What Trump’s victory means for AI policy, chips, and AI development

Subscribe to policy.ai
Government Updates

What a Second Trump Term Means for AI Policy: Donald Trump’s victory in this month’s presidential election has many in the AI world wondering: what will the next four years look like for AI policy? Trump’s first term as president coincided with the start of the current boom in AI research and development. In 2017 — his first year in office — Google researchers published “Attention Is All You Need,” the hugely influential research paper that introduced the transformer architecture underlying today’s most powerful generative AI systems. In 2020, near the end of his first term, OpenAI released GPT-3, the company’s first model to make serious mainstream waves (a later version of GPT-3 powered ChatGPT when it launched in 2022). The Trump administration didn’t ignore the burgeoning technology: In 2019, President Trump signed the country’s first executive order on AI, which laid out a strategy to promote U.S. leadership in AI development. Then in December 2020, he issued the first executive order on federal government AI use. Just as the intervening four years have seen an unprecedented surge in AI development and investment, there has been a corresponding boom in AI policy efforts from the White House under Joe Biden. Highlights include:
On the campaign trail, Donald Trump said he would repeal Biden’s AI executive order on “day one” of his term, a goal that made its way into the Republican Party’s official 2024 platform. As Mozilla Fellow Deb Raji pointed out on social media, the Biden administration’s executive branch-heavy approach to AI policy — while more easily enacted than a traditional legislative strategy — is much more easily undone than policies that have passed through Congress. With Republicans’ victory in both the House and Senate, it seems unlikely that much of Biden’s AI agenda will be resuscitated on Capitol Hill. While both chambers of Congress have pursued some notable bipartisan AI initiatives in recent years — including Senate Majority Leader Chuck Schumer’s AI Insight Forums and House Speaker Mike Johnson and Democratic Leader Hakeem Jeffries’ bipartisan AI Task Force — neither chamber has shown much appetite for sweeping legislation.

What a Second Trump Term Means for Semiconductors: Since taking office in 2021, President Biden’s administration has pursued a two-pronged semiconductor strategy: incentivizing domestic manufacturing through the CHIPS and Science Act and limiting China’s access to high-end chips with strict export controls. On export controls, it seems unlikely that President Trump will roll back the Biden administration’s restrictions — announced in October 2022 — that aimed to significantly cut off China from the semiconductors needed for high-end AI applications. Those controls were part of a trend that began under President Trump’s first administration to restrict China’s access to critical technology, and Trump’s picks for key positions in his new administration — including noted “China hawks” Marco Rubio and Michael Waltz — don’t indicate a change in tack. Subsidizing domestic semiconductor manufacturing could be a different story. Though the 2022 CHIPS and Science Act passed through Congress with significant bipartisan support, Trump criticized the law during an October interview with the podcaster Joe Rogan. Soon after, House Speaker Mike Johnson indicated that repealing the law could be on his agenda, but quickly walked back those comments. While most observers don’t expect the Trump administration to try to repeal the law, the White House is working to allocate and finalize the remainder of the $39 billion in CHIPS Act incentives before President Biden leaves office. Last week, the Commerce Department finalized a $6.6 billion award for TSMC’s investment in new Arizona facilities, finalized $1.5 billion for GlobalFoundries on Wednesday, and is expected to finalize more awards soon.

What a Second Trump Term Means for AI Development: One area where questions remain most pronounced is what Donald Trump’s victory means for AI investment and development. Last year, the United States led the way in private investment — of the $95 billion invested worldwide, the United States accounted for $67 billion of it. From one angle, it looks like a second Trump term could be a boon for the technology companies financing the AI boom: President Trump earned the support of some of the tech world’s most outspoken voices, including Tesla, SpaceX, and xAI CEO Elon Musk and prominent venture capitalists like Marc Andreessen and David Sacks. After the election, Andreessen — who has championed a “Techno-Optimist” agenda and pointed to concerns about “hostile” overregulation as a key driver of his political activities — said Trump’s victory “felt like a boot off the throat.” But it’s far from clear that Trump’s second term will be a bacchanal for AI developers. During the first Trump administration, the Justice Department and Federal Trade Commission launched antitrust cases against Google and Facebook, respectively, and Amazon alleged it lost out on a lucrative DOD cloud computing contract due to political differences between the president and Amazon founder and then-CEO Jeff Bezos. Vice President-elect JD Vance has expressed support for current FTC Chair and Big Tech critic Lina Khan, and Attorney General nominee Matt Gaetz has previously called to break up Big Tech companies. Time will tell which voices — the tech accelerationists or the tech skeptics — will win out in the second Trump administration.

Worth Knowing

AI Wall? Hints of Diminishing Returns, But Developers Don’t Seem Worried: A report from The Information has sparked intense debate in AI circles about whether the field is approaching another “winter.” According to the outlet, OpenAI’s next flagship model, codenamed Orion, has shown more modest improvements over GPT-4 than the company saw between previous generations. And according to Bloomberg, OpenAI rivals Google and Anthropic have faced similar issues. The AI world is primed to worry about an AI winter — the history of AI research and development has been characterized by boom periods and fallow periods — and the year and a half since GPT-4’s release has seen OpenAI’s competitors, including Google, Anthropic, and Meta, release models that generally matched or only slightly exceeded GPT-4’s performance. But many in the world of AI development have dismissed the rumors of a scaling wall, including OpenAI CEO Sam Altman, who posted on social media that “there is no wall” and Anthropic CEO Dario Amodei, who seemed to dismiss the idea of a slowdown on a recent podcast. The truth could be that the scaling debate is growing less relevant — while companies are likely to hit a point at which ever-larger computing clusters become prohibitively expensive relatively soon, new approaches to AI development like OpenAI’s o1 reasoning model could open up new pathways for AI research that aren’t solely reliant on brute-force compute.
Big Tech’s LLMs for the Pentagon (and the People’s Liberation Army?): A handful of stories about military use of LLMs caught our attention:
  • Three of the biggest generative AI developers have begun working more closely with the U.S. military and intelligence agencies: Anthropic announced its Claude 3 and 3.5 family of language models will be available to the Pentagon and U.S. intelligence agencies through a partnership with Palantir and Amazon Web Services. Meta, meanwhile, gave the green light to U.S. government agencies and contractors — including those working on defense applications — to use its free, open-weight Llama models. And OpenAI, which had already worked with NASA, the IRS, and USAID, among other federal users, announced a limited ChatGPT Enterprise partnership with the Air Force Research Laboratory. The Pentagon and intelligence agencies have expressed interest in using generative AI tools since soon after OpenAI launched ChatGPT in 2022, and the DOD set up a task force last year to investigate the potential for generative AI integration across the military.
  • Chinese researchers linked to the People’s Liberation Army adapted Meta’s open-weight, 13 billion-parameter Llama model for military purposes, according to recent reports from Reuters and the Jamestown Foundation. The prospect that U.S. adversaries could repurpose open-weight or open-source models for nefarious purposes has been a significant concern for policymakers, and the latest reports seemed to confirm those fears — U.S. Rep. John Moolenaar issued a statement saying the news “[demanded] a critical reassessment of how we protect our technological edge from exploitation.” But, as CSET’s Kyle Miller argued on social media, the military usefulness of the Chinese-adapted Llama may be significantly overstated — “The Chinese researchers created a simple Q&A bot that could answer basic questions about the military,” Miller wrote. The model was fine-tuned on public information about the military and could answer relatively simple prompts about military issues, such as “describe the U.S. Army Research Laboratory,” but doesn’t appear radically different from other publicly available language models.

  • More: A Chinese lab has released a ‘reasoning’ AI model to rival OpenAI’s o1 | Open Foundation Models: Implications of Contemporary Artificial Intelligence
Anthropic Paces the AI Agent Race with Computer-Using Claude: Anthropic announced a significant upgrade to its Claude 3.5 Sonnet model that enables it to interact with computers like a human user — moving cursors, clicking buttons, and navigating between applications. So-called AI agents — sophisticated tools that can carry out actions autonomously in complex environments — have been a top goal of AI researchers and developers. OpenAI CEO Sam Altman called them “the next giant breakthrough” in a Reddit “Ask Me Anything” last month. While the updated Claude leads the field in computer-use benchmarks, scoring 14.9% on the OSWorld evaluation compared to the next-best competitors’ 7.7%, the technology remains far from human-level performance (typically 70-75%). Anthropic shared one charming example of the model’s shortcomings — at one point during a coding demo being recorded for the model’s launch, Claude stopped coding to scroll through pictures of Yellowstone National Park. Anthropic’s announcement has intensified competition in the AI agent space, with both OpenAI and Google reportedly preparing similar tools. According to Bloomberg, OpenAI plans to launch its “Operator” agent in January, while The Information reports that Google is developing “Project Jarvis,” a Chrome-focused tool for automating web-based tasks.
Job Openings

We’re hiring! Please apply or share the role below with candidates in your network:
  • Program Specialist: We are currently seeking a detail-oriented Program Specialist to join our growing operations team. This role will provide essential administrative, organizational, and project management support. The Program Specialist will coordinate various CSET programs, including internships, external fellowships, and student employment, while ensuring smooth logistical operations. Ideal candidates will bring strong organizational and project management skills, experience in program management, and the ability to work effectively in a collaborative, cross-matrix environment. If you are motivated, detail-oriented, and passionate about the intersection of technology and security, we encourage you to apply. Apply by Monday, November 25th
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 THE NEWS

What We’re Reading

Paper: Artificial Intelligence, Scientific Discovery, and Product Innovation, Aidan Toner-Rodgers (November 2024) 

Report: 2024 Annual Report to Congress, U.S.-China Economic and Security Review Commission (November 2024) 

Paper: Larger and more instructable language models become less reliable, Lexin Zhou, Wout Schellaert, Fernando Martínez-Plumed, Yael Moros-Daval, Cèsar Ferri, and José Hernández-Orallo, Nature (September 2024)