AI Policy Weekly #27
Apple’s Apple Intelligence (AI), Epoch’s analyses of training data and compute, and the EPIC 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.
Apple Integrates New AI Features Into Its Products
“Every once in a while, a revolutionary product comes along that changes everything.”
So said Steve Jobs when he unveiled the iPhone in the Bay Area on January 9th, 2007.
Today, seventeen years after the iPhone, approximately 99% of US adults aged 18–49 own a smartphone, and over half of these people “can’t imagine life without it.”
So the iPhone did not literally change everything, but it has changed modern life for just about everyone in the US.
The current wave of AI may cause a similar transformation. During this week’s Worldwide Developer Conference (WWDC), Apple announced plans to integrate AI (“Apple Intelligence”) into iPhones, iPads, and Macs.
Here are some of Apple’s new AI-powered features:
Writing Tools can “proofread your text, rewrite different versions until the tone and wording are just right, and summarize selected text with a tap.”
Text-to-image tools allow users to generate their own customized emojis known as “Genmojis.”
Image Wand can “transform your rough sketch into a related image in the Notes app.”
The Photos app can find appropriate photos or video frames in response to user queries.
The Clean Up tool can remove unwanted objects from photos and replace them with AI-generated filler.
Speech-to-text tools can transcribe audio recordings and phone calls.
Siri has “richer language understanding and an enhanced voice.”
Siri also has “onscreen awareness, so it can understand and take action with things on your screen.”
Siri and Writing Tools can access ChatGPT when users give explicit permission.
Many of the AI computations will run on the Apple device itself, but highly intensive computations will run on Apple’s “private cloud compute.”
Partly due to the on-device processing requirements, most of these new AI features are only available on the latest Apple devices.
But as time goes by and users upgrade their products, many more people will gain access to Apple Intelligence. The world currently contains over 2 billion active Apple devices, so Apple’s embrace of AI—not to mention similar projects at Microsoft and Google—could affect over a billion people in the coming years.
Zooming out, all of these features are using the current wave of AI, which largely looks like an exciting new tech product. But AI capabilities have rapidly expanded over the years and will continue to do so, so future waves of AI will probably have even larger societal impacts.
Epoch AI Analyzes AI Scaling Trends and Potential Data Constraints
Epoch AI is a research institute studying trends and questions that will shape AI’s trajectory and governance. They recently published three papers that cover three critical factors fueling AI progress: spending, compute, and data.
First, Epoch found that the hardware and energy cost for the final training run of frontier models has grown by approximately 2.4x per year since 2016.
Besides hardware and energy, AI companies also need expensive R&D talent for training these models. Epoch analyzed a few frontier models in depth and found that that R&D staff costs ranged from 29–49%.
The researchers concluded that “if the trend of growing training costs continues, the largest training runs will cost more than a billion dollars by 2027.”
These exponentially growing expenses should be visible in trends in the computational resources (“compute”) that AI models consume during development.
Indeed, a separate Epoch analysis found that “the amount of compute used to train notable models has grown about 4.1x/year” between 2010 to May 2024. A “notable” AI model is one that either earns many citations, achieves state-of-the-art performance, has historical significance, or receives significant use.
This growth in training compute has been a major driver of AI progress over the past decade. But can it continue?
One potential constraint is the availability of human-generated text data, the subject of Epoch’s third paper. The researchers predict that large-scale AI training projects will run out of public, human-generated text data at some point between 2026 and 2032.
However, Epoch also identified several promising innovations that could surmount this looming “data wall,” such as using AI-generated data or using multimodal data from other domains besides internet text.
Concerningly, last week’s newsletter detailed how leading AI companies currently lack the information security necessary to prevent such insights from leaking. If Congress fails to address AI security, the US could make major AI breakthroughs without gaining much ground over foreign adversaries in AI.
Stevens, Obernolte Introduce the EPIC Act
Last Friday, Representatives Haley Stevens (D-MI) and Jay Obernolte (R-CA) introduced the Expanding Partnerships for Innovation and Competitiveness (EPIC) Act. The bill received cosponsor support from Representatives Frank Lucas (R-OK) and Zoe Lofgren (D-CA), chairs of the House Committee on Science, Space, and Technology.
The EPIC Act would establish a nonprofit Foundation for Standards and Metrology to support the National Institute of Standards and Technology (NIST) in its efforts to improve measurement science, technical standards, and US technology.
Notably, the Foundation would be able to accept donations from private entities and philanthropic organizations, bringing resources to critical standards-setting efforts at a time when Congressional funding is hard to come by.
News at CAIP
Save the date: we’re hosting a panel discussion on Wednesday, June 26th from 11am–12pm titled “Protecting Privacy in the AI Era: Data, Surveillance, and Accountability.”
ICYMI: transcripts for all seven episodes of the Center for AI Policy Podcast are now available on Substack.
Jason Green-Lowe wrote a new blog post: “Apple Intelligence: Revolutionizing the User Experience While Failing to Confront AI's Inherent Risks.”
We expressed support for the EPIC Act.
We’re hiring for an External Affairs Director.
Quote of the Week
They have this ability to individually call specific members of their family with a unique call. [...]
A lot of interesting stuff is going on in the rumbles.
—Dr. Mickey Pardo, an acoustic biologist at the Cornell Lab of Ornithology, describing results from a paper he coauthored on using AI to decode elephant rumbles
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