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.
OpenAI’s Updated GPT-3 Takes Human Evaluations Into Account: OpenAI announced an updated version of its GPT-3 natural language processing model that it says helps to mitigate some of the issues associated with the original. A persistent problem of large language models like GPT-3 has been their tendency to reproduce the toxic language and biases found in their training data, much of which is taken from the internet. OpenAI’s new model, dubbed InstructGPT, used reinforcement learning to incorporate and learn from the feedback of a group of 40 human reviewers. As the reviewers gave high scores to good responses and low scores to responses that they judged to be false or toxic in some way, the model learned to generate better responses. While this method of training — reinforcement learning from human feedback — is not new, observers say its application to a model the size of GPT-3 is significant. According to the results published by OpenAI, the effects of the training, while not massive, moved the model in a direction that is slightly less toxic, generates fewer false statements, and is deemed more helpful by the human evaluators. InstructGPT has now supplanted GPT-3 as the default model available through its API.
- More: Solving (Some) Formal Math Olympiad Problems | CSET: Truth, Lies, and Automation: How Language Models Could Change Disinformation
- More: Europe kept waiting as Intel commits to new US chip factories | ASML Warns Chinese Rival May Be Infringing its Trade Secrets
House Passes COMPETES Act — Conference With Senate Up Next: Last week, the House passed the America COMPETES Act, a major tech competitiveness bill that would direct billions of dollars to domestic semiconductor production and critical supply chains. The bill’s 222 to 210 passage was a near-party-line vote, with all but one Democrat voting for and all but one Republican voting against. The House and Senate — which passed its own tech competitiveness bill, the U.S. Innovation and Competition Act, last June — must now negotiate a compromise version to send to the President’s desk. While the two bills contain a number of similar provisions — including $52 billion in incentives for domestic semiconductor manufacturing and research and development — they differ in key ways, including their plans for the NSF and immigration policy changes, among other things. While the Senate’s bill earned Republican support, passing with 68 votes, observers note that the more divisive vote in the House could indicate an uphill battle in reaching a final deal. A timeline for negotiating a final agreement has not been laid out, but the Biden administration has urged Congress to act quickly.
The Pentagon’s CTO Lists Tech Priorities and Previews Tech Strategy: DOD Chief Technology Officer Heidi Shyu identified the Pentagon’s technological priorities and announced plans for an upcoming science and technology strategy in a memo issued last week. The memo listed 14 “critical technology areas vital to maintaining the United States’ national security,” including:
- Trusted AI and autonomy
- Advanced computing and software (including supercomputing, cloud computing, and data storage and processing)
- Human-machine interfaces
- Integrated sensing and cyber
- and Biotechnology
DOD’s New Lead AI Office Is Up and Running: Last week, the DOD announced that the office of the Chief Digital and Artificial Intelligence Officer had reached initial operating capacity and that it had named an acting CDAO — key milestones for the office meant to coordinate the Pentagon’s data, AI and digital efforts. The new office, which was announced in December, will oversee the work of three of the Pentagon’s primary AI and data offices: the Joint Artificial Intelligence Center, the Defense Digital Service and the office of the Chief Data Officer. The acting CDAO — the DOD’s Chief Information Officer, John Sherman, who will serve in both roles until a CDAO is appointed — told reporters that the new office will help create a “collective ecosystem” capable of shepherding the DOD’s AI and data projects from “end-to-end.” Two memos issued last week by Deputy Secretary of Defense Kathleen Hicks (available here and here) clarified the CDAO’s functions and reporting structures. The office — which will reportedly have a budget of approximately $500 million and 200 to 300 employees — is set to reach full operating capacity by June 1.
The IRS Ditches Its Facial Recognition Plan: The Internal Revenue Service dropped a plan that would have required taxpayers to submit to a facial recognition check to access their online records. As part of the proposed process, which was set to take effect this summer, IRS website users would have recorded a video selfie and uploaded it to ID.me — a private identity verification company — for a facial recognition check. The plan set off a wave of criticism from privacy advocates and members of Congress, who raised concerns both about how the company would handle taxpayers’ data and private information, and about the growing use of facial recognition by state and federal agencies. A GAO report released last year found widespread facial recognition use among surveyed federal agencies, and ID.me already handles identity verification for a number of state and federal agencies. In a press release, the IRS said it would “transition away from using a third-party service for facial recognition” and find another way of authenticating taxpayers’ identities.
CSET’s translations of significant foreign language documents on AI
CSET’s translations of significant foreign language documents on AI
PRC Big Data Five-Year Plan: “14th Five-Year” Plan for the Development of the Big Data Industry. This document, a spinoff of the March 2021 14th Five-Year Plan, describes China’s strategy for the growth of its big data industry through 2025. Although the big data plan includes very few quantitative targets for the industry, it does qualitatively describe the Communist Party’s near-term vision for China’s big data industry, which involves both making big data more broadly accessible and improving data security.
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.
Join us today at 2PM ET for a (virtual) Hiring Info Session!
We’re hiring! Please apply or share the roles below with candidates in your network:
- Research Analyst (multiple): CSET RAs are vital to our work across a range of lines of research. Research Analysts collaborate with Research and Senior Fellows to execute CSET’s research. Apply by February 25 and be sure to list your areas of research interest in your cover letter.
- Data Research Analyst (multiple): DRAs work alongside our analysis and data teams to produce data-driven research products and policy analysis. This role combines knowledge of research methods and data analysis skills. Those with experience in common data visualization, programming languages, and/or statistical analysis tools may find this position of particular interest. Apply by February 25.
- Business Operations and Management Specialist: Reporting to CSET’s Director of Operations, the management specialist will have responsibility for sub-grant processing, contracts management and grants management for the entirety of CSET. Excel/gsheets skills are a must. Apply by March 4.
What’s New at CSET
- Exploring Clusters of Research in Three Areas of AI Safety by Helen Toner and Ashwin Acharya
- Reshoring Chipmaking Capacity Requires High-Skilled Foreign Talent: Estimating the Labor Demand Generated by CHIPS Act Incentives by Will Hunt
- The Hill: The US can compete with China in AI education — here’s how by Kayla Goode and Dahlia Peterson
- All Sides with Ann Fisher: Department Of Justice’s China initiative And Consequences For Academic Freedom, featuring Emily Weinstein
- We are pleased to announce that Foretell — the crowd forecasting project started by CSET in 2020 — is now part of a larger forecasting program to support U.S. government policy decisions called INFER, which will be run by the Applied Research Laboratory for Intelligence and Security (ARLIS) at the University of Maryland and Cultivate Labs. Since its founding, Foretell has hosted over 27,000 forecasts on more than 100 questions, with contributions from nearly 2000 individual forecasters. INFER, short for INtelligent Forecasting of Events and Risks, is designed to generate valuable signals and early warning about the future of critical science and technology trends and events. To follow along with INFER’s forecasting work, visit the public portion of the program online.
- Axios: Research Analyst Will Hunt discussed the risks of an even bigger chip shortage with Margaret Harding McGill of Axios.
- FedScoop: For an article about the DOD’s list of critical technology areas, Jackson Barnett reached out to Associate Director of Analysis and Research Fellow Margarita Konaev for her thoughts.
- Vox: Konaev also discussed Moscow’s strategy of Ukrainian destabilization and humanitarian implications of urban warfare with Jen Kirby of Vox.
- Protocol: For a story on U.S.-China technological decoupling and how it might affect China in particular, Protocol’s Shen Lu spoke with Research Analyst Emily Weinstein.
What We’re Reading
Report: China’s Model of Science: Rationale, Players, Issues, China Aerospace Studies Institute (February 2022)
Article: Xi Jinping and the CCP’s expanding technology agenda, Kai von Carnap, The Strategist, Australian Strategic Policy Institute (February 2022)
Report: When the Chips Are Down: Gaming the Global Semiconductor Competition, Becca Wasser, Martijn Rasser and Hannah Kelley, Center for a New American Security (January 2022)
- February 10: CSET Virtual Hiring Info Session
- February 16: CSET Webinar, More than Deepfakes: AI and the Future of Disinformation Campaigns, featuring Katerina Sedova and John Bansemer
- February 17: U.S.-China Economic and Security Review Commission, Hearing on “China’s Cyber Capabilities: Warfare, Espionage, and Implications for the United States”, featuring testimony by Dakota Cary
What else is going on? Suggest stories, documents to translate & upcoming events here.