Meta’s powerful language models leak online, the U.S. Commerce Department opens up applications for CHIPS funds, and a PRC reorganization will impact R&D and data

Sign up
Worth Knowing

Meta’s Language Models — Released to Researchers — Leak Online: On February 24, Meta released a family of language models — dubbed LLaMA (Large Language Model Meta AI) — to the broader research community. The performance of the models, which range in size from 7 billion to 65 billion parameters, is not particularly exceptional — in an accompanying paper, Meta researchers say the 13B-parameter model outperforms OpenAI’s GPT-3 on “most benchmarks” while the 65B-parameter model is “competitive with the best models,” such as DeepMind’s Chinchilla and Google’s PaLM. But while those comparable models are either inaccessible to the public or only accessible through their developers’ APIs, Meta released both the source code and the model weights to researchers from academia and industry on a case-by-case basis. Much as with last year’s semi-public release of the less-sophisticated OPT-175B, Meta said the relatively unrestricted LLaMA release was part of its “commitment to open science” and expressed hope that it would help researchers identify and mitigate the problems of large language models. The restrictions that did exist were quickly rendered moot, however, as the model weights were (predictably) leaked on 4chan soon after the release — making it, in effect, the most powerful model available “in the wild.” Meta has defended its release strategy, but it is worth wondering whether the company will continue to be so open with new models in the future. CEO Mark Zuckerberg announced plans last week to create a “top-level product group” focused on generative AI. It remains to be seen whether Meta’s “commitment to open science” can coexist with a more AI-dependent business strategy.
Rotterdam’s Welfare Fraud Algorithm Had Major Bias and Accuracy Issues: This week, Wired and Lighthouse Reports published the results of a joint investigation into a machine learning algorithm used by the Dutch city of Rotterdam to identify potentially fraudulent welfare recipients. The city began using the risk-scoring algorithm — developed by the consulting firm Accenture and trained on nearly 13,000 previous welfare fraud investigations — in 2017, but paused its use in 2021 after auditors raised concerns about bias in the model. The Wired-Lighthouse investigation found that Rotterdam’s algorithm assigned higher risk scores to younger people, parents and residents who spoke a language other than Dutch. Those findings mirrored similar claims of algorithmic bias surfaced in 2016 by a ProPublica investigation of a common recidivism-prediction algorithm. The contrast between the two cases shows that some progress has been made: Rotterdam officials seem to have been relatively cooperative in working with Wired and Lighthouse Reports, in contrast to the company at the center of ProPublica’s investigation — though much of the analysis was only possible because the city inadvertently revealed most of the model’s training data to the researchers.

Expert Take: “The evaluation that the city of Rotterdam produced of its own model’s performance is one of the worst I’ve ever seen, and shows a model that should never have been trusted to make impactful decisions about individuals. What’s interesting about both this article and the 2016 ProPublica report is the decision to frame these models’ problems as problems of bias — which invites complicated statistical and philosophical arguments about fairness — instead of as an even worse failure to function with meaningful accuracy in the first place.” — Micah Musser, Research Analyst
A Planned PRC Reorganization Will Impact Tech and Data Ecosystems: Proposed reorganizations of the Chinese government could have a significant impact on the country’s technological development and data management ecosystems. A State Council plan submitted to the National People’s Congress (NPC) on Tuesday would, among other things, restructure the Ministry of Science and Technology in order to help “accelerate the realisation of high-level scientific and technological self-reliance.” The proposal would also establish a Central Commission on Science and Technology, which observers say would allow for greater party control over China’s technological development and funding decisions. Another plan under discussion at the NPC would concentrate China’s data management in a new national data bureau. Data regulation in China is currently handled by a patchwork of separate agencies; observers say the new bureau — which would be led by a vice-minister-level official — would streamline the country’s data oversight, though concerns remain that the new bureau could amount to an added layer of bureaucracy. Both plans are expected to be approved before the end of the NPC’s session, which runs through March 13.
Government Updates

The Commerce Department Opens Up Applications for CHIPS Funds: Last week, the Commerce Department announced the first funding opportunity for CHIPS for America semiconductor manufacturing incentives — a key step toward disbursing the $39 billion in incentives provided by the CHIPS and Science Act of 2022. The funding opportunity specifically seeks applications for projects related to the construction, expansion or modernization of U.S. facilities that produce “leading-edge, current-generation, and mature-node” chips (the full Notice of Funding Opportunity details the differences between each type of chip). Future funding opportunities — which the Commerce Department said would be announced later this year — will solicit applications for facilities related to R&D and semiconductor materials and equipment. In a speech by Commerce Secretary Gina Raimondo at Georgetown’s School of Foreign Service (with a subsequent question-and-answer session moderated by CSET’s Emily Weinstein) and in a “Vision for Success” document that accompanied last week’s announcement, the Commerce Department laid out the “cross-cutting themes” guiding its implementation plan and specific objectives the CHIPS Program Office has targeted. The former includes “catalyzing private investment,” “building a skilled and diverse workforce,” and “enforcing guardrails” (such as blocking stock buybacks), while the latter includes plans to have two or more “new large-scale clusters of leading-edge logic fabs,” two or more “high-volume advanced packaged facilities,” high volume DRAM production, and increased current-generation and mature-node chip production in the United States by the end of the decade. The Commerce Department will begin accepting full applications for leading-edge facilities on March 31 and for current-generation, mature-node, and back-end facilities on June 26.

The FTC Warns Companies About Overselling Their AI Products: In an item posted on its blog last week, the Federal Trade Commission issued new guidance warning businesses to “keep [their] AI claims in check.” AI hype cycles are nothing new, but the current boom — fueled in large part by generative AI systems like those powering OpenAI’s ChatGPT and the image-generating Stable Diffusion — has sparked what some observers are calling “a new gold rush.” Officials from the FTC’s Division of Advertising Practices seem concerned that marketers might oversell their products’ capabilities in their rush to cash in on the AI hype. The agency’s new guidance outlines four questions the FTC will look to answer while scrutinizing potentially misleading AI marketing claims: whether companies are exaggerating their AI products’ capabilities, whether they are promising (without sufficient evidence) that their products perform better than a non-AI competitor, whether they have made an effort to address the “reasonably foreseeable risks and impact” of putting their AI products on the market, and whether the product even uses AI. This isn’t the first time the FTC has weighed in on issues concerning AI — in 2021, the agency warned companies about the potential legal ramifications of deploying biased AI systems — and it is unlikely to be the last. As we covered last month, the FTC launched a technology office to support its tech-related enforcement efforts, including those related to AI.

The Air Force Wants to Ramp Up Its “Loyal Wingman” Program in 2024: President Biden’s FY2024 budget request — which he is set to unveil later today in Philadelphia — is expected to include plans for a “significant investment” in collaborative combat aircraft for the Air Force. In a speech earlier this week, Secretary Frank Kendall said the Air Force plans to eventually deploy approximately 1,000 CCAs — uncrewed “loyal wingman” aircraft that work together with planes piloted by humans to carry out elements of a mission — though he later said in comments to the press that the number was a “reasonable starting point” and could go even higher. In previous comments, Kendall has said that as many as five CCAs could ultimately be assigned to a manned aircraft, but the current plan calls for fewer: two each for 300 F-35s and 200 still-in-development “Next Generation Air Dominance” fighters. Kendall added that the CCA-related request will be related to the development of the program and “not the ultimate CCA.” Exact details will almost certainly be classified, but last year Kendall said that part of the budget request will go towards funding a 2024 competition to choose the best-suited CCA.

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

PRC ML Security Standard: Information Security Technology-Security Specification and Assessment Methods for Machine Learning Algorithms. This is a draft Chinese national standard for the security of machine learning (ML) algorithms. Drafted by Chinese government entities and tech companies, the standard provides security guidelines for developers of ML models and procedures for assessing ML algorithms’ susceptibility to cyberattacks.

South Korean Industrial Technology Protection Law: Act on Prevention of Divulgence and Protection of Industrial Technology. This document is South Korea’s industrial technology protection law, as amended in January 2023. The law aims to prevent technologies vital to South Korean national security or economic competitiveness from being divulged to or shared with foreign countries or corporations without the government’s knowledge.

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.

What’s New at CSET

REPORTS

PUBLICATIONS, PODCASTS AND MORE

TESTIMONY

  • U.S.-China Economic and Security Review Commission: On February 24, Director of Biotechnology Programs and Senior Fellow Anna Puglisi and Research Analyst Dahlia Peterson testified before the U.S.-China Economic and Security Review Commission at a hearing on “China’s Challenges and Capabilities in Educating and Training the Next Generation Workforce.” Read Puglisi’s testimony for the panel on “Advancing Growth, Knowledge, and Innovation through Higher Education,” read Peterson’s testimony for the panel on “The Role of Education in Promoting China’s Strategic and Emerging Industries” and watch the full hearing.  

EVENT RECAPS

  • On February 23, CSET and Georgetown’s Walsh School of Foreign Service hosted U.S. Secretary of Commerce Gina Raimondo, during which she delivered a speech on “The CHIPS Act and a Long-term Vision for America’s Technological Leadership,” and participated in a moderated conversation with CSET Research Fellow Emily Weinstein. Watch a full recording of the event.

IN THE NEWS

What We’re Reading

Paper: High-Resolution Image Reconstruction With Latent Diffusion Models From Human Brain Activity, Yu Takagi and Shinji Nishimoto (December 2022)

Article: Special Operations Forces Require Greater Proficiency in Artificial Intelligence, Kelley Jhong, War on the Rocks (February 2023)

Article: Algorithmic Black Swans, Noam Kolt, Washington University Law Review (February 2023)

Upcoming Events


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