[dofm]

Serious question: do you think the NSA aren't training their own LLMs? (With or without Anthropic and OpenAI's help)

It's a perfect technology for their uses, they get a big chunk of a $100 billion black budget, and they've had access to the research for at least as long as we have.

[xeubie]

I can't say what they're doing now because I worked for the NSA 15 years ago but the view of them as an omnipotent power is a product of Hollywood. The government is good at throwing an ungodly amount of resources at something to get a result through sheer attrition, and so they are often the source of original development of technologies. The private sector has always been much better at building a technology to greater sophistication and efficiency. There may be blue badgers in Fort Meade trying to train models but there is no chance they are competitive with the frontier AI companies. It's like saying the government has an amazing home-grown fighter aircraft that is beyond what Lockheed has ever made...they delegate that stuff to private companies for a reason.

[LPisGood]

I’ve heard of “blue suiters” for air force brass, but never blue badgers.

Anyways, isn’t NSA one of the largest employers of mathematicians in the world? Surely they’re doing something useful.

[rob74]

Cryptography, I guess? Not really related to LLMs...

[zhoBEENG]

Crypto and AI are deeply connected, and you see similar structures/problems in both. Shannon, the “Father (or whatever) of AI”, worked for the NSA and published many papers there that were later declassified.

Here is a banger quote on this by Shannon’s boy Warren Weaver, keeping in mind LLMs came from translation problems:

“One naturally wonders if the problem of translation could conceivably be treated as a problem in cryptography. When I look at an article in Russian, I say: 'This is really written in English, but it has been coded in some strange symbols. I will now proceed to decode.”

[xeubie]

Blue badges were for government employees (like I was), and green badges were for private contractors. And yes they have a lot of math and physics guys; my own physics lecturer was in my orientation class, actually. He was there for quantum computing, which reinforces my point. The government can be good at pioneering unproven / uncommercialized technologies, but in general they are like a blunt weapon; the profit motive and lack of bureaucracy eventually makes the private sector far better for improving the technology later. In the case of LLMs, they didn't even originate in government, and I don't think there's any chance they are being developed there at a more advanced level.

[ben_w]

> Serious question: do you think the NSA aren't training their own LLMs?

Given the evergreen discussion of "are these companies making a profit"*, I think any LLMs that the NSA (or any other government agency worldwide) may be making are quite far from the leading edge.

* Person A: "they are making a loss!" Person B: "Only if you count training, they make a profit on inference, look at what it costs to run comparable open models on generic cloud servers" A: "Sure, but if they don't train new models they'll be left behind, so they're still making a loss"

That and the way compute is now measured in GW, I think even random low budget vloggers just getting started would be able to spot if the NSA was doing anything significant just from the extra heat emissions or power plants getting built.

[ACCount37]

Model training does NOT dominate the model costs.

The rate of inference compute to training compute is ~10:1, for popular frontier models. Models are routinely overtrained past the Chinchilla optimum now because it makes an immense amount of economic sense to do so.

Worse the more niche and unused your models get, but when this "making a loss" fuckery pops up, it's usually about the big guys like Anthropic, OpenAI, GDM and maybe xAI and Meta. Of which only the latter can be accused of not selling enough inference to offset the training runs.

The real money sinks are: R&D and infrastructure buildouts.

[HlessClaudesman]

I don't think there is much overlap between people capable of building cutting edge LLM's and the people who want to build a cutting edge LLM for the government.

[dofm]

The NSA managed to deliberately insert a backdoor into elliptic-curve cryptography right under the noses of everyone capable of making elliptic-curve cryptography.

I wouldn't count them out.

[tux3]

Mathematicians in academia are paid a little less than AI researchers. Companies are willing to pay billions to steal the few people capable of driving development of frontier LLMs from each other. Cryptographers don't quite enjoy the same popularity.

[mpyne]

> The NSA managed to deliberately insert a backdoor into elliptic-curve cryptography right under the noses of everyone capable of making elliptic-curve cryptography.

That sort of proves the opposite point, assuming you're referring to Dual EC DRBG, because the flaw was noticed very early on, by people who weren't even involved in its development.

[curt15]

Would they be able to hire top ML talent with US government salaries?

[polytely]

They probably also have an insane dataset

[doug_durham]

The NSA is government agency. They are certainly not training any world class LLMs. They probably have some specialized fine tunings of existing models, but that's it. They don't have the capacity.

[stronglikedan]

> do you think the NSA aren't training their own LLMs?

They probably already have access to Sentinel, so they wouldn't need to train their own.

[segmondy]

Serious question, do you realize that the NSA are mere mortals? Do you realize how much it takes to train a model? Does the NSA make their own chips or planes? The NSA buys a lot of technology because they can't make their own.

[dofm]

You mean "Rhetorical question," and I didn't need patronising.

They have at least one pretty vast, largely classified data centre in Utah, with a sizeable chunk of the black budget and they also have pretty large data sets.

[halJordan]

Whats in Utah is data storage.

[convolvatron]

NSA has had their own supercomputing program for decades. they design and produce their own large scale machines. chips, fabrics, arithmetic units, all of it. they also employ quite a number of hardcore mathematicians, computer scientists, and systems wranglers. if they decided it was of strategic importance there is absolutely no reason they couldn't train their own models.

[distill17801]

I guess we're just conspiracy theorists for landing at the objective conclusion that three letter government agencies:

- find "modern AI" to have strategic importance

- have ways to spend loads of money while having a front-facing budget on the record

- could be running a PR program to have Americans think they "buy" access to models like they do, but the AI companies were taken over by these agencies long ago

Look at Google, Microsoft...Apple got away with it by having as much on-device operation as possible so they could wash their hands, honestly saying "We don't have it."

This is the world's largest data gathering operation. Remember after 9/11 when the NSA copied as much Internet back bone traffic as they could?

I'm not for or against, even as a resident, but we certainly shouldn't be naive.

[convolvatron]

as someone who actually worked at the NSA pointed out earlier in this thread, they have plenty of resources, but also plenty of politics and some execution problems. so I wouldn't put money on them making a great model, but to say that they are completely incapable of doing anything is probably quite wrong.

the issue here that is a forgone conclusion, regardless of where the model comes from and which chips it runs on, is that now they can reasonably comb through all the stuff that they've been collecting. that's a pretty huge operational change.

[dgellow]

You cannot really hide the amount of compute required to train an LLM. Do we have actual clues that NASA is training their own frontier model?