A new EU law to rival the GDPR, the DOD names an AI chief, and the mystery drones in the most recent U.S. aid package to Ukraine

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

EU Leaders Agree on Another Big Tech Act: On April 23, European Union leaders reached agreement on a new law, the Digital Services Act, that would impose significant moderation and transparency requirements on many tech companies, including some of the world’s biggest tech giants. While the final text of the DSA has not yet been released, earlier drafts and public comments shed light on most of the law’s key characteristics. Of particular interest to AI policy observers are a requirement that so-called “very large online platforms” (platforms with more than 45 million monthly EU users) share internal data with EU authorities and vetted academic or non-profit researchers upon request, mandatory annual risk assessments and independent audits for companies, a provision that VLOPs must label deepfakes when they are discovered, and a requirement for “recommender system transparency.” That last mandate (Article 24a in an earlier draft), which requires online platforms to make information about their “recommender systems” available to users, could create major headaches for companies that use complex algorithms — or many complex algorithms acting in concert — to drive user engagement. Together with its companion bill — the Digital Markets Act, agreed to in March — the DSA’s impact has been compared to the GDPR, the massive privacy regulation that went into effect in 2018. Should the DSA pass, as is widely expected, it could go into effect as soon as next year.
Opening Twitter’s Algorithm: Last week, Elon Musk reached a deal to buy Twitter for approximately $44 billion. Of the many changes Musk has hinted at pursuing once the deal goes through (as it is expected to later this year), of particular relevance to an AI-focused audience is a potential plan to make Twitter’s algorithms open source. In a TED talk last month, Musk floated the idea of uploading the algorithm’s code to GitHub, where independent observers could suggest changes, critique design choices or point out mistakes. But a number of observers have noted that increasing Twitter’s algorithmic transparency isn’t as simple as copy-pasting to GitHub. As Will Knight wrote in Wired, Twitter has no “master algorithm” — instead, user-engagement decisions are determined by the complex interactions of many different algorithms, the (presumably exabytes of) data Twitter has access to, and the inputs of human users. What’s more, decisions made by machine learning-based algorithms, like those used by Twitter, are notoriously difficult to parse, even with complete access. In March, for instance, the Verge’s Alex Heath reported that it took Facebook engineers years to discover a bug that had caused the system powering Facebook’s News Feed to promote misinformation instead of suppressing it, and several months more to locate the bug. If Twitter’s algorithms are at all similar, it seems unlikely that access to its code would be particularly illuminating for outside observers without access to other crucial inputs.
Meta Releases 175B Parameter Language Model — Source Code and All: On Tuesday, Meta’s AI unit announced a new 175 billion parameter language model — dubbed OPT-175B — and released its source code and information about how it was trained to the broader research community. Trained using open-source datasets, the model’s performance is deliberately non-revolutionary compared to other models its size. According to Meta AI’s new research paper, the team aimed to match OPT-175B’s performance with that of OpenAI’s GPT-3 in order to make comparative study as accurate as possible. The team did, however, use recent advances in data parallelism to train the model with one-seventh the carbon footprint of GPT-3 — an important demonstration as the energy and compute demands of large language models (LLMs) become increasingly unsustainable. For now, OPT-175B’s code will be available only to vetted researchers, but Meta has made smaller OPT models ranging in size from 30 billion to 125 billion available to the public. The decision to release OPT-175B’s code earned praise, but also raised some eyebrows — not least because of Facebook and Meta’s less-than-transparent history. Meta AI’s Joelle Pineau told Will Douglas Heaven that the decision to release the code was part of a “commitment to open science” and a desire to establish a “state-of-the-art in terms of responsibility.” AI ethics expert Margaret Mitchell — whose 2021 article with Emily M. Bender, Timnit Gebru and Angelina McMillan-Major explored the dangers of LLMs — told Heaven that the release was a positive step, but raised concerns about how the model might be used in the wild and its potential to generate harmful content.
AI Interpretability Company Attracts Major Funding: Anthropic AI, a San Francisco-based company focused on AI safety, reliability and interpretability, has raised $580 million in Series B funding. Founded last May by a number of former high-ranking OpenAI staffers (including non-resident CSET Research Fellow Jack Clark), the company’s impressive fundraising — its Series A round netted $124 million — attests to the perceived importance, and potential profitability, of interpretable AI systems. As observers have frequently noted (and the EU and Twitter stories above illustrate), interpretable systems continue to be the white whale of the AI industry. Anthropic is registered as a “public benefit corporation” and says it plans to focus mostly on research at present, with plans to commercialize “down the road.” What those plans would be is unclear, but with possible fines for violating the EU’s Digital Services Act (including its provisions on recommender system explanation) reaching as high as 6 percent of companies’ annual revenue — Alphabet earned $257 billion last year, for reference — the value of AI interpretability to regulatory compliance alone could be staggering.       
Government Updates

DOD Names Chief Digital and Artificial Intelligence Officer: Last week, the Pentagon announced Dr. Craig Martell as the new Chief Digital and Artificial Intelligence Officer, putting him at the helm of the office meant to coordinate the Pentagon’s data, AI and digital efforts. The office of the CDAO, announced in December, was established to oversee the work of three of the Pentagon’s primary AI and data offices: the Joint Artificial Intelligence Center (though according to a recent GAO report, the JAIC will stand down by the end of fiscal year 2022), the Defense Digital Service and the office of the Chief Data Officer. Martell’s background spans industry and academia — he was most recently the head of machine learning for Lyft — and while his Pentagon experience may be lacking, the DOD has taken steps to ensure the office of the CDAO isn’t short on Pentagon expertise. Margaret Palmieri — the recently named deputy CDAO — founded the Navy Digital Warfare Office and has years of Pentagon experience under her belt. A DOD spokesperson touted the match of an “industry outsider” with “inside the department” expertise as a key consideration in the appointments.   

Project Maven Moves to NGA: The National Geospatial-Intelligence Agency will soon take over Project Maven, one of the DOD’s earliest and best-known AI initiatives. Established in 2017, the project — which is tasked with fielding AI-powered computer vision systems capable of analyzing drone footage — had been run as an “Algorithmic Warfare Cross-Functional Team” out of the office of the undersecretary of defense for intelligence and security. The move to the NGA, which is part of the DOD’s FY2023 budget request, shouldn’t come as a surprise. Discussions about giving Project Maven a more permanent operational home stretch back years — in 2020, for example, the Air Force was angling to take it over — and the NGA — the intelligence and combat support agency responsible for collecting and analyzing geospatial intelligence — has been involved with the project since its inception.  

Biden Sends More Aid to Ukraine — Including “Mystery Drones”: On April 21, the Biden administration announced an additional $800 million security assistance package for Ukraine, which includes 121 Phoenix Ghost drones — a new type of drone about which little is publicly known. The mystery surrounding the weapon and the Pentagon’s tight-lipped communication about its capabilities (plus, let’s face it, a cool name) set DOD watchers on the hunt for more details. Pentagon officials said the drone had been in development for the U.S. Air Force for some time and was tapped for the most recent assistance package based on conversations with Ukrainian officials about their needs. According to DOD officials, the new UAV is similar to the Switchblade drones — a line of loitering munitions that possess some autonomous capabilities but are not fully autonomous — of which the United States has sent more than 700. Last week, President Biden called on Congress to authorize $33 billion more in aid for Ukraine, including $20 billion in additional security assistance. While the White House has not yet specified how those funds would be spent, it seems likely, based on past packages, that future aid could include 1000 or more UAVs.

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

PRC Draft Dual-Use Export Control Regulation: Export Control Regulations for Dual-Use Items (Draft for Feedback). The following document is China’s draft regulation on export controls on dual-use items. When finalized, it will replace existing, separate regulations governing export controls on missile-related dual-use items and on nuclear, biological, and chemical weapons-related dual-use items. The PRC Ministry of Commerce is soliciting public feedback on this draft regulation through May 22, 2022.

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 May 6.
  • 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.
  • 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 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

REPORTSPUBLICATIONSTESTIMONY
  • Senate Armed Services Subcommittee on Cybersecurity: On May 3, CSET Senior Fellow Andrew Lohn testified on artificial intelligence applications in cyberspace operations at a Senate Armed Services cyber subcommittee hearing. He discussed AI’s capabilities and vulnerabilities in cyber defenses and operations. Read his testimony or watch it here.
IN THE NEWS
What We’re Reading

Paper: Planting Undetectable Backdoors in Machine Learning Models, Shafi Goldwasser, Michael P. Kim, Vinod Vaikuntanathan and Or Zamir, arXiv (April 2022) 

Report: Machine Learning Explainability & Fairness: Insights from Consumer Lending, FinRegLab (April 2022)  

Interactive: China Index: Measuring PRC Influence Around the Globe, Doublethink Lab (April 2022) 

Report: U.S.-China Technological “Decoupling”: A Strategy and Policy Framework, Jon Bateman, Carnegie Endowment for International Peace (April 2022) 

Upcoming Events


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