Electrical engineer Gilbert Herrera was named research director of the US National Security Agency in late 2021, just as an AI revolution was brewing within the US tech industry.
The NSA, sometimes jokingly said to stand for No Such Agency, has long hired top talent in math and computer science. The tech leaders were early and avid adopters of advanced computing and AI. And yet, when Herrera spoke to me by phone about the implications of the latest AI boom from NSA headquarters in Fort Meade, Maryland, it seemed that the agency, like many others, was stunned by the recent success of the big language models behind ChatGPT. and other popular AI products. The conversation has been lightly edited for clarity and length.
Gilbert HerreraWith thanks to the National Security Service
How big was the ChatGPT moment for the NSA?
Oh, I thought your first question would be, “What did the NSA learn from the Ark of the Covenant?” That’s been a recurring problem since about 1939. I’d love to tell you, but I can’t.
What I think everyone learned from the ChatGPT moment is that if you throw enough data and enough computing resources at AI, these emergent properties show up.
The NSA really sees artificial intelligence as the frontier of a long history of using automation to accomplish our missions with computers. AI has long been seen as a way for us to operate smarter, faster and on a larger scale. And so we have been involved in research for over twenty years that has led to this moment.
Large language models have existed long before generative pre-trained (GPT) models. But this “ChatGPT moment” – once you could ask him to write a joke, or once you can strike up a conversation – really sets it apart from other work we and others have done.
The NSA and its counterparts among America’s allies have occasionally developed important technologies unlike any other, but have kept them secret public key cryptography in the 1970s. Could the same have happened with large language models?
At the NSA we couldn’t have made these big transformer models because we couldn’t use the data. We cannot use the data of US citizens. Another thing is the budget. I was listening to a podcast where someone shared a Microsoft earnings call, and they said they were spending $10 billion per quarter on platform costs. (The total US intelligence budget by 2023 this would amount to $100 billion.)
They really need to be people who have enough money for capital investments of tens of billions and who have access to the kind of data that can produce these emerging properties. And so, in reality, it’s the hyperscalers (the largest cloud companies) and possibly governments that don’t care about personal privacy, don’t have to abide by personal privacy laws, and have no problem with data being stolen. And I’ll leave it to your imagination who that could be.
Doesn’t that put the NSA – and the United States – at a disadvantage in gathering and processing intelligence?
I’ll push back a bit: it doesn’t put us at a major disadvantage. We have to work around it a little bit, and I’ll get back to that.
It is not a major disadvantage to our responsibility, which is to deal with national objectives. If you look at other applications, this could make things more difficult for some of our domestic intelligence colleagues. But the intelligence community will have to find a way to use commercial language models and respect privacy and personal freedoms. (The NSA is prohibited from collecting domestic intelligence, although several whistleblowers have warned that it does collect U.S. data.)