AI Policy Weekly #20
AI Policy Weekly #20
Meta shares Llama 3, AI CEOs outline what’s next, and Senators release the Future of AI Innovation 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 AI policy professionals.
Meta Releases Llama 3
Among AI models that are widely available for downloading and modifying—often called “open models” or “open source AI”—there is a major new frontrunner: Meta’s Llama 3.
The model is currently available in small and medium sizes, with 8 billion and 70 billion parameters, respectively. Both are among the best in their size class.
To request access to the 70-billion (70B) version, all you need to do is provide your name, date of birth, email, country, and organization. With this information alone, Meta will likely struggle to effectively identify and block access for malicious actors aiming to repurpose the model for large-scale phishing attacks or disinformation campaigns.
Alternatively, if you don’t feel like waiting to download 100+ gigabytes of weights, you can speak with the model using Meta’s new chatbot interface website, meta.ai. However, to generate images, you’ll need to sign in with your Facebook account.
How did Meta build these models?
First, using an enormous amount of data. The Llama 3 training set is seven times larger than the Llama 2 training set. In fact, Meta went far beyond what is typical for the 8B model, using ~75x more training data than the “Chinchilla scaling” practices that were in fashion just a year ago.
Second, as always for AI leaders, using an enormous number of mathematical operations, which requires an enormous stockpile of high-performance hardware. The 8B model appears to have trained with just over 10^24 operations, whereas the 70B model used almost 10^25 operations. Thus, the 70B model is among the ten most computationally intensive AI systems ever trained.
Meta certainly has the chips to fuel that training, since it recently announced two computing clusters containing 24,576 H100 GPUs each, plus plans to scale to hundreds of thousands of GPUs this year.
Meta is not stopping here: the company is still finishing training an enormous Llama 3 variant that has 405 billion parameters. That version is already scoring on par with the most generally capable AI models available today, such as OpenAI’s GPT-4 and Anthropic’s Claude 3 Opus.
However, those models are “closed,” in that they are not available for the public to download and tweak. It’s unclear whether Meta will keep the 405B model closed, or release it openly to the public like the 70B and 8B variants.
An open release of the 405B model would seriously threaten to steal customers from rival AI companies that sell chatbot access. Who wants to pay when you can get something even better for free?
Whether open or closed, the 405B model and the entire Llama 3 release are a clear victory for Meta, who is also poised to benefit from the potential demise of American TikTok.
Top AI CEOs Paint Diverging Pictures of AI Timelines
Three CEOs of leading AI companies outlined contrasting sketches of the speed of upcoming AI progress.
First, Anthropic CEO Dario Amodei spoke in a fascinating interview with Ezra Klein of The New York Times.
Leading AI systems today cost between $33 million and $300 million to build, according to Amodei, a claim confirmed by Stanford’s latest AI Index. And the models that will come out later this year or early next year are “closer in cost to $1 billion.”
Further, Amodei expects that in 2025 and 2026, model costs will rise as high as $5 billion or $10 billion.
In line with his perspective on rapidly rising costs, Amodei expects AI models to quickly grow more powerful and risky. He predicts that capabilities that create significant biological and cyber misuse risks "could easily happen this year or next year.”
In contrast, Meta CEO Mark Zuckerberg offered a vision of bottlenecks keeping AI progress slow and steady, in an interview with Dwarkesh Patel.
For example, Zuckerberg highlighted that companies will be reluctant to invest vast sums of money into building AI systems unless they can recoup the costs.
Additionally, Zuckerberg highlighted potential energy constraints preventing companies from building ever-larger AI models. “I don't think anyone's built a gigawatt single training cluster yet,” said the CEO.
However, Amazon already bought a data center near a nuclear power plant that might use nearly a gigawatt of electricity, so Zuckerberg should arguably be using higher numbers.
Somewhere between Amodei and Zuckerberg’s perspectives is Mustafa Suleyman, the newly minted CEO of Microsoft AI, who took the stage at TED 2024.
Suleyman believes about five to ten years remain before AI systems gain the ability to self-improve their own capabilities without human assistance.
“I think AI should best be understood as something like a new digital species,” said Suleyman, arguing that this is “the most fundamentally honest way of describing what’s actually coming.”
Senators Introduce the Future of AI Innovation Act
Senators Maria Cantwell (D-WA), Todd Young (R-IN), John Hickenlooper (D-CO), and Marsha Blackburn (R-TN) introduced new legislation that would solidify and create several new AI projects in the federal government.
First, the Future of AI Innovation Act would formally establish the US AI Safety Institute, giving it a mission to develop voluntary best practices for evaluating AI systems, provide technical assistance for AI adoption across the federal government, and develop guidelines to promote AI standards and innovation.
Another section would establish programs focused on evaluating AI systems, including a project specifically targeting advanced general-purpose AI models.
Separately, the bill would direct several agency heads to coordinate with foreign allies to develop and harmonize international standards for AI.
The remaining programs in the bill largely focus on promoting AI innovation. There are testbeds for synthesizing new materials with AI, prize competitions to stimulate AI research, a report to identify “regulatory impediments” to AI innovation, and more.
The Center for AI Policy welcomes the release of this bill. We are thrilled to see several of the bill’s programs, such as efforts to test for biological, cyber, and nuclear threats. However, we are concerned that the bill relies exclusively on voluntary standards for reducing AI’s risks.
News at CAIP
On Tuesday, April 23rd, we hosted a discussion on AI, Automation, and the Workforce in SVC 212 inside the Capitol Visitor Center. The speakers were Professor Simon Johnson of MIT and Professor Robin Hanson of GMU. If you missed it, you can watch a recording here.
To accompany our briefing, we released a research report on AI’s workforce impacts.
We’re hiring for two different roles: External Affairs Director, Government Relations Director.
We released a statement on the Future of AI Innovation Act.
We were proud to sign a letter to the Senate and House Appropriations Committees calling for increased NIST funding.
Quote of the Week
A few years ago I heard a wonderful definition of intelligence, which I’ve been mulling over ever since. And it comes from Jean Piaget, who once said that intelligence is ‘knowing what to do when you don’t know what to do.’ [...]
A raccoon or a bear is a very intelligent animal, but put them in a really novel situation, and they're stumped. And there’s no sign that they know how to do what we know how to do, namely, to think it over and figure out some novel response. [...]
If you're going to have intelligence in Piaget’s sense, you’re going to have to be an agent like us. Not just a machine, but an agent with an agenda. And agents like us are dangerous. We're dangerous because we're not readily controlled by other agents.
—Daniel Dennett, a renowned philosopher and cognitive scientist who passed away last week, speaking at an event in 2020
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