TThis week, the National Science Foundation (NSF), in collaboration with 10 other federal agencies and 25 private sector and nonprofit organizations, launched a new initiative to democratize access to expensive infrastructure needed for cutting-edge AI research. As a first step, they announced the launch of a pilot program.
The National Artificial Intelligence Research Resource (NAIRR) pilot aims to provide expensive computing power, datasets, AI models, and other tools to academic AI researchers who are increasingly struggling to access the resources they need. The purpose is
Chipmaker Nvidia, one of the companies participating in the program, announced it would donate $30 million worth of cloud computing resources and software to the pilot over two years. Meanwhile, Microsoft announced that it would donate his $20 million in cloud computing credits in addition to other resources. . His OpenAI, Anthropic and Meta, leading companies in the field, are reportedly providing access to their AI models.
The NAIRR pilot comes at a pivotal time for AI research. As tech companies spend billions on acquiring computational resources, datasets, and hiring skilled talent, researchers in academia and the public sector are being left behind. As a result, important research directions and basic scientific research remain unexplored. But commentators warn that this pilot is just an early stage and that closing the AI gap will require sustained and ambitious government investment.
industry moves forward
AI systems have three inputs: computational power (often referred to as “computing”), data, and algorithms. More data and computing, and better designed algorithms will result in more capable AI systems. Industry’s access to all three AI inputs will become increasingly privileged, resulting in a widening gap between AI systems built by companies and those built by researchers in academia. did.
read more: Four charts showing why AI progress is unlikely to slow down
Decades ago, most of the breakthroughs came from researchers in academia, said Nur Ahmed, a researcher at MIT’s Sloan School of Management. “Rather than trying to push the boundaries, scholars are now conducting further follow-up and continuing research.”
Whereas in the past, the most capable AI for a given task would have been built by academics, almost all cutting-edge AI systems now involve at least some collaboration with industry, and many Completely built by industry.
![](https://api.time.com/wp-content/uploads/2023/09/AI-benchmarks-chart.jpg?quality=85&w=113)
In reality, computing means access to specialized semiconductor chips that are expensive and rare. As access to computing power becomes more economical over time, the amount used to train AI systems has steadily increased, and since the dawn of his AI in 1950, once every 20 months At the rate he has doubled. More computing power greatly increases its power, so AI developers started training even larger models, and the amount of computing used doubled every six months. .
Since then, the amount of money spent on training AI systems has skyrocketed. Epoch researchers found that from 2009 to 2022, computing costs increased by approximately three times each year. Epoch’s data shows that academic researchers are effectively priced out of their developmental state. art models.
Much of the data used to train AI systems, especially language models that use large amounts of data collected from the internet, is publicly available. But Neil Thompson, director of MIT’s FutureTech research project, says industry still has two advantages over academia and the public sector.
First, processing the vast amounts of data used to train state-of-the-art AI models requires a lot of compute, which is made easier by having a team dedicated to cleaning and preparing the data. Both are available in industry, but not in academia. Second, companies often have access to unique datasets that are particularly valuable for specific purposes.
Researchers design algorithms. Therefore, organizations with access to the most talent tend to have access to more sophisticated algorithms. Following the release of ChatGPT and the ensuing artificial intelligence boom, the AI labor market has become incredibly hot, creating intense competition for skilled researchers and engineers, Thompson says. Companies are increasingly offering higher salaries to attract these workers, with job openings at Netflix last year offering salaries of up to $900,000. Pay disparity aside, Thompson says researchers are also attracted to the greater access to data and computing that the industry provides.
This dynamic may be bad for society as a whole, says MIT’s Ahmed. Commercial AI developers have their own incentives, and fewer academic research resources may mean less research is being done on socially important issues such as addressing bias in AI systems. , says Mr. Ahmed. A paper published in 2020 by researchers at the National Endowment for Science, Technology and the Arts confirms Ahmed’s concerns, stating that “private sector AI researchers are “There is a tendency to specialize in computationally intensive deep learning methods.” We will also consider the ethical implications of AI and its applications in areas such as health. ”
Left unchecked, private companies tend to lack funding for basic research, Thompson said. And without sufficient computing, academics and public sector researchers won’t even be able to check the work of industrial researchers.
bridge the gap
The pilot released this week has been a long time in the making. The NAIRR Act, passed in 2020, established a task force to create a roadmap for a national program to improve access to computing, data, and educational tools. The NAIRR working group’s final report, released in January 2023, estimated that NAIRR would require $2.6 billion over six years to operate and proposed a pilot as a way forward in the absence of full funding. President Biden’s AI Executive Order, signed on October 30, gave NSF 90 days (until January 28) to begin piloting NAIRR.
Divyansh Kaushik, deputy director for emerging technologies and national security at the Federation of American Scientists, who advised the NAIRR task force, said the pilot is welcome but not enough. Congress needs to pass legislation authorizing NAIRR and making the necessary funding available, he said, adding that most lawmakers support the program. “There’s really no opposition,” he says.
Such legislation was proposed in July when the leadership of the Congressional Artificial Intelligence Caucus introduced the CREATE AI Act to establish NAIRR. Sens. Martin Heinrich, Todd Young, Cory Booker, and Mike Rounds have introduced a companion bill in the Senate. “We pretty much stuck to implementing what the task force recommended. In my view, they did a very good job,” said California Democrat and co-chair of the Congressional Artificial Intelligence Caucus. Rep. Anna Eshoo told TIME in September 2023.
“NAIRR will provide researchers in universities, nonprofits, and government with the powerful datasets and computing resources they really need,” Eshoo said. “Ensuring everyone has access to the tools they need to research and develop AI systems in a safe, ethical, transparent, and inclusive manner.”
In addition to the NAIRR Act, Kaushik said lawmakers should take steps to expand government access to computing power. This could include building a new government supercomputer, in line with a U.S. Department of Energy report released in May with support from Sen. Joe Manchin, D-West Virginia. He suggests that there is.
“NAIRR is a very important first step, but it is just a first step. There’s not enough to meet that,” says MIT’s Thompson. “We need to continue to invest to scale further here.”