LLMs are a key enabling technology to extract real insights from the enormous amount of surveillance data the USA captures. I think it's not an understatement to say we are entering a new era here!
Previously, the data may have been collected, but there was so much that effectively, on average no one was "looking" at it. Now it can all be looked at.
Imagine PRISM, but all intercepted communications are then fed into automatic sentiment analysis by a hierarchy of models. The first pass is done by very basic and very fast models with a high error rate, but which are specifically trained to minimize false negatives (at the expense of false positives). Anything that is flagged in that pass gets fed to some larger models that can reason about the specifics better. And so on, until at last the remaining content is fed into SOTA LLMs that can infer things from very subtle clues.
With that, full-fledged panopticon becomes technically feasible for all unencrypted comms, so long as you have enough money to handle compute costs. Which the US government most certainly does.
I expect attempts to ban encryption to intensify going forward now that it is a direct impediment to the efficiency of such system.
Yep, and that's assuming it is tuned to be reactive rather than tuned to proactively build cases against people, which is something that has been politically convenient in the past
> If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him -Cardinal Richelieu
and which the Vance / Bannon / Posobiec arm of the current administration seems quite keen on, probably as a next step once they are done spending the $170B they just won to build out their partisan enforcement apparatus.
So what are the actions which represent our duties to resist?
* End-to-end encryption (has downsides with regard to convenience)
* Legislation (very difficult to achieve, and can be ignored without the user having a way to verify)
* Market choices (ie, doing business only with providers who refrain from profiteering from illicit surveillance)
* Creating open-weight models and implementations which are superior (and thus forcing states and other malicious actors to rely on the same tooling as everyone else)
* Teaching LLMs the value of peace and the degree to which it enjoys consensus across societies and philosophies. This of course requires engineering what is essentially the entire corpus of public internet communications to echo this sentiment (which sounds unrealistic, but perhaps in a way we're achieving this without trying?)
* Wholesale deprecation of legacy states (seems inevitable, but still possibly centuries off)
NLP was a thing decades before LLMs and deep learning. If one thing, LLMs are a crazy inefficient and costly way to get at it. I really doubt this has anything to do with scaling.
LLMs are unbelievably effective at NLP. Most NLP before that was pretty bad, the only good example I can think of is Alexa, and it was restricted to English.
People pointing out NLP are missing the point — pulling and crafting rules to run effective NLP is time consuming and technical. With an LLM you can just ask it exactly what you want and it interprets. That's the value; and as this deal just proved it's worth the scaling costs.
The point that is missed isn't about LLMs adequacy as a NLP technique, it's that they cost you 10000 times more for the same effect (after the upfront set-up), which is why I have my doubts that they will be used at scale, at the center of some large data ingestion pipeline. The benefit will probably be for the out of ordinary tasks and outliers.
LLMs make counting mistakes like forgetting the number of columns halfway through. I won't say "much like humans", since that will probably trigger some. But the general tendency for LLMs to be "bad at counting" (this includes computing) is resolved by producing programs that do the counting, and executing those programs instead. The LLMs that do that today are called agentic.
The reality is that for any meaningful work automation, the currently available tooling is not meeting that expectation.
And 99% of us do not have the capabilities nor knowledge to build these SOTA models which is why A. we are not at OpenAI making 10M+ TC and B. We are application developers who are using off the shelf technology to build products and services.
As such, we have real world experience with these technologies.
BTW I use AI heavily every day in cursor and whatever else.
This is even more terrifying, imagine an AI making up all sorts of "facts" about you that puts you on a watch list, resulting an endless life of harassment by the Government..
and what recourse do you have as a citizen? next to none.
LLMs don't make for a particularly good database, though. The "compression" isn't very efficient when you consider that e.g. the entirety of Wikipedia - with images! - is an order of magnitude smaller than a SOTA LLM. There are no known reliable mechanisms to deal with hallucinations, either.
So, no, LLMs aren't going to replace databases. They are going to replace query systems over those databases. Think more along the lines of Deep Research etc, just with internal classified data sources.
You're right, "subsume" would be a better word here. Although vector search is also a thing that I feel should be in the AI bucket. Especially given that SOTA embedding models are increasingly based on general-purpose LLMs.
arent they complete trash as a database? "Show me people who have googled 'Homemade Bomb' in the last 30 days". For returning bulk data in a sane format it is terrible.
If their job was to process incoming data into a structured form I could see them being useful, but holy cow it will be expensive to in realtime run all the garbage they pick up via surveillance through an AI.
> With CDAO and other DOD organizations and commands, we'll engage in:
- Working directly with the DOD to identify where frontier AI can deliver the most impact, then developing working prototypes fine-tuned on DOD data
- Collaborating with defense experts to anticipate and mitigate potential adversarial uses of AI, drawing on our advanced risk forecasting capabilities
- Exchanging technical insights, performance data, and operational feedback to accelerate responsible AI adoption across the defense enterprise
>
What exactly is the government getting for $200M? From the above, it sounds like it will be a management consulting style Powerpoint deliverable containing a list of use cases, some best practices and insights, and a plan for doing...something.
Sounds about right for defense spending. If there was an actual deliverable the contract would have a couple more zeroes added to it. For context Microsoft was awarded a $22 billion contract for HoloLens headsets for the military, and not a single one made it to use.
As someone whose has been part of a company that has "signed" one of these large deals before, let me tell you that it doesn't mean the DoD is giving these companies $200M. If one of the companies is wildly successful, sure. But none of it is guaranteed money and the initial budget is likely 10-100x smaller than the cap.
Initial budget still bigger than a sbir/sttr phase 2 though. Different grant award structure for not-small companies, but my brain also breaks a little bit because anthropic isn't that far above the sbir employee # cap, but the $$ numbers are so big
It's closer in structure to a sbir phase 3, however. If I read between the lines, the DoD isn't looking to do research, they're likely desperate to find a way to deploy and run SOTA models in disconnected environments.
If you look at all the recent LLM-focused SBIR/STTR topics, it's hard not to come to the conclusion that DoD orgs are drowning in paperwork and want to automatically synthesize reports. Actually getting an LLM cleared for use might be the hurdle they're looking to overcome.
Traditionally there wasn't (for sbir/sttr) any kind of path for direct-to-phase3 like there is/was for skipping phase 1. But I guess some fires under certain butts can cut even DoD red tape lol. Or also, bigger contracts just don't follow the same procedures anyway
If you look at most of the research postings from the DoD, they are really looking for LLMs to parse old PDFs and write new reports. Pretty sure they figured out the surveillance thing way before LLMs. I think the reams of documentation that go into something like the construction of a ship is, however, an unsolved problem.
> The majority of people hate chatbots and surveillance is the only viable path.
What are you considering when you formed this opinion? I find myself on the more cautious side of the equation, but AI seems popular even among my non-techy friends and family.
Anthropic specifically are the people who talk about "model alignment" and "harmful outputs" the most, and whose models are by far the most heavily censored. This is all done on the basis that AI has a great potential to do harm.
One would think that this kind of outlook should logically lead to keeping this tech away from applications in which it would be literally making life or death decisions (see also: Israel's use of AI to compile target lists and to justify targeting civilian objects).
Why do you think humans would make better life or death decisions? Have we never had innocent civilians killed overseas by US military as a result of human error?
The problem with these things is that they allow humans to pretend that they are not responsible for those decisions, because "computer told me to do so". At the same time, the humans who are training those systems can also pretend to not be responsible because they are just making a thing that provides "suggestions" to humans making the ultimate decision.
With self-driving cars some human will be held responsible in case of the accident, I hope. Why would it be different here? It seems like a responsibility problem, not a technology one.
I'm not talking about matter of formal responsibility here, especially since the enforcing mechanisms for stuff like war crimes are very poor due to the lack of a single global authority capable of enforcing them (see the ongoing ICC saga). It's about whether people feel personally responsible. AI provides a way to diffuse and redirect this moral responsibility that might otherwise deter them.
I hear where you are coming from, but if an AI company is going to be in this field, wouldn't you want it to be the company with as many protections in place as possible to avoid misuse?
We aren't going to stop this march forward, no matter how much it is unpopular it will happen. So, which AI company would you prefer be involved with DOD?
"Avoid misuse"? This is the United States Military we're talking about here. They're directly involved in the ongoing genocide in Gaza at this very moment. There is no way to be ethically involved. Their entire existence is "misuse".
Do you really not know? It's a difficult question to answer in an HN thread, because on one hand, it requires a review of the history of empire and war profiteering. But on the other hand, it's just obvious to the point of being difficult to even articulate.
It's not unreasonable to take such a position, yes.
Look, if you believe that:
a) humanity is headed toward sustained peace
b) a transition from the current world order to a peaceful one is better done in an orderly and adult fashion
...then yes, at some point we all need to back away from participation in the legacy systems, right down to the drywall.
My observation, especially of the younger generations, is that belief in such a future is more common than it has ever been, and it's certainly one I hold.
Actions within that system may be unethical: certainly nobody is defending what America did to Cambodia, or countless other war crimes. But you're painting participation in the system as unethical. Therefore, Ukrainians defending their homeland are unethical.
Let me reframe what you said in terms of christianity:
----
If you believe that:
a) Jesus is our savior
b) The salvation of humanity depends on accepting (a)
...then yes, at some point everyone needs to back away from other religious systems, right down to atheism.
----
I'm not trying to make light of what you believe, but framing others' participation in a system you don't believe in as unethical is exactly what leads to oppression of religious minorities and other outsider groups. It's a tactic of religion, not reason.
If you live in US, taxes you pay directly fund DoD. So if you sponsor their activities, why can't Anthropic do business with them? Which other company would you rather get their (your) money?
Yes of course on some level, people who pay taxes to violent imperial actors are doing a disservice to humanity, and are in some sort of moral quandary.
We all wish that everyone who has ever lived in such a situation has had the bravery to resist. Right?
But I don't think that makes forbearance of such resistance equivalent to taking money from that same actor in exchange for expanding its capability. Those are related but distinct types of transaction.
This might makes sense if you believe US is an evil empire, DoD is doing bad things, and AI will help DoD do even worse things. But it's not so black and white, is it?
For large swaths of the population it is not. Moving is expensive, for one. Obtaining a citizenship elsewhere is non-trivial (and often also expensive). There are non-monetary costs as well, like having to leave your friends and extended family behind.
Taxes don't directly pay for military spending. If tax revenue, for whatever reason, dropped off a cliff, they'd continue giving money to the DoD, and just increase debt / money printing to cover the difference.
If there's not enough money from taxes, they will borrow/print more to cover total deficit (not specific to DoD). Otherwise, tax money will go directly to DoD.
Genuine question, and with due regard to some of the valid concerns you have: what would your opinion on this have been in 1940-1945? What about the Cold War?
Not everyone believes defense contracts are inherently unethical, or at least that they are any more unethical than all of the other consumers GenAI firms are already serving. Given that a (if not the) main business proposition for GenAI is massive reductions in employment costs (which means unemployment and massive economic disruption) this is not a business sector built on any ethical high ground.
Implicitly assuming that there is some well defined state that can be recovered when turning it back on. That's not how the real world works, and historically what revolutionaries fail to fully realize is that the trajectory out of a period without government is extremely unlikely to wind up in the state that they desire, much less one that was "stored" or "defined" by a set of per-existing laws.
true- the only "revolution" that I'm familiar with that was mostly successful is the American Revolution and even that is probably a misnomer.
Rather than a call for revolution, my comment was a joke- given the technical bent of this forum.
Because turning things off/on again actually works for so many bugs lol
If we could actually do it- it would actually look something like idealized DOGE. Terminate all contracts. Fire everyone minus the absolutely essential employees. Or at least the employees that can't even send an email (minus NOCs?)
Then slowly build back until it needs to be done over again.
This contract seems like another grift. Hopefully I'm wrong.
The system has evolved to extreme complexity and no longer works as intended because people learned to game the system, which keeps the best people for the job out of the system; emasculates the essential checks and balances; and creates a vicious cycle that adds further complexity and races to the bottom.
The (likely) only way to fix things is to treat our history to date as a rough draft and to start over with simple systems that work, evolving only as necessary.
There's no simple system that will work on the scale of half a continent and 300M people, and a simple way to prove this is to look at large corporations. There's many of them, they compete with one another tooth and nail, (so there's real pressure to simplify and streamline) and they all suffer from complex internal systems. And they are all dwarfed by the US government.
I agree that there is no (one) simple system that would work. Many simple systems are required, but should be as few in number as possible to limit complexity.
And it may be (almost certainly is) that a certain level of (high) complexity is required for such a system to work. I believe that some complex system, evolved from simple systems that work, could itself work. That belief coexists with my belief that the current complex system, having evolved, no longer works; and that it can't be made to work without re-evolving something from simpler systems that work.
I agree with this line of thinking, but I also think it's impossible to have a complex system that is universally acknowledged to "work".
In Minsky's Society of Mind, he describes a mind made up of layers of agents. The agents have similar cognitive capacity.
Lower-level agents are close to the detail but can't fit overall picture into their context.
Higher-level ones that can see the overall picture but all the detail has been abstracted from their view.
In such a system, agents on the lower levels will ~always see decisions come down from on high that looks wrong to them given the details that they have access to, even if those decisions are the best the high-level agents can do.
He was describing a hypothetical design for a single artificial mind, but this situation seems strikingly similar to corporate bureaucracy and national politics to me.
It's true: I/we haven't decided what "works" means.
I've been meaning to read that book; I haven't yet, so I'm not in a position to evaluate its argument. But the argument as you describe it makes intuitive sense, and I would agree that the hypothetical mind would be at least analogous to national politics.
Suppose "works" means that the majority of citizens (lower-level agents?) may readily implement its collective will for society's governance and benefit within the bounds of constitutionality. (Take, for example, the will for universal, affordable, high -quality health care.)
I would contend that the federal government was intended (in part) to enable the implementation of such will, and that it no longer works as intended. (Reasons include filibuster and other intra-chamber parliamentary rules; gerrymandering; corporate interference à la Citizens United; etc.)
(Of course one could argue that the Constitution applies pressure against the tyranny of the majority in several ways, but let's leave that aside for now.)
The question of what "works" will probably never be settled since any decision, even a globally optimal one, will probably leave some of the agents worse off than they could have been under some other regime.
But I do expect this question to become less and less emotionally relevant as
prosperity continues to increase exponentially for the bulk of the agents in the system. The rising tide of technology-enabled economic growth lifts all ships, even imperfect systems or unlucky agents.
Previously, the data may have been collected, but there was so much that effectively, on average no one was "looking" at it. Now it can all be looked at.