OpenAI’s new chatbot takes the internet by storm, China looks to bolster its chip industry, and a NeurIPS 2022 roundup

Sign up
Worth Knowing

OpenAI’s New Chatbot Wows (and Worries): Late last month, OpenAI debuted a prototype AI-powered chatbot — dubbed ChatGPT — capable of performing a number of natural language tasks with an impressive level of sophistication. As in the aftermath of OpenAI’s DALL-E 2 release earlier this year, the internet was soon awash in screenshots of ChatGPT conversations. In addition to being a seemingly capable code debugger, essay writer, and conversationalist, users have prompted ChatGPT to design a weight loss plan, write fake Bible verses, and take the SATs. And while ChatGPT’s moment in the spotlight has spawned a robust conversation about what such innovations mean for human creativity, the system — as OpenAI CEO Sam Altman cautioned — is still far from perfect. Its guardrails seem to have kept it from going the way of Microsoft’s Tay — the 2016 chatbot that was shut down after it started praising Hitler — but some users have found clever ways of circumventing ChatGPT’s various protections. And, predictably, it appears to suffer from many of the same bias issues that are practically universal among large language models. That said, more capable bots with fewer (or just different) weaknesses could be on the way soon. ChatGPT is based on a version of OpenAI’s GPT-3.5 (see our coverage from earlier this year for more on how OpenAI used reinforcement learning from human feedback to update GPT-3), but industry insiders expect that the company will debut GPT-4 as soon as 2023.
NeurIPS 2022 Showcases the Latest in AI and ML Research: NeurIPS 2022 — the 36th edition of the prestigious annual AI and ML research conference — wrapped up last week in New Orleans. After a year in which AI art and language tools like DALL-E and ChatGPT took the internet by storm, it’s perhaps no surprise that generative AI research played a starring role at NeurIPS. Of the 15 “Outstanding” papers recognized by the conference (out of 2,905 accepted), a significant portion were related to image or text generation. These included Google AI’s paper about its Imagen model, DeepMind’s research on “compute-optimal” large language models, and the paper introducing LAION-5B — the open dataset used to train Stable Diffusion. Other award-winning submissions included Nvidia’s paper about its Minecraft-playing agent, Microsoft’s work on a “Neural Corpus Indexer” and a paper from researchers at Princeton about how AI agents can learn human inductive biases (for a helpful overview of the award-winning papers, see Jim Fan’s excellent Twitter thread). The winner of this year’s “Test of Time” award — which recognizes the best-aged paper from the NeurIPS conference 10 years ago — went to one of the most important papers in machine learning history: ImageNet Classification with Deep Convolutional Neural Networks by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, which kicked off the so-called deep learning revolution, won the award by a unanimous vote.      
Cut Off From Chipmaking Tools, China Looks to Build Up Its Own Industry: China is working on a $143 billion support package for its domestic semiconductor industry, according to a Reuters report. The move appears to be part of Beijing’s response to sweeping U.S. export controls, announced in October, that aimed to hamstring China’s domestic semiconductor industry and limit its access to high-end chips. According to the Reuters report, the package could be rolled out as soon as the first quarter of 2023. Initial indications suggest it would primarily consist of subsidies and tax credits to subsidize purchases of Chinese-made semiconductor manufacturing equipment (SME) and other Chinese semiconductor production and research activities. The U.S. export controls put significant limits on China’s access to the tools needed to make high-end semiconductors. Now the Netherlands and Japan, both key SME exporters, appear ready to join the United States in cutting off China’s access to its most advanced tools. Without access to foreign inputs and expertise, Chinese-made SME cannot yet come close to replacing the best Dutch, Japanese, or U.S. equipment, and the reported package is unlikely to change that in the short term. Time will tell if China’s investment will be enough to help bridge the gap.
Government Updates

AI and Emerging Tech in the FY2023 NDAA: On December 8, the House of Representatives passed a compromise version of the Fiscal Year 2023 National Defense Authorization Act by a vote of 350-80. The bill (full text available here) authorizes roughly $858 billion for defense and defense-related activities and includes a number of notable provisions relevant to AI:
  • The Advancing American AI Act, which directs federal agencies to regularly identify and report on AI use cases within the government, requires the Office of Management and Budget to identify four new use cases for AI to support interagency or intra-agency modernization initiatives, and mandates that the Department of Homeland Security develop policies and procedures for the acquisition and use of AI in a manner that supports privacy and safe use (Title LXXII, Subtitle B).
  • A directive to the Deputy Secretary of Defense to identify priority enterprise projects for using AI and other digital solutions to improve business efficiency or warfighting capabilities, and assigns various responsibilities for implementing the projects (Sec. 1513).
  • A requirement for the Commander of U.S. Cyber Command and the DOD’s Chief Information Officer, in coordination with the Chief Digital and AI Officer, DARPA, and the Under Secretary of Defense for Research and Engineering, to develop a five-year plan for adopting AI to support DOD cyber missions (Sec. 1554).
  • A pilot program to evaluate the intelligence community’s ability to provide intelligence support for the export control and investment screening activities of the Department of Commerce and DHS using open-source, publicly and commercially available information (Sec. 6311).
  • An authorization of efforts to expand the participation of minority-serving institutions in DOD innovation programs and increase DOD STEM talent diversity (Sec. 222 and Sec. 1083, among others).
The Senate is expected to pass the bill this week, after which it will be sent to the president to be signed into law.

DOD Awards Cloud Contract to Amazon, Google, Microsoft and Oracle: Last week, the Pentagon announced it had awarded a multi-vendor cloud-computing contract to Amazon, Google, Microsoft and Oracle in what amounts to one of the biggest federal cloud computing contracts to date — as much as $9 billion over five years. The new Joint Warfighter Cloud Capability is expected to act as the backbone for the military’s cloud computing needs, serving “across all security domains and classification levels.” When the DOD announced plans for the JWCC last year, officials said that going with a multicloud approach — multiple clouds from different providers — would let the military remain flexible by taking advantage of vendors’ “varying capabilities.” The Pentagon had originally awarded a 10-year, $10 billion cloud computing contract to Microsoft in 2019, but it canceled that deal over concerns about its growing computational needs and, observers speculated, because of lawsuits from Amazon and Oracle alleging an unfair process. While the contract could be a boon to the four companies, only a small portion of the $9 billion is guaranteed — the rest will be doled out over the next five years as military users bid on the provider best-suited to the task at hand. DOD officials say they hope this competition will help keep overall prices down.  

The U.S.-EU Trade and Technology Council Holds Its Third Meeting: On December 5, representatives from the United States and the European Union — including Secretary of State Antony Blinken, Commerce Secretary Gina Raimondo and European Commission Executive Vice President Margrethe Vestager — convened in Maryland for the third meeting of the U.S.-EU Trade and Technology Council (TTC). A joint statement by the participants outlined key takeaways from the meeting and offered updates from the ten working groups previously established by the council. Developments of note include:
  • The United States and EU issued a first “Joint Roadmap on Evaluation and Measurement Tools for Trustworthy AI and Risk Management.” Meant to inform and align both parties’ approaches to AI risk management and trustworthy AI, the roadmap lays out key short- and long-term objectives, including the development of shared terminologies and developing tools to “map, measure, manage, and govern AI risks.”
  • The TTC released a joint study on AI’s impact on the U.S. and EU workforces. The study charts the current state of AI adoption and discusses potential future impacts.
  • The U.S. Commerce Department and the European Commission announced plans to implement an early warning system for semiconductor supply chain disruptions.
  • The United States and EU announced a pilot project to evaluate the use of synthetic data and “privacy enhancing technologies” for health and medicine research.
According to the joint statement, the TTC’s next meeting will be held in Europe in the middle of next year.

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

Shenzhen AI Regulations: Regulations for the Promotion of the Artificial Intelligence Industry in Shenzhen Special Economic Zone. This document is a set of regulations for the artificial intelligence industry in Shenzhen City in China’s Guangdong Province. The regulations aim to promote AI in Shenzhen, one of China’s tech hubs, but also stress the importance of ethical guidelines for the use of AI technology, so as to prevent AI systems from aiding discriminatory practices or infringing on individuals’ privacy.

Shanghai AI Regulations: Regulations for the Promotion of the Development of the Artificial Intelligence Industry in Shanghai Municipality. This document is Shanghai’s set of regulations for the artificial intelligence industry. The regulations, which took effect in October 2022, aim to marshal supercomputing infrastructure, data centers, financing, and talent for the benefit of the city’s AI industry, and to apply AI technology in municipal governance.

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 and New Funding

The Patrick J. McGovern Foundation has increased its investment in CSET’s success with a $500,000 grant to support diverse talent development and the protection of democratic ideals in the cybersecurity and information sphere. In addition to its initial grant in 2021, the Foundation’s new gift will  help CSET continue its work on education and workforce policy at the confluence of AI and cybersecurity and expand into new research areas. These will include cutting-edge AI research for cybersecurity and the study of mitigation strategies to counter disinformation.

And we have more exciting news — we’re hiring! Please apply or share the role below with candidates in your network:
  • People Operations Specialist: We are currently seeking a People Operations Specialist to play a key role in helping to build and develop the CSET team, with a particular focus on furthering our diversity, equity and inclusion initiatives. This Specialist will provide administrative, organizational and project management support to ensure that our people-focused operations run smoothly. Applications due by January 30.
  • Fellow – Emerging Technology Supply Chains: We are currently seeking candidates to lead and coordinate our Emerging Technology Supply Chains Line of Research, either as a Research Fellow or Senior Fellow (depending on experience). This fellow will shape priorities, lay out an overall research strategy, oversee execution of the research and production of reports, and help hire and manage supporting researchers. Applications due by January 2.
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

PUBLICATIONSEMERGING TECHNOLOGY OBSERVATORYEVENT RECAPS
  • On December 5, Zachary Arnold, analytic lead for CSET’s Emerging Technology Observatory, demonstrated the newly launched ETO’s capabilities and led a discussion with Melissa Flagg and John VerWey on the key policy takeaways that can be derived from its tools.
IN THE NEWS
What We’re Reading

Report: Human-Machine Teaming in Intelligence Analysis, Anna Knack, Richard J. Carter and Alexander Babuta, Centre for Emerging Technology and Security, The Alan Turing Institute (December 2022) 

Commentary: What China’s Algorithm Registry Reveals about AI Governance, Matt Sheehan and Sharon Du, Carnegie Endowment for International Peace (December 2022) 

Article: Abandoned: The Human Cost of Neurotechnology Failure, Liam Drew, Nature (December 2022)