Wednesday, May 27, 2020
|
by REBECCA KAGAN
|
|
Worth Knowing
TSMC Announces Plan to Build US Foundry: Taiwan Semiconductor Manufacturing Company, the world’s largest contract chipmaker, plans to open an advanced foundry in Arizona. TSMC will invest $12 billion in the foundry, which should begin producing 5 nanometer chips in 2024. However, experts have noted that TSMC’s leading-edge processes will remain in Taiwan: the company expects to produce a more advanced 3 nm chip in Taiwan in 2022, and the U.S. facility’s anticipated production capacity will be less than a fourth of some of TSMC’s largest foundries. The Trump administration had pushed for TSMC to develop a U.S. foundry; the federal government and State of Arizona have committed to supporting the facility.
China to Invest $1.4T in Domestic Technology: China plans to invest $1.4 trillion in its domestic tech sector over six years to reduce dependence on foreign technology companies, Bloomberg reports. The proposal, first outlined in a state think tank white paper in March, is part of a broader fiscal package China’s legislature is expected to approve this week. The plan includes investments in AI — including surveillance, automated factories and autonomous driving — as well as 5G, internet of things and high-speed rail. Chinese state-backed funds also recently announced a combined $2 billion investment into domestic chipmaker Semiconductor Manufacturing International Corp., following U.S. export controls announced on May 15 that jeopardize China’s imports of chips.
Microsoft and Nvidia Announce Advances in AI Supercomputers: Microsoft has partnered with OpenAI to create one of the world’s top five supercomputers, the company announced. Built exclusively for OpenAI, the supercomputer is designed to train OpenAI’s large AI models. Nvidia also debuted hardware products including a DGX A100 computer and the DGX Superpod, a cluster of DGX A100 computers that collectively function as a top AI supercomputer. The Department of Energy’s Argonne National Laboratory has already ordered DGX A100 systems and plans to use them to better understand and fight COVID-19.
Machine Learning Spotlight — Adversarial Attacks in NLP: Even the most advanced natural language processing models are vulnerable to adversarial attacks, according to recent research. Adversarial attacks are often inconspicuous tweaks in input data designed to trigger the wrong result from a machine learning model. Researchers developed TextFooler, a system that generates subtle alterations in wording — for example, changing “contrived” to “engineered” — that don’t alter a sentence’s meaning, but confuse machine learning models. TextFooler’s attacks reduced the top NLP models’ accuracy levels from above 90 percent to below 15. The researchers hope TextFooler will facilitate the development of more robust NLP models.
|
|
Government Updates
Commerce Department Limits Huawei’s Ability to Import Semiconductors: The Department of Commerce amended export controls to restrict Huawei’s ability to use U.S. technology to design and manufacture semiconductors abroad. While direct U.S. technology exports to Huawei were already limited, the company could still outsource manufacturing of its designs to foreign foundries that use U.S. semiconductor manufacturing equipment. The latest restrictions aim to close that loophole, though the exact implications remain unclear. Huawei said the decision threatens the company’s survival, and top chipmaker TSMC reportedly halted new orders from Huawei after the rule changed.
JAIC Awards $800M AI Contract to Booz Allen: The DOD’s Joint AI Center and the General Services Administration awarded Booz Allen Hamilton a five-year, $800 million contract for AI support to the DOD. The contract is the largest awarded by the JAIC to date. The task order requires Booz Allen to deliver “AI enabled products to support warfighting operations” and embed AI decision-making and analysis across all tiers of DOD operations. The contract was made possible through the JAIC’s partnership with the GSA, though outgoing JAIC Director Lt. Gen. Jack Shanahan has advocated for the JAIC to have its own acquisition authorities.
Bill Investing $100B in Emerging Tech Introduced: Last week, Senate Minority Leader Schumer, Sen. Young, Rep. Khanna and Rep. Gallagher introduced bipartisan legislation to increase U.S. investment in emerging technology research and development. The Endless Frontier Act would allocate $100 billion over five years for research grants, education fellowships and support for bringing new technologies to market, and establish a Directorate for Technology within the National Science Foundation. It emphasizes investment in key technologies including AI, computing hardware, cybersecurity and robotics. After Sen. Schumer announced the idea last fall, the four co-sponsors previewed the bill in a recent op-ed.
Deepfake Identification Bill Passes Committee: The Senate Committee on Commerce, Science, and Transportation approved legislation to support the development of technologies to improve deepfake detection. Sens. Cortez Masto and Moran introduced the bipartisan Identifying Outputs of Generative Adversarial Networks Act in November 2019. The bill requires the National Science Foundation to fund research into tools to identify manipulated media and directs the National Institute of Standards and Technology to develop standards to accelerate detection tool development. IOGAN awaits consideration by the full Senate; the House has approved a companion bill.
|
|
In Translation
CSET's translations of significant foreign language documents on AI
AI Security Standardization White Paper: Artificial Intelligence Security Standardization White Paper. This white paper by a PRC information security standardization group describes the current state of AI security and safety standards in China. Appendices list all of China’s AI security standards as of October 2019, and provide examples of AI security innovations pioneered by Chinese tech companies and IBM.
|
|
What’s New at CSET
REPORTS
PUBLICATIONS
IN THE NEWS
|
|
|
|