OpenAI and Google show off new AI features, Meta’s Llama 3 is the best in open-source, DARPA’s AI dogfight, and U.S.-China AI risk talks

Sign up
Worth Knowing

OpenAI and Google Make A Big Push for Broader Consumer AI Adoption: This week, OpenAI and Google debuted a number of new products and features:
  • On Monday, OpenAI announced a new flagship model — dubbed “GPT-4o” — that, while only slightly more sophisticated than the previous flagship, GPT-4 Turbo, is significantly faster and cheaper to use than its predecessor. GPT-4o’s speed and native multimodality — it can take text, audio, photos, and videos as input with fewer steps than earlier models — allow for more seamless “conversations” with the model. Much of the announcement event focused on demonstrations of this “Voice Mode” (which drew inevitable comparisons to the 2013 movie “Her”). While, to date, access to OpenAI’s most advanced model has been paywalled (for the most part), the company said that GPT-4o would be rolled out to free users in the coming weeks. It is already available to paying users through ChatGPT Plus and the OpenAI API.
  • On Tuesday, Google debuted a host of new AI-powered products and features at its yearly Google I/O conference. Many of these new tools seem positioned to compete with OpenAI’s products: its Gemini 1.5 Flash language model is a low-latency model similar (at least in terms of speed) to GPT-4o; its Imagen 3 text-to-image generator is in the same ballpark as OpenAI’s DALL-E 3; and its new text-to-video generator — “Veo” — comes a few months after OpenAI announced its own video generator, Sora (though neither company has said when it will roll the tools out to the general public). Notably, Google extended its advantage in massive LLM context windows (the amount of input data a model can process at once): while Gemini could already handle an industry-leading 1 million tokens, Gemini 1.5 Pro will have a 2 million token context window.
That neither event debuted a generational performance leap — such as a GPT-5 or a Gemini 2 — may have come as a bit of a letdown to some, but the improvements shouldn’t be ignored — they are more likely to drive widespread AI adoption than even a next-generation model. Price, availability, and speed have been some of the clear impediments to widespread AI use, and reducing or removing those obstacles will be essential to breaking more completely into the mainstream. Additional announcements could be on the horizon that would help flesh out both companies’ strategies — such as GPT-5 or a deal between Apple and OpenAI or Google to put a state-of-the-art LLM on an iPhone — but, for now, it seems that both Google and OpenAI are focused on expanding their user bases rather than pushing LLM benchmarks to their absolute limits.
A New Top Open-Source Model — Meta Releases Llama 3: Last month, Meta announced “Llama 3,” the next generation of its family of open-source language models. At launch, the Llama 3 family has two models: a larger 70B-parameter model that nears (or even exceeds) the performance of some leading proprietary models like Google’s Gemini and Anthropic’s Claude on some benchmarks, and a smaller 8B-parameter model that can be run locally (not connected to the internet) on relatively inexpensive hardware. While OpenAI’s GPT-4o and Google’s Gemini 1.5 are undoubtedly more sophisticated models, Llama 3 is probably more important from a security-policy perspective. Its open weights are great for law-abiding businesses and individuals who can now run near-state-of-the-art models at a fraction of the cost, as well as for researchers who can learn much more about how such models work, but they are also a boon to propagandists, scammers, and other nefarious actors who can spin up campaigns at low cost and with very few guardrails. And while cloud computing providers can exert some control over the use of the larger 70B Llama 3, the ability to run the 8B model locally means the only ongoing practical constraint to running an influence campaign will be the cost of electricity. More releases in the Llama 3 family could be coming soon. Meta says it has 400B-parameter models still in training and teased more releases soon, though it’s not clear whether those releases will also include open weights.
Stanford AI Index Recaps a Big Year for Generative AI & Foundation Models: In April, Stanford University’s Institute for Human-Centered Artificial Intelligence released the 2024 edition of its AI Index. Now in its seventh year, the highly anticipated report (to which CSET contributes data), provides a comprehensive overview of the state of AI’s impact on research, investment, the workforce, and more. Key findings from this year’s report include:
  • 2023 was big for foundation models, both in terms of the number released (149 — more than double the previous year) and the cost to train them; the authors estimate that the compute used to train GPT-4 and Gemini Ultra cost $78 million and $191 million, respectively.
  • Open-source AI research boomed in 2023, with AI-related GitHub projects increasing by nearly 60% compared to 2022.
  • China maintained its significant lead in AI patents, accounting for more than 60% of the world’s AI patents in 2022 (the most recent year with data). The United States was a distant second, with 21% of the total.
  • Despite a massive increase in generative AI investment — nearly nine times higher in 2023 than in 2022 — overall private investment in AI continued to fall from its 2021 peak, dipping under $100 billion for the first time since 2020.
  • Academic brain drain continues, with the rate of newly minted AI PhDs turning to industry over academia increasing by 5.3% last year to more than 70%.
  • Gender diversity in AI postsecondary education remains an issue. While there was a slight uptick in the percentage of female bachelor’s graduates in 2022 (the most recent year with data), graduate and PhD levels saw a minor downturn. In both cases, the rates have remained largely steady since 2010.

  • More: ETO: Red-Hot Topics in AI Research: Insights from the Map of Science
Government Updates

Schumer-Led AI Working Group Releases Roadmap: Yesterday, the bipartisan Senate AI Work Group, led by Majority Leader Chuck Schumer, released its long-awaited roadmap for U.S. AI regulation. The final product attempts to translate lessons learned from last fall’s nine “AI Insight Forums” and other related efforts into a clear agenda for legislative action. It recommends that Congress move to, among other things:
Senator Schumer said the working group does not intend to put all of its recommendations into one sweeping package. Instead, he and the other members of the AI working group hope that relevant committees will move ahead with smaller bills based on these recommendations in line with their respective areas of jurisdiction.

USAF and DARPA Held Real-World AI Dogfighting Test: The Air Force and the Defense Advanced Research Projects Agency pitted an autonomously piloted plane against a human pilot in live dogfight exercises last year, military officials confirmed. While several countries — including the United States and China — have conducted virtual AI-vs-human testing over the last decade, the September 2023 test at Edwards Air Force Base is the first confirmed real-world exercise. The test used an X-62A VISTA — an F-16 modified to fly autonomously as part of DARPA’s Air Combat Evolution (ACE) program — flying against a human-piloted F-16. Maneuvers eventually ramped up to nose-to-nose engagements, during which the planes flew as close as 2,000 feet at 1,200 miles per hour. Officials said the test was about demonstrating “that we can safely test these AI agents in a safety critical air combat environment,” rather than seeing which side would come out on top, and the “winner” of the tests wasn’t disclosed. But if previous tests are any indication, the AI-powered planes are more than capable of holding their own; back in 2020, an autonomous fighter beat a human pilot 5-0 in a virtual DARPA-run test. The tests are an important step toward the Air Force’s goal of fielding 1,000 collaborative combat aircraft — autonomous aircraft flown side-by-side with human-piloted aircraft — by the end of the decade.  

U.S. and PRC Representatives Hold AI Risk Talks in Geneva: On Tuesday, representatives from the United States and China met in Geneva for high-level talks on AI risk and safety. The talks were arranged after last November’s meeting between Presidents Biden and Xi in San Francisco. The U.S. delegation for the Geneva talks was led by White House official (and former CSETer) Tarun Chhabra and the State Department’s Seth Center. Their counterparts were officials from China’s Foreign Ministry and the National Development and Reform Commission. According to a PRC Foreign Ministry readout, the talks retread some old ground: both sides agreed that AI presents opportunities and risks, China reiterated that it would like the United Nations to take the lead in setting up a global framework for AI governance, and the PRC representatives “expressed a stern stance on the U.S. restrictions and pressure” in the AI space. But observers said that breaking new ground wasn’t the aim of the initial “get-to-know-you” meeting; as CSET’s Sam Bresnick told the Associated Press, the meetings are an important opportunity for the United States to “get a better sense of China’s approach to defining and mitigating AI risks” and could help to build trust between the two superpowers in the long run.

Commerce Is Developing Export Controls on AI Models: The Commerce Department is in the process of developing new rules that would control the export of advanced AI models, according to a Reuters report. In recent years, the Biden administration has imposed controls on the powerful chips used to train AI systems, but exports of the models themselves are still entirely legal. It’s unlikely the discussed restrictions would control any current models — according to the Reuters report, the Commerce Department is considering using the computing power threshold established by President Biden’s October 2023 executive order: 10^26 floating-point operations. As of today, no AI systems are thought to have hit that mark, but it’s likely that the next generation of models will surpass it. But a source at the Commerce Department said that the threshold could ultimately be lower, or coupled with other factors such as a model’s use cases or underlying training data. While Reuters says the Commerce Department is not yet close to finalizing a rule, some supporters in Congress are laying the necessary groundwork for such a rule when it does come. On May 8, a bipartisan group of House representatives introduced the Enhancing National Frameworks for Overseas Critical Exports Act, which would give the Commerce Department clear authority to control the export of AI systems or national security-related technologies. According to Reuters, the newly proposed bill was developed with input from the Biden administration.

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

PRC Draft AI Law: Artificial Intelligence Law of the People’s Republic of China (Draft for Suggestions from Scholars). This document is a preliminary draft of China’s proposed AI Law that has circulated among legal scholars. The draft law specifies various scenarios in which AI developers, providers, or users are liable for misuse of AI tools. It also allows for the use of copyrighted material for model training in most cases and provides intellectual property protections for content created with the assistance of AI technology.

Computing Power White Paper: White Paper on China’s Computing Power Development Index (2022). This annual white paper by a Chinese state-run think tank analyzes the computing power landscape, in China and globally, as of late 2022. The report begins by summarizing global developments in compute over the past year and then provides statistics on China’s compute industry. The authors find that, despite disruptions caused by the COVID-19 pandemic, China’s computing power growth outpaced the world average in 2022.

PRC Industrial Policy Paper: Implementation Opinions on the Innovative Development of Pilot Testing in the Manufacturing Industry. This Chinese industrial policy directive aims to speed up the process of getting manufacturing prototypes into commercial production. One of the goals of this directive is to establish five homegrown, world-class pilot testing platforms in China by 2025. It prioritizes the development of more precise metrology technology as one means to this end.

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 role below with candidates in your network:
  • Strategic Engagement Specialist: The Center for Security and Emerging Technology at Georgetown University (CSET) is seeking applications for a Strategic Engagement Specialist to advance CSET’s development goals through cultivating relationships with strategic stakeholders, preparing funding requests, and providing critical administrative support. Apply by Monday, June 10.
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

REPORTS

PUBLICATIONS

EMERGING TECHNOLOGY OBSERVATORY

EVENT RECAPS

  • In April, CSET’s Director of Strategy and Foundational Research Grants Helen Toner delivered a talk at TED2024 on the importance of developing smart AI policy, even in the face of uncertainty. Watch her full talk.
  • On April 29, CSET hosted DHS Assistant Secretary Iranga Kahangama and AI Policy Specialist Noah Ringler for a discussion, moderated by CSET’s Jessica Ji, on the department’s AI efforts. Watch a full recording of the event.

IN THE NEWS

What We’re Reading And Listening To

Report: Supercharging Research: Harnessing Artificial Intelligence to Meet Global Challenges, President’s Council of Advisors on Science and Technology (April 2024)

Podcast: The Road to Accountable AI, Kevin Werbach 

Article: In Arizona, election workers trained with deepfakes to prepare for 2024, Sarah Ellison and Yvonne Wingett Sanchez, the Washington Post (May 2024)