How sure are you about that? The basic theories have been around since the 70s, have been proven at scale in the last decade, and the NSA has more data and compute than anybody else. I’d be shocked if they aren’t very far along in solving many problems.
This isn’t really a “throw money at it” problem like the government is good at.
Take drones for example. The government got really good at those because they made them jet-powered (lol) and blew a bunch of money on server-grade FPGA’s in each one of them.
You can’t really just buy a lot of GPUs to make an LLM work, you need iterative development of architecture and training methods.
Like maybe the government invented self-attention before 2017, but if they didn’t, then the constraint is training time, and the government has the same number of seconds as the rest of us.
The military invented the nuclear bomb yes. But Fermi did most of his thinking work in Italy before the Manhattan project. He got money thrown at him once he got here.
As for semiconductor devices, it was Bell Labs and TI.
Coding is an ambiguous concept that wasn’t really invented, but if it were, it would have first appeared in programmable looms.
The military likes to take credit for things, but really all they do is throw money at existing inventions.
I’m sure they’re throwing a bunch of money at Transformers now, but who are all these uncredited super geniuses who invent things and then let randos at Google take the credit/earn the money?
I used to agree with you. I used to think the military was kinda dumb. But after doing a deeper dive into past military technology, and present - I've come to realize this is just an intelligence ruse.
They made some mistakes in the 40's and 50's to where they had nuclear secrets stolen by the Russians. And ever since has been hyper compartmentalized.
It would not surprise me if in the 90's or 00's they had an internal working LLM, considering all the puzzle pieces. You will never hear about classified tech unless it's a bomb, gets leaked. (See code breaking machines declassed after 70+ years)
In a hypothetical scenario, a military organization might want to conceal its use of a large language model (LLM) for intelligence gathering and analysis.
Another scenario is the military's current interest in everything quantum. Quantum computers for example (you wouldn't want another nation being first and pirate baying out our secrets, would you) so there is an extreme national security importance of being first.
And to be first, you need to have the smart people, which the military has. There is a reason China struggles with jet engines 80 years after their invention, and still can't make nuclear carriers. While the US navy works on things like this: https://www.navair.navy.mil/foia/sites/g/files/jejdrs566/fil...
The military invented AI and NLP which underpins LLMs.
The military is responsible for most the technology we use and talk about today. The government may appear incompetent, but we’re living off military hand-me-downs, the entire world is
Your argument reminds me of the discussions about the moon landing.
If today's hardware was available 20 yrs ago, this would've been possible just like the moon landing could've been faked if it took place 20+ yrs later. The technology wasn't available at the time (GPUs in this case, and generally no experience in doing such advanced trick techniques for movies back then)
These models are having such a strong effect now because we've finally got the hardware to run them
I'm not so sure about that. I mean, maybe not since the Snowden leaks but how do we know that governments haven't been running their own LLMs for the last five years or so? We know that they're using sockpuppets[1]. We know that they're astroturfing[2]. Integrating LLMs into their toolkits seems like an obvious move, so obvious that they would be stupid not to do it.
>haven't been running their own LLMs for the last five years
Because the hardware has not existed.
This said by accident I've seen hardware that was brought to a testing company by federal marshals that was massively parallel custom hardware that was likely for signal processing a lot of channels at once. So there is plenty of custom hardware out there, but these items have not been produced at the scale needed (from what anyone can tell) and, again from what we can tell, they don't have the general processing capability that GPU/TPU driven LLMs have.
Yes, it has. Consumer-wise, we've had Dragon Naturally Speaking since the late 90s. It's pretty simple to have a script read what it outputs text-wise and look for key words. No AI is even needed to do this.