A proposal to make AI training more efficient, Biden’s EO on Chinese tech investment, and an influential new report on U.S.-China tech competition

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

Chip Giants Propose a Way to Speed Up AI Training: Last week, researchers from Nvidia, Arm and Intel published a whitepaper detailing a proposed type of computer number format they say will offer significant improvements in training efficiency without major loss of accuracy. Historically, most AI training required allocating 32 bits of memory to store each model parameter with a high degree of precision. Many AI developers have switched to using 16 bits of memory per parameter, which reduces memory requirements and makes model training less computationally costly. Some researchers have experimented with reducing precision even further, down to as few as 8 bits, though this risks degrading accuracy. The new whitepaper presents a 8-bit floating point (FP8) specification for deep learning that Nvidia, Arm and Intel claim can avoid significant loss in accuracy relative to FP16 approaches. This proposed format, which could allow training larger models on the same amount of compute, could help address a key issue facing AI developers: compute usage in AI is growing at a potentially unsustainable rate. Nvidia, Arm and Intel announced they had made the FP8 format open and license-free with the hopes that it would encourage broader industry adoption.    
Nvidia’s New Consumer Graphics Cards — What They Mean for AI: On Tuesday, Nvidia launched the first in its new line of 4000-series consumer graphics cards, which it says offer significant performance gains compared to its 3000-series line. Consumer GPUs helped launch the deep learning revolution — Alex Krizhevsky used two Nvidia GTX 580 GPUs to train the influential “AlexNet” convolutional neural network in 2012 — but their role has been largely usurped by data center-specific chips like Nvidia’s H100, which offer more memory bandwidth and greater networking capabilities (we covered the H100 announcement back in March). But while the biggest breakthroughs in AI will likely come from large models trained on massive clusters of H100s or equivalent data center chips, consumer cards (which are primarily marketed to video gamers) still play an important role in AI research. For researchers looking to experiment with smaller models and less complex applications, consumer GPUs can offer ample performance at a fraction of the price of data center cards (while Nvidia prohibited the use of consumer cards for “data center deployments” in 2017, it often makes exceptions for university users). Additionally, consumer GPUs will likely play a significant role in the way non-researchers engage with AI — the growing popularity of open-source AI models means that consumer GPUs can be used to fine-tune large language models, generate art locally, and even generate deepfakes. With Nvidia touting twice the performance of the previous generation’s chips, those workloads can be processed much more quickly.
A New Report on U.S.-China Tech Competition Gets D.C.’s Attention: A report published last week arguing that the United States is at risk of losing its technological competition with China has struck a chord with policymakers and other influential voices in Washington. The report was issued by the Special Competitive Studies Project, a bipartisan organization led by former Google head Eric Schmidt and Yll Bajraktari, who were both previously with the National Security Commission on Artificial Intelligence. While SCSP is an independent organization — unlike the congressionally established NSCAI — the report nevertheless seemed to earn the attention of many political leaders and policymakers in both parties. A day-long SCSP summit held in Washington last week included appearances by Deputy Defense Secretary Kathleen Hicks, Deputy Secretary of State Wendy Sherman, former Secretary of State Condoleezza Rice, National Security Advisor Jake Sullivan, former  National Security Advisor H. R. McMaster, Speaker of the House Nancy Pelosi, and Senators Schumer, Young and Rounds.
Government Updates

Chinese Tech Investments — Biden Provides Additional Guidance on CFIUS: Last week, President Biden signed an executive order meant to provide additional guidance on reviewing Chinese investment in critical U.S. tech companies. The order (fact sheet available here) directs the Committee on Foreign Investment in the United States — an interagency panel tasked with reviewing foreign investments in U.S. companies — to consider the following factors in its reviews:
  • A transaction’s effect on critical supply chains.
  • A transaction’s effect on U.S. leadership in critical technologies, including AI, microelectronics, and biotechnology.
  • A transaction’s effect in terms of broader investment trends.  
  • Investments that may create a cybersecurity risk.
  • Investments that pose a risk to “U.S. persons’ sensitive data.”
According to The New York Times’ David Sanger, that last factor is “potentially the most important element in coming months” and reflects growing U.S. government concerns about Chinese access to U.S. citizens’ data through popular platforms like TikTok. The move is the latest in a string of U.S. government efforts to limit Beijing’s access to U.S. technology. But more could be on the way soon — as Reuters reports, the Biden administration plans to codify restrictions on high-end chip and chipmaking equipment exports to China sometime next month.

NIST Partners with Google to Design Dozens of Research Chips: The National Institute of Standards and Technology announced a partnership with Google to develop and produce up to 40 new chip designs optimized for research applications. NIST will work with partners from a number of universities to develop and tailor the chip designs to suit their needs, and will make the resulting designs open-source so they can be further adapted and refined by other academic or industry researchers. The chips will be manufactured by Minnesota-based SkyWater Technology, with Google paying for the initial setup costs and subsidizing the first production run. While the chips will be manufactured using SkyWater’s relatively antiquated SKY130 node, as the Register’s Tobias Mann notes, these types of simple chips can be helpful for testing out experimental uses. Designing and manufacturing chips for such niche applications would normally be prohibitively expensive for most academic and small business researchers, but NIST hopes that the project will bring down costs significantly.      

NIH to Spend $130M to Advance AI Health Research: The National Institutes of Health announced plans to spend $130 million over four years to accelerate AI adoption in biomedical and behavioral research. The awards are the first issued by the NIH Common Fund for its “Bridge to Artificial Intelligence” (Bridge2AI) program, which aims to make the biomedical and behavioral research fields AI-ready. To date, AI has had a mixed record when it comes to healthcare. In some cases, it has proved genuinely useful. But in other instances, it has been ineffective or worse, with poorly trained AI systems producing biased results that exacerbate existing inequities. Encouragingly, Bridge2AI’s first grants reflect a considered approach to the potential pitfalls of AI adoption: four of the seven grants are for data generation projects meant to produce “high-quality ethically sourced data sets” on which AI systems can be trained, and the other three are to create a Bridge Center that will develop best practices, disseminate information on the use of data and AI systems, and evaluate their application.    

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

Demand for AI Talent in China: Artificial Intelligence Industry Talent Development Report (2019–2020 Edition). This document, published by China’s Ministry of Industry and Information Technology in 2020 on the basis of job posting data and surveys, describes the country’s strong demand for — and insufficient supply of — artificial intelligence talent. The authors make several recommendations to increase China’s supply of AI talent and to ensure that AI training in universities prepares graduates with the practical skills valued by employers.

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

We’re hiring! Please apply or share the roles below with candidates in your network:
  • Visual Communications Specialist: The Visual Communications Specialist will support the work of the External Affairs Team to raise the profile of CSET’s research through a variety of outreach activities such as graphics, multimedia, and publications. Apply by October 3.
  • Research Fellow — AI Applications: This Research Fellow will focus on helping decision makers evaluate and translate new and emerging technologies, particularly in the field of AI, into novel capabilities by separating real trends and strategic opportunities from technological hope and hype. Rolling application — Apply today.
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

PUBLICATIONSCONFERENCE PRESENTATIONS
  • American Political Science Association: Director of Data Science Catherine Aiken presented a CSET working paper entitled “Persuasive AI: An Experimental Test of GPT-3’s Targeted Political Statements”  with co-authors Data Scientist James Dunham and Senior Fellow Andrew Lohn at the American Political Science Association annual meeting last weekend in Montreal.
IN THE NEWS
What We’re Reading

Report: Rewire: Semiconductors and U.S. Industrial Policy, Chris Miller, Center for a New American Security (September 2022)

Report: Software-Defined Warfare: Architecting the DOD’s Transition to the Digital Age by Nand Mulchandani and Jack Shanahan, Center for Strategic & International Studies (September 2022) 

Paper: Robust Speech Recognition via Large-Scale Weak Supervision, Alex Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey and Ilya Sutskever, OpenAI (September 2022)

Upcoming Events


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