This is no way to develop advanced A.I. [Remember the bomb?]

Dan Rockmore, We Didn’t Build the Atomic Bomb This Way, NYTimes, July 3, 2026.

Artificial intelligence, funded overwhelmingly by private capital, has careened forward despite immense concerns about the effects it will have on labor, education, science, defense and civic life. A.I. companies have outpaced public oversight and, at times, successfully lobbied against it. The central achievements of the industry, the proprietary “frontier models” developed by companies like Anthropic, OpenAI and Google, are guarded intellectual properties even as they are incorporated into schools, offices, hospitals, courts, commerce and our everyday devices. The public did not ask for these A.I. tools and now can hardly opt out of them.

There is no one-size-fits-all solution to addressing the impacts that may be coming. But the scale of the concerns requires ambitious responses that serve the public. The United States can start by building a national A.I. laboratory.

Think about the atomic age. The early work in the United States was carried out across a classified but federally accountable archipelago with national laboratories in four sites: Berkeley, Oak Ridge, Hanford and, most famously, Los Alamos. They were expensive and sometimes secretive, but connected to universities as well as the rest of our nation’s research apparatus.

From its inception, the nuclear enterprise was regulated, imperfectly but deliberately, by the federal government. [...] Research in applied nuclear physics continues apace both in the open and in secret.

Imagine if, instead of a national effort to make an atomic bomb, a small number of private companies had gotten there first and began selling powerful nuclear technologies back to the nation, with the government trying to catch up afterward. Would we have accepted that arrangement? Would we have said that because innovation was fast, regulation should wait? Would we have trusted a few executives and technologists to decide what the public needed to know?

The analogy is imperfect — A.I. is not a bomb and we’re not fighting a world war. But like all technologies, it can be exploited for dark purposes. Yet the United States apparently has no public institution capable of building, testing and understanding frontier A.I. at the same scale as the companies that now define it.

A federal A.I. lab could change this by ensuring that the American public has at least one institution that can build, see and test the most advanced A.I. systems in the public interest. Such a lab should be built by, and provide controlled access for, university researchers, and it would be a training ground for our next generation of A.I. architects. It could publish research openly where possible, while classifying work only where necessary. It could study risks without needing to protect a new product launch. It could build models for science, education, medicine and national security — not as corporate afterthoughts, but as public missions that aren’t subverted by investor concerns.

The article goes on from there.