OpenAI and Google announce impressive new systems, AI in Biden’s budget request, a big DOD news roundup, and data scraping in the South Carolina courts

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

Google’s Joke-Explaining Language Model and the Future of LLMs: Earlier this week, Google debuted its new 540 billion parameter Pathways Language Model (PaLM), which appears to show significant improvements over state-of-the-art large language models across a number of benchmarks and tasks. In its announcement, Google included several impressive demonstrations of the model’s abilities, including joke explanation, chain-of-thought and cause-and-effect reasoning, and code generation. PaLM’s size and performance were made possible by Google’s Pathways system, which enabled training across thousands of accelerator chips at high levels of efficiency. GPT-3 and subsequent large autoregressive language models have shown that increasing the size of a model results in improved performance — which accounts for the last few years’ dramatic rise in model size. However, most of them were trained on a single system or used pipeline parallelism to train across multiple GPU clusters or TPU pods (with a maximum scale of 4096 TPU chips). Pathways allowed Google to conduct pipeline-free training at high levels of efficiency across two TPU pods containing 6144 chips. While these scaling advances make it feasible to use even larger compute infrastructures to build better models, given the high cost of compute, they may also centralize cutting-edge AI development even further among a small group of well-financed institutions. 
OpenAI Announces AI System That Turns Text Into Original Images: Yesterday, OpenAI announced DALL-E 2, an AI system that can convert natural language prompts into original images. The Microsoft-affiliated research lab, best known for GPT-3, introduced the original DALL-E last year, but that system was only capable of relatively simple, specific text-to-image tasks. According to OpenAI — and the examples it has shared appear to bear it out — the new version generates more realistic, caption-accurate, and higher-resolution images than the original system. DALL-E 2 can also edit existing images based on natural language prompts from a user. While the examples OpenAI has provided are undoubtedly impressive, a number of observers have raised concerns about the tool’s potential nefarious uses. The tool — or one like it — could be used to fuel disinformation campaigns, whether through entirely fabricated images or deceptively altered ones. OpenAI has acknowledged those concerns, and says it will be studying DALL-E 2’s limitations and capabilities and implementing “safety mitigations,” such as removing explicit content from its training data. For now, the system will not be available through OpenAI’s API, though it is intended to be released to the public eventually.    
The EU’s New Digital Act Could Mean Big Changes for Tech Giants: On March 24, European Union leaders reached agreement on a new law, the Digital Markets Act, that would dramatically reshape the business practices of many of the world’s biggest tech companies. The law would require designated “gatekeeper” companies providing “core platform services” — such as Alphabet, Meta, Microsoft and Apple — to make dramatic changes to the way they do business. In particular, its ban on aggregating user data across platforms without explicit consent could have a major impact on some of the world’s leading AI developers. Tech giants like Meta and Alphabet have leveraged their massive aggregated data resources to help train AI used for advertising, personalization, and other purposes. The threat of penalties of up to 10 percent of global revenue (20 percent for repeat offenses) may make that kind of data pooling too risky to be worth pursuing. Despite a fierce big tech lobbying campaign in Brussels and the Biden administration’s concerns that the law unfairly targets U.S. companies, the DMA is expected to be approved and take effect as soon as October.        
NAACP and ACLU File Suit Over South Carolina Data Scraping Ban: Last week, the South Carolina chapter of the NAACP filed a lawsuit against the state’s Court Administration over its ban on data scraping. The organization, represented by lawyers from the ACLU, argued in its lawsuit that the ban — which prohibits the automated data collection of publicly available information on the court’s Public Index — impairs its ability to provide resources and support to South Carolinians facing eviction, who are disproportionately Black. Although the Supreme Court has ruled that data scraping does not violate existing cybersecurity law, this case challenges the ability of public entities to impose anti-scraping technical measures — such as automation detectors that freeze browser access — and include scraping prohibitions in terms of service agreements for data they are required to make publicly accessible. In recent years, the data scraping practices of companies like Clearview AI — which has scraped together a vast database of more than 10 billion images to train its facial recognition tool — have come under fire from privacy advocates, politicians and private companies. But blanket bans affect more than Clearview and its ilk — as the South Carolina and other recent cases show, they can make some forms of academic research and non-profit advocacy work significantly more difficult.  
Government Updates

Biden Budget Includes AI Funding: President Biden announced his FY2023 budget request late last month, a $5.8 trillion proposal with substantial funding proposed for AI and emerging technology:
  • The budget includes a $773 billion request for the Pentagon, approximately $30 billion higher than Congress approved for the FY2022 budget. Under the proposed budget, the DOD’s total Research, Development, Test and Evaluation (RDT&E) spending would soar to its highest level ever — $130.1 billion, more than 9 percent higher than FY2022 levels. While the Pentagon’s Science and Technology budget would fall by 13 percent — to approximately $16.5 billion — DARPA would see a 7 percent boost to $4.1 billion. That builds on the 10 percent increase it already received for FY2022. The request also establishes and funds the office of the Chief Digital and Artificial Intelligence Officer — the new unit meant to oversee the work of three of the Pentagon’s primary AI and data offices and its key AI efforts.
  • It would direct $880 million to the NSF Directorate for Technology, Innovation, and Partnerships — a new directorate, announced last month, that will focus on advancing critical and emerging technologies and bringing them to market.  
  • It includes $187 million in new NIST research initiative spending, focused on speeding the adoption of AI and other emerging technologies by developing new standards.
For now, the request is just a proposal, and Congress will develop actual spending levels in the coming months. For more information on science funding in the President’s budget request, check out the American Institute of Physics’ excellent FYI Federal Science Budget Tracker.

Third Time the Charm for U.S.-EU Data Deal?: On March 25, the United States and the European Union announced they had reached agreement on a framework for transatlantic data sharing. While exact details about the agreement still have to be hammered out, a White House fact sheet says the United States has made “unprecedented commitments” to bolster privacy and civil liberty protections and provide redress to EU individuals through a multi-layer redress mechanism, which will include an “independent Data Protection Review Court.” The announcement comes almost two years after Europe’s top court struck down the previous agreement that governed data transfers, the EU-U.S. Privacy Shield, on the grounds that it did not adequately protect EU citizens’ privacy, especially from U.S. intelligence authorities’ activities. Since that ruling, the transatlantic data environment has stood on ambiguous legal ground, and a number of U.S. companies, including Mailchimp and Google, have had their services banned by regulators in EU member states over privacy concerns. Both the Privacy Shield agreement and its predecessor, the Safe Harbor accord, were struck down by the European Court of Justice because of legal challenges brought by Austrian lawyer and privacy advocate Max Schrems. Schrems said he will wait for more details to be released but is ready to challenge the new deal “if it is not in line with EU law.”   

A DOD News Roundup: It’s been a busy two weeks for DOD news, with several recent stories that could impact the future of AI and emerging technology in the military:
Senate Vote Moves Tech Competitiveness Bill Forward: Last week, the Senate voted 68-28 to pass a revised version of the House of Representatives’ America COMPETES Act, substituting the original House text with the Senate’s previously passed U.S. Innovation and Competition Act. The vote clears the way for the House and Senate to establish a conference committee that will aim to negotiate a compromise bill both chambers can pass. Over the last year, both bodies have approved versions of a major tech competitiveness bill that would authorize $52 billion in funding to boost domestic semiconductor production and billions more to bolster critical supply chains. Despite support from President Biden and generally bipartisan backing in both chambers, differences in other aspects of the legislation — such as plans for the new NSF Directorate for Technology, Innovation, and Partnerships — have produced delays. While the Senate vote is an important step toward the bill’s final passage, observers don’t expect a compromise before the end of May.

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

PRC State Laboratory Recap: 2016 Annual Report on State Key Laboratories. This document lists all of China’s state key laboratories and summarizes their key accomplishments in 2016. An appendix includes contact information for all 254 state key labs and all seven pilot national laboratories. The document also briefly notes the results of a 2016 evaluation of the 75 biology- and medicine-related state key labs, which gave one lab a failing grade and identified another eight as in need of improvement.  

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:
  • Data Scientist: We are currently seeking applications for a Data Scientist to explore research questions leveraging CSET’s unique data holdings. Apply by April 18.
  • 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 products. Apply by April 18.
  • Software Engineer: Software Engineer to assist with data cleaning and normalization, data pipeline development and/or web development, and creating automated test suites for CSET’s new Emerging Technology Observatory. Apply by May 6.
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 AND PODCASTSEVENT RECAPS
  • Director of Biotechnology Programs and Senior Fellow Anna Puglisi joined Craig Singleton, Bonnie Glick, Phelim Kine and RADM (ret.) Mark Montgomery for a Foundation for Defense of Democracies event, “Disrupting China’s Military-Academic Complex.” Senator Marco Rubio and Jonathan Schanzer provided opening remarks. Watch a recording of the event here.
IN THE NEWS
What We’re Reading

Report: Artificial Intelligence: DOD Should Improve Strategies, Inventory Process, and Collaboration Guidance, United States Government Accountability Office (March 2022)

Commentary: The New Arms Race: Sanctions, Export Control Policy, and China, Jeannette Chu, CSIS (March 2022) 

Upcoming Events


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