The NSA has long hired leading math and computer science talent, often jokingly referred to as No Like Agency. Its complex leaders have used advanced computing and AI frequently and first. However, Herrera seemed to have been stunned by the agency’s recent success of the big language models that have been the inspiration for ChatGPT and other popular AI products when he spoke with me by phone about the relevance of the most recent AI increase from the NSA office in Fort Meade, Maryland. For precision and length, the conversation has only been casually edited.
How surprising was the NSA’s reaction to the ChatGPT?
Oh, I did n’t realize your first inquiry would be “what did the NSA learn from the Ark of the Covenant”? That’s been a recurring one since about 1939. I’d love to tell you, but I ca n’t.
What I believe everyone at the ChatGPT moment discovered is that these emerging characteristics can be attributed to adequate data and computing resources to AI.
The NSA truly sees artificial intelligence as the pinnacle of a long tradition of utilizing technology to carry out our computation missions. AI has long been thought of as a means of operating at level, faster, and smarter. And therefore we’ve spent well over 20 years conducting research that is relevant to this situation.
Long before generative pretrained ( GPT ) models, large language models have been around. But it really stands out from other job that we and others have done because it can be asked to write a joke or engage in conversation with it.
In the 1970s, the NSA and its US allies have sometimes developed significant technologies before anyone else, keeping them a secret, similar to public key cryptography. Has the same thing possibly happened with big language versions?
At the NSA we could n’t have created these big transformer models, because we could not use the data. We don’t use US individual’s data. Another factor is the expenditure. Someone mentioned in a podcast that they were spending$ 10 billion a third on program costs. ]The total US intelligence budget in 2023 was$ 100 billion. ]
People who have access to the kind of data that can create these emerging properties must actually have the tens of billions of dollars to invest in these things. And so it really is the hyperscalers]largest cloud companies ] and potentially governments that do n’t care about personal privacy, do n’t have to follow personal privacy laws, and do n’t have an issue with stealing data. And I’ll leave it up to you to figure out who that might be.
Does n’t that put the NSA—and the United States—at a disadvantage in intelligence gathering and processing?
II’ll push back a little bit: It does n’t put us at a big disadvantage. We might need to operate around it, and I’ll do that.
It’s never a great disadvantage for our obligation, which is dealing with country- state targets. If you look at other uses, it does make it more difficult for some of our acquaintances that offer with local knowledge. However, the intelligence community will have to find a way to respect privacy and private liberties while using commercial language models. Although numerous reporters have warned that the NSA does gather US data, it is forbidden from doing so.
How might the NSA benefit from significant language models that are readily available for commercial use?
Slow engineering and simplifying cyber defenses are two areas that these massive models have shown they are competent at. And those points can be accomplished without being extremely restrictive in terms of privacy legislation because it can be taught on less sensitive program code.
Let’s say we wanted to make a” copilot” analyst tool that aids an analyst in their data analysis. If we wanted to accomplish that. Given the various regulations [about accessing US data], then we would have someone with logical skills in American society and the English language.
Hypothetically, we could use an LLM to just look at information that had passed our compliance testing using things like RAG [retrieval augmented generation, a method in which a language model reacts to a comment by summarizing trusted information].
How might the NSA’s language versions be made more difficult by the law?
It raises a question about the data engagement issues that we might have to deal with when we use some data that were used to teach models for really long periods of time. Another problem is that the whole web was flooded with information. You might also include copyrighted or personal information from US citizens. But you do n’t look at it]when feeding it to an AI model]. What day do all the rules take effect?
Because we already know the extent of their expenditure, I believe it will be challenging for the intelligence community to recreate something like GPT-10. And they can use data to accomplish things that no one in the state would ever consider doing.
Does the US face new security issues as a result of common AI apply?
On day one of the transfer of ChatGPT, there was evidence of improved phishing problems. And if it improves their victory price from one in 100, 000 to one in 10, 000. That’s an order of magnitude progress. People who do n’t have to worry about quantifying margins and uncertainties in the use of the product will always benefit from artificial intelligence.
So, does AI create a new frontier in data security?
They’re going to be big new security risks. One of the reasons we established an AI Security Center is because of this. There are many ways to harm a unit, according to the law. There are rotation assaults where you can try to steal some of the personal data from models and expert on them.
The second line of defense in AI safety is good security. It means protecting your types, protecting the information that’s in there, protecting them from being stolen or manipulated.
