AI Policy Weekly #33
Llama 3.1, EU AI Act, and the VET AI Act
Welcome to AI Policy Weekly, a newsletter from the Center for AI Policy. Each issue explores three important developments in AI, curated specifically for US AI policy professionals.
Meta Releases Llama 3.1
In May 2020, OpenAI released GPT-3, a large language model that used approximately 10^23 operations during training.
About 23 months later, Meta released OPT-175B. The model was, roughly speaking, a replica of GPT-3.
Now, about 16 months after OpenAI’s March 2023 release of GPT-4, Meta has again succeeded in catching up to OpenAI with the release of Llama 3.1, a “herd” of three Llama models with 8 billion (8B), 70B, and 405B parameters.
Meta expended immense resources to build Llama 3.1-405B, which performs competitively with the best AI systems in the world.
16,384 NVIDIA H100 chips executed over 10^25 operations to train the model. 16 chips cost about $400,000, so this full arsenal could easily be worth $400 million.
Well over 200 employees worked on the project as “core contributors,” plus hundreds of additional employees in less involved roles. Top AI engineers can command multi-million dollar pay packages, so Meta may have paid over $100 million in labor costs.
Over 15 trillion tokens served as text data for training the model. Meta offers limited information about the “variety of sources” behind this dataset. Nonetheless, data access deals can cost over $100 million per year.
Additionally, varying hardware demands occasionally caused “instant fluctuations of power consumption across the data center on the order of tens of megawatts, stretching the limits of the power grid.” For reference, 10 megawatts would supply over 80,000 megawatt-hours in a year, enough to power thousands of US households.
In total, Meta may have spent half a billion dollars on this project’s resources. And they are poised to spend more; CEO Mark Zuckerberg told Bloomberg that “we’re basically already starting to work on Llama 4.”
The Llama 3.1 models—including versions without safety guardrails—are available for anyone on the internet to download. Facebook and Instagram users can access the 405B version for free through a chatbot interface at meta.ai.
Meta tested Llama 3.1 for risks related to cyber, chemical, and biological weapons. It appeared to be safe.
However, the testing failed to account for ways that malicious actors might modify, tune, and specialize the model to cause harm. This raises national security risks, as actors in China, Russia, and Iran are already using less-modifiable AI models to assist in influence operations and cyber attacks.
This serious oversight in Meta’s safety testing demonstrates the need for stronger safety practices at billion-dollar AI companies.
EU AI Act Enters Into Law
Earlier this month, the European Union’s AI Act was published in the Official Journal of the EU. This means the legislative text is essentially final.
AI companies are now on a timer to comply with the Act, which will enter into force on August 1st.
At that point, companies are encouraged to comply voluntarily, but they will still have time before various provisions become binding requirements. For example, rules on general-purpose AI will take effect in August 2025, and many other provisions will take effect in August 2026.
The Act includes requirements for general-purpose AI models with “capabilities that match or exceed the capabilities recorded in the most advanced general-purpose AI models.” In other words, “frontier” models that extend the frontier of AI capabilities.
Frontier AI providers must perform model evaluations, assess systemic risks, report serious incidents, and ensure cybersecurity protections.
The Center for AI Policy applauds the European Union for mandating these common-sense safety practices in frontier AI development.
Hickenlooper, Capito Introduce the VET AI Act
Senators John Hickenlooper (D-CO) and Shelley Moore Capito (R-WV) recently introduced the Validation and Evaluation for Trustworthy Artificial Intelligence (VET AI) Act.
The bill would direct the National Institute of Standards and Technology (NIST) to develop voluntary guidelines for conducting internal and third-party evaluations of AI models to mitigate harms, protect privacy, and more.
Additionally, the Commerce Department would study the current state of the AI assurance ecosystem, including topics like market demand, technical tools, and best practices for protecting confidential information.
Furthermore, the VET AI Act would establish an AI Assurance Qualifications Advisory Committee, which would recommend qualifications for parties seeking to conduct adequate evaluations.
The Center for AI Policy strongly endorses the VET AI Act. If enacted, the bill would develop clear standards for identifying qualified evaluators, which would help ensure effective AI safety assessments.
CAIP Event
On Monday, July 29th at 3 p.m. ET, the Center for AI Policy hosts a webinar titled “Autonomous Weapons and Human Control: Shaping AI Policy for a Secure Future.”
The event will feature a presentation and audience Q&A with Professor Stuart Russell, a leading researcher in artificial intelligence and a coauthor of the field’s standard textbook. We’re also privileged to begin the webinar with an introductory address from Mark Beall, co-founder of the American Security Fund and former Director of Strategy and Policy at the DoD’s Joint AI Center.
RSVP here.
News at CAIP
Jason Green-Lowe wrote a blog post titled “How to Advance ‘Human Flourishing’ in the GOP’s Approach to AI.”
Kate Forscey wrote a blog post about the Microsoft-CrowdStrike outage: “America Needs a Better Playbook for Emergent Technologies.”
Jason Green-Lowe wrote a blog post about five senators’ letter to OpenAI: “US Senators Demand AI Safety Disclosure From OpenAI.”
Brian Waldrip wrote a blog post introducing CAIP’s 2024 AI Action Plan, which hopes to build consensus on cybersecurity standards, emergency preparedness, and whistleblower protections.
Claudia Wilson wrote a blog post about covert malicious fine-tuning: “Researchers Find a New Covert Technique to ‘Jailbreak’ Language Models.”
Jason Green-Lowe wrote a LinkedIn article titled “5 Things I Learned in My First Year with CAIP.”
Jason Green-Lowe issued a press statement in response to OpenAI CEO Sam Altman’s AI governance op-ed in the The Washington Post.
We issued a statement of support for the VET AI Act from Senators Hickenlooper and Capito.
Quote of the Week
Does OpenAI plan to honor its previous public commitment to dedicate 20 percent of its computing resources to research on AI safety? [...]
Can you confirm that your company will not enforce permanent non-disparagement agreements for current and former employees? [...]
What security and cybersecurity protocols does OpenAI have in place, or plan to put in place, to prevent malicious actors or foreign adversaries from stealing an AI model, research, or intellectual property from OpenAI?
—US Senators Brian Schatz, Ben Ray Luján, Peter Welch, Mark Warner, and Angus King, in a letter to OpenAI CEO Sam Altman
This edition was authored by Jakub Kraus.
If you have feedback to share, a story to suggest, or wish to share music recommendations, please drop me a note at jakub@aipolicy.us.
—Jakub