Democracy and artificial intelligence can work hand-in-hand if industry leaders build it out that way — and they should, according to two local AI experts.

“If we harness these democratizing forces in the AI marketplace, we could produce an AI ecosystem that is more accessible, more affordable and potentially even more trustworthy and secure than a more centralized and concentrated marketplace would be,” said Harry Krejsa, the director of studies at the Carnegie Mellon Institute for Strategy & Technology who previously worked as National Cyber Director at the White House.

Mr. Krejsa and Thomas Şerban von Davier, an AI research scientist at CMU’s Software Engineering Institute, wanted to better understand why AI is proliferating and getting cheaper and easier to use. 

The answers they found were recently published in a policy brief titled “Agents of Change,” in which they offer a vision forward for AI, one where accessibility and smart policy reign supreme.

Chloe Jad

State legislators are pushing for artificial intelligence companies to take root in Pa.

“This transition toward more flexible, affordable, and distributed AI systems is also reshaping how we have envisioned their deployment, from the way we build data centers and power infrastructure to how we will manage these systems’ cybersecurity and data governance,” the paper says.

Their conclusions grew partly out of AI success in China.

DeepSeek, a Chinese artificial intelligence startup, set the pace with its innovations in cost and efficiency — being able to run DeepSeek (a chatbot platform similar to ChatGPT) at a fraction of the price without losing much performance quality.

Then the company made its system free to download to allow for modifications and customization by the general population.

DeepSeek’s cost-cutting innovations made some investors in the industry uneasy, but “Agents of Change” emphasized that China’s innovations were good news for everyone — that “a more inclusive AI economy is a rising tide that lifts all boats.”

Making difficult tasks ‘more plausible’

With DeepSeek’s efficiency breakthroughs came the understanding of something called MoE, or Mixture-of-Experts. Instead of employing an advanced AI model in its entirety for any given task, the MoE design technique allows for the model to “selectively activate” the parts of it that are specialized, or “experts,” to the task, the paper says.

As AI evolves, it is expected to become more specialized, efficient and smaller, which would impact the infrastructure around the deployment of AI — moving AI models closer to the user, the paper argues. Imagine a company like Google having its own expert models for smaller, more manageable tasks, and leaving the complex computes to more powerful data centers. That setup can also allow for a degree of security, the experts argue, with less data leaving facilities.

Chloe Jad

New AI data center opens in Robinson, riding a wave of such developments

Although DeepSeek didn’t invent the idea of openly distributing software, they “reinvigorated interest in such models after demonstrating how quickly the market can transform when performance, efficiency, and accessibility align,” the paper says. 

In other words, it showed that openness can be central to the harmony between innovation and security.

“A lot of security-oriented institutions in government and around the world thought that doing secure stuff, building security and cryptography, secret codes in darkness and in private was the only way that they could be made secure,” Mr. Krejsa said. “And it was a long and hard-learned lesson that actually doing big parts of that in the light, in an open source ecosystem, produced better, more secure outcomes.”

Upgrades in efficiency, affordability and accessibility are essential for the “agentic” future of AI — that is, the ability of AI to solve problems, create, and more without human oversight — has long colored the imaginations of everyone from filmmakers to computer scientists. And Mr. von Davier and Mr. Krejsa envision systems that work together to problem solve and increase accuracy through “multi-agent” collaboration, or systems operating together that will also create security protocols.

“If you have multiple models at your disposal, multiple agents, able to pursue complementary tasks at the same time in an affordable way,” Mr. Krejsa said, “it makes more ambitious and difficult tasks more plausible.”

But there’s a long way to go.

“This is still research that’s being done,” Mr. von Davier said. “And it is research that requires the combination of focus and effort from policy makers in the government, it requires business leaders to make certain business decisions about how they want to invest their time, their energy and their resources, and of course, supporting research institutions.”

‘No one has it solved yet’

With the landscape of AI always changing, Mr. von Davier said that continued research — and investment in that research — is crucial to mapping out AI’s future, and what policy makers should expect.

“This synergistic relationship between research institutions, universities and industry, providing positive feedback loops into each other of innovation, of market demand signals, create durable hubs of growth and technological advancement,” Mr. von Davier said. “And I think Pittsburgh has long been at the forefront of that ecosystem, of that dynamic.”

The Steel City is a “poster child for post-Rust Belt renewal” in the arena of advanced and emerging tech, he said — so it makes perfect sense that the Energy and Innovation Summit is coming to town on Tuesday.

State lawmakers and industry leaders now must evaluate the capacity and health of Pennsylvania’s energy infrastructure to keep up with AI’s forever-shifting demands and capabilities — especially if Pittsburgh wants to be at the forefront of that growth, Mr. Krejsa said.

“One of the reasons why the AI and energy conversation is so prominent right now is because electricity growth in the United States has been flat for decades,” Mr. Krejsa said.

And now AI has come along, hungry for water, electricity and land.

To those nervous of the implications, both environmentally and economically, of developing infrastructure to support AI, Mr. Krejsa said that “more electricity and more compute gives us the power to solve more problems.”

Indeed, this could be humanity’s chance to build a better grid, he argued, one that is “more sustainable and secure and resilient than the one that we had before.”

In the report, Mr. Krejsa and Mr. von Davier gave four main recommendations to capitalize on AI while mitigating risks: Swift deployment of more electricity; strengthening open-source AI development; establishing interoperability protocols for agent collaboration; and incentivizing business models and equipment that support locally-executable AI.

“Some cool and amazing things can come out of China through DeepSeek; some cool and amazing things can come out of Silicon Valley,” Mr. von Davier said. “But no one has it solved yet.

“And in order to actually get a few steps closer to that, research has to keep going, and research investment has to keep going.”

First Published: July 14, 2025, 8:00 a.m.
Updated: July 15, 2025, 5:51 p.m.