Deepfakes, facial recognition and UAVs — AI and autonomy during the invasion of Ukraine, chip news from Nvidia and Intel, and the DOD gets rolling on JADC2

Sign up
Work at CSET

We’re hiring! If you’re interested in joining our team, check out the positions in the “Job Openings” section below or consult our careers page.

Worth Knowing

Deepfake Showing Zelenskyy Surrendering Quickly Debunked: Unidentified hackers gained access to a Ukrainian news website and posted a deepfake video of Ukrainian President Volodymyr Zelenskyy appearing to surrender to Russian forces. While deepfakes have been used in a political context before, the Zelenskyy video is particularly notable because it is the first deepfake used to spread disinformation during an active conflict — something experts have long warned about. The video was not particularly well done and was quickly identified as a fake, but observers warned that future deepfakes are likely to be more sophisticated as the technology improves. Experts also noted that the Ukrainian government’s preemptive warnings about the likelihood of such deepfakes and its swift response to this particular video likely helped blunt its impact.
Facial Recognition Technology Plays A Role In Russia And Ukraine: Reports indicate that the Russian and Ukrainian governments, as well as private actors conducting open source intelligence research, have made use of facial recognition tools. According to the CEO of Clearview AI — a U.S. company with a vast database of more than 10 billion facial images — the company has offered its powerful facial recognition tool, a favorite of U.S. law enforcement, to the Ukrainian government, free of charge (though the Ukrainian government has not confirmed it is using Clearview’s tool). Wired’s Tom Simonite reports, meanwhile, that private individuals and organizations have used facial recognition tools to conduct open-source intelligence analysis and identify Russian combatants. In Russia, a number of anti-war protesters allege they were identified and arrested by police using facial recognition tools. Those allegations have not been confirmed, but they would not be surprising — Russia has built up a robust facial-recognition-powered surveillance apparatus over the last few years.
Thought Experiment Raises Red Flags About Dual-Use AI: Researchers repurposed a generative model used for drug discovery to quickly predict tens of thousands of potentially toxic molecules, raising concerns that simple, seemingly harmless AI tools could be deployed for nefarious purposes. In a paper published in the journal Nature Machine Intelligence, the researchers explained that by inverting the logic of their machine learning-based generative model — which is trained to penalize toxicity in predicted molecules — the model generated predictions for 40,000 potentially toxic molecules in under six hours. Several of the molecules were known chemical warfare agents — indicating the model’s accuracy — and others were predicted to be more toxic than the nerve agent VX, one of the world’s best known and most toxic chemical compounds. While the research was part of a thought experiment for a Swiss conference on developments relevant to the Biological and Chemical Weapons Conventions (and the process of actually synthesizing the model’s predicted molecules would be difficult, time consuming and expensive), its results have raised serious concerns about the future of seemingly benign AI systems as dual-use tools.          
Nvidia Announces New GPU Architecture And Datacenter Chip: At its AI-focused GTC event this week, Nvidia debuted its next-generation “Hopper” GPU architecture, its new H100 datacenter GPU, and gave an update on its Arm-based “Grace” CPU. Manufactured using a custom TSMC 4N process, the H100 delivers substantial performance increases in high-powered computing, AI inference and AI training tasks, compared to current state-of-the-art A100 GPUs, which power many of today’s fastest supercomputers. On a chip-to-chip basis, the H100’s Tensor Cores are six times faster than the A100’s, and the company says the H100’s new “Transformer Engine” leads to nine times faster AI training. Bandwidth and networking improvements, meanwhile, allow for significant scaling. Nvidia announced plans to build a supercomputer later this year dubbed “Eos,” comprising 576 DGX H100 systems and capable of 18.4 exaflops of AI computing performance, which would make it the fastest AI supercomputer in the world. Regarding its Arm-based “Grace” CPU, designed for advanced applications such as AI and natural language processing, and originally announced last year, CEO Jensen Huang said it is on track for 2023 and announced a “Grace CPU Superchip,” which combines two CPUs comprising 144 CPU cores total. The company says its H100 chips will be available beginning later this year.          
Stanford Institute Publishes Yearly AI Report: Stanford University’s Institute for Human-Centered Artificial Intelligence released the 2022 edition of its AI Index earlier this month. The yearly report (to which CSET contributed analysis) aims to provide a comprehensive account of the state of AI research and development. Among its findings:
  • Private investment in AI more than doubled between 2020 and 2021, from approximately $46 billion in 2020 to more than $93.5 billion in 2021. The number of newly funded AI companies continued its drop, however, from a 2018 peak of nearly 1,200 companies to 746 in 2021. The United States maintained its significant lead in both overall private investment and newly funded AI companies.
  • China retained its global lead in AI research publications, producing 63 percent more than the second-place United States.
  • The number of AI patent filings continued its explosive growth from fewer than 20,000 in 2018 to more than 140,000 in 2021. East Asia and the Pacific continued to lead in AI patent filings (though this lead diminished for the second year in a row), while North America continued to lead in granted AI patents (though its lead continued to fall from a 2015 peak).
  • As language models have grown, they have also become more toxic — a state-of-the-art large language model developed in 2021 had a 29 percent increase in “elicited toxicity” over a 2018 model.
  • More: Global AI Vibrancy Tool | HAI Seminar with Jack Clark on the 2022 AI Index Report 
Intel Makes Major Investment In European Chipmaking: Last week, Intel announced plans to invest more than 33 billion euros across a number of semiconductor manufacturing and R&D sites in the European Union. The announcement is a significant victory for the bloc, which has recently emphasized the importance of growing its domestic semiconductor industry and attempted to attract manufacturers with a 43 billion euro “Chips Act,” currently under consideration by the European Parliament and member states. Intel’s initial investment, which the company says could grow to 80 billion euros, includes plans for two new semiconductor fabrication plants in Germany, an R&D hub and European foundry design center in France, and other facilities in Ireland, Italy, Poland and Spain. The German fabs — set to cost 17 billion euros — will produce some of the company’s most advanced chips (equivalent to the 2 nm process node and smaller, though Intel moved away from the nanometer naming scheme last year) when production comes online in 2027.
Government Updates

DOD To Begin Implementing JADC2 Strategy: Last week, Deputy Secretary of Defense Kathleen Hicks signed the implementation plan for the military’s Joint All Domain Command and Control (JADC2) strategy. The strategy, which was approved in May 2021 by Defense Secretary Lloyd Austin, is meant to help the DOD coordinate and connect the sensors used by all of the services, allowing for rapid data processing — augmented by AI and machine learning — and enabling improved and accelerated decision making. Both the strategy and the implementation plan are classified, so exact details are unavailable. But according to a DOD press release, the implementation plan outlines specific actions, milestones for delivery and funding plans. A declassified summary of JADC2 strategy, also released last week, outlines the three “guiding C2 functions” of the strategy — “Sense,” “Make Sense,” and “Act” — and designates five “Lines of Effort” around which the strategy is organized, each guided by an Office of Primary Responsibility: Data Enterprise, Human Enterprise, Technology Enterprise, Integrating with Nuclear C2 and C3, and Modernizing Mission Partner Information Sharing.

A New NSF Directorate — But Competitiveness Bill Could Mean Changes: Earlier this month, the National Science Foundation announced a new directorate — the first new NSF directorate in more than 30 years — that will focus on advancing critical and emerging technologies and bringing them to market. The Directorate for Technology, Innovation and Partnerships will take over a number of existing NSF programs and serve as a “crosscutting platform” meant to “coordinate and enhance the impact of investments.” But it could undergo a swift transformation if Congress adopts the major tech competitiveness bill it’s currently negotiating. Both the House-passed America COMPETES Act and the Senate-passed United States Innovation and Competition Act contain plans for the directorate but with significant differences in funding levels and missions. This week, the Senate is formally considering the COMPETES Act, setting the stage for negotiations to reconcile the bills.

United States Sends Switchblade Drones To Ukraine: As part of its $800 million security assistance package, the United States will send 100 Switchblade drones to Ukraine. The Switchblade is a loitering munition that comes in two models — the 300, which is small enough to fit in a rucksack and is used for targeting personnel, and the larger 600, which carries an anti-armor warhead — though it is not clear which the United States is sending. With the Russian military unable to establish air superiority, the drones could prove an effective asset for the Ukrainian military. The decision comes as Russia has begun to use more of its UAVs, which had been conspicuously absent during the early days of the invasion. The systems currently in use in Ukraine possess some autonomous capabilities — the Switchblade 600, for example, can fly autonomously, while promotional materials for the Russian KUB-BLA claim AI-powered targeting capabilities that are unconfirmed — sparking concerns about the use of so-called “killer robots.” But as observers have pointed out, these systems likely do not meet the generally accepted threshold for Lethal Autonomous Weapons Systems because they cannot carry out a mission without significant human input.

Job Openings and New Funding

We’re excited to announce that CSET has received an additional $3 million in grant funding from the Hewlett Foundation to continue cyber and AI research, contributing to the CyberAI Project’s work at the intersection of artificial intelligence and cybersecurity.

And we have more exciting news — we’re hiring! Please apply or share the roles below with candidates in your network:
  • Data Scientist: We are currently seeking applications for a Data Scientist to explore research questions leveraging CSET’s unique data holdings. Apply by April 1.
  • UI/UX Designer: We are currently seeking applications for a UI/UX Designer to perform user interviews, write user stories, create user interface mockups, and conduct usability testing for public-facing Emerging Technology Observatory (ETO) products. Apply by April 1.
Please bookmark our careers page to stay up to date on all active job postings. You can also subscribe to receive job announcements by updating your subscription preferences in the footer of this email.

What’s New at CSET

REPORTSPUBLICATIONS, PODCASTS AND EXPERT TESTIMONYEVENT RECAPSIN THE NEWS
What We’re Reading

Report: Towards a Standard for Identifying and Managing Bias in Artificial Intelligence, Reva Schwartz, Apostol Vassilev, Kristen Greene, Lori Perine, Andrew Burt and Patrick Hall, NIST (March 2022)   

Report: The Tangled Web We Wove: Rebalancing America’s Supply Chains, Megan Lamberth, Martijn Rasser, Ryan Johnson and Henry Wu, CNAS (March 2022) 

Blog Post: Why DoD’s New Approach to Data and Artificial Intelligence Should Enhance National Defense, Michael C. Horowitz and Lauren Kahn, CFR (March 2022)  

Upcoming Events


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