more than 7,000 square kilometers of satellite imagery. The exercise used software from the Air Force’s Project Maven, but unlike Maven — which uses AI to scan real-time drone footage — Scarlet Dragon used AI to process satellite images from the National Geospatial Intelligence Agency. While the exercise originally included plans to drop two GBU-32 bombs on the selected targets, those were scuttled due to weather and safety concerns. But the exercise served as an important step toward demonstrating the military’s integration of AI as it attempts to narrow the space between sensor and shooter. Next month, the military is scheduled to begin the second edition of its Project Convergence exercises. Unlike that exercise — which aims to test and implement future uses of technologies like AI under battlefield conditions as part of the military’s long-term Joint All-Domain Command and Control (JADC2) strategy — Scarlet Dragon was reportedly meant to test how current AI capabilities could be used in a present-day conflict.

CIA Creates New Tech Mission Center and CTO Role: Earlier this month, CIA Director William J. Burns announced the creation of both a chief technology officer position and a new technology-focused mission center. A press release from the agency, which also announced the creation of a China-focused mission center, said the Transnational and Technology Mission Center will “address global issues critical to US competitiveness — including new and emerging technologies, economic security, climate change, and global health.” While the announcement did not mention AI explicitly, a source told The New York Times that the new center will help track and counter the technologies — among them facial recognition and other forms of biometric surveillance — that have made traditional espionage much more difficult. As part of the agency’s effort to attract and retain talent, Burns also announced the creation of a “CIA Technology Fellows” program meant to bring outside experts into the agency for 1-2 year postings.

In Translation
CSET’s translations of significant foreign language documents on AI

PRC AI Ethical Guidelines: Ethical Norms for New Generation Artificial Intelligence Released. This document, issued by a state AI governance committee, lays out ethical norms for the use of AI in China. The norms cover areas such as the use and protection of personal information, human control over and responsibility for AI, and the avoidance of AI-related monopolies. The document does not specify how these norms are to be enforced, nor does it mention any punishments for those who violate the norms.

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.

Job Openings

Please apply or share the roles below with candidates in your network:
  • Survey Research Analyst: CSET’s Survey Research Analyst will lead survey design and execution in collaboration with CSET analysts. They will work closely with data scientists, software engineers, a translation lead, and other data research analysts to produce, collect, and analyze original survey data. Strong applicants will have experience with survey design, online survey distribution platforms (e.g., Qualtrics), project management, and quantitative data analysis. Submit your application by December 1.
  • AI Research Subgrant (AIRS) Program Director: CSET’s AIRS program will promote the exploration of foundational technical topics that relate to the potential national security implications of AI over the long term via research subgrants. The Director of AIRS will manage all technical, programmatic, and financial aspects of the new AIRS program.
  • Research Fellow – Cyber/AI: CSET’s CyberAI project is currently seeking Research Fellow candidates to focus on machine learning (ML) applications for cybersecurity to assess their potential and identify recommendations for policymakers (background in ML programming or cybersecurity highly desired). 
  • Senior Fellow: CSET’s Senior Fellows provide mentorship and intellectual leadership; shape and lead lines of inquiry and research projects aligned to our research priorities; and facilitate engagements with government, military, academic, and industry leaders.
  • Georgetown’s Walsh School of Foreign Service, the home institution for CSET, is hiring for a Professor of the Practice in Security Studies and Director of External Education and Outreach: This three-year, non-tenure-line faculty position in the Security Studies Program will have teaching administrative responsibilities. The candidate would teach four courses a year and oversee SSP’s external education and outreach activities. A Ph.D. with a specialization in a security-related area is preferred. The start date for this position is January 1, 2022, though flexibility on start date is possible.
  • Gracias Family Chair in Security and Emerging Technology: This non-tenure-track, rank-open position will have teaching and administrative responsibilities in the Security Studies Program, with potential for affiliations with CSET and the Science, Technology and International Affairs Program. The successful candidate will have a record of professional or teaching experience focused on security and emerging technology, with a particular focus on AI and its implications for national and international security.
Please visit CSET’s careers page to stay up to date on all active job postings.

What’s New at CSET

REPORTSPUBLICATIONS, VISUALIZATIONS, AND PODCASTSFORECASTS
Foretell has launched a new project that combines expert and crowd judgment. You can read more about the experts’ views, including how they think trends like China’s military aggression, political polarization, and the strength of the tech sector affect the DOD-Silicon Valley relationship. See all 20 forecast questions associated with this project here.

EVENTSIN THE NEWS
What We’re Reading

Article: How China Is Planning For a Tech Decoupling, Alex Stone and Peter W. Singer, Defense One (October 2021) 

Article: Virtual idols are a hot commodity in China, Shen Lu, Protocol | China (October 2021)  

Article: Multitask Prompted Training Enables Zero-Shot Task Generalization, Victor Sanh et al., arXiv (October 2021) 

Upcoming Events


What else is going on? Suggest stories, documents to translate & upcoming events here.