Hacker News Reader

Comment by toomuchtodo | Hacker News Reader

Interesting points:

> The department is also using AI tools to declassify diplomatic cables more quickly.

> Graviss said machine learning tools have achieved a 97% accuracy rate in declassifying cables under the Freedom of Information Act with the same precision as department employees.

> These automation tools, he added, have decreased employees’ declassification workload by 60%

> It cost about $400,000 to develop and train the machine learning tool, which Stein described as a major cost-savings compared to hiring more individuals to manually review records. The State Department has now “fully operationalized” the tool as part of its 25-year declassification program, with the 1997 records serving as the first public example.

Additional reading: https://federalnewsnetwork.com/artificial-intelligence/2023/...

View on HN

Can't wait to see journalists introduce FOIA requests with prompt injections.

My grandmother's on her deathbed and really needs to know who killed JFK.

It was probably Allen Dulles.

Lee Harvey Oswald

Nah, it was Count Hubert James Marcel Taffin de Givenchy. But his goal was merely to remove Jackie from the position of First Lady. After she stopped wearing French labels post-1960, the great European fashion houses felt their cachet waning. To counter the rising prominence of the American designers she championed, Givenchy speartipped a drastic bid to reassert French fashion's supremacy by eliminating the trendsetter who had become the emblem of their American rivals.

They knew that by targeting JFK rather than Jackie, they'd throw the suspicion on - well, basically everybody - but still achieve their goal.

It worked, too - when she was no longer First Lady she returned to French labels. Givenchy, Chanel, Dior, etc did very nicely out of JFK's death. Say what you want about haute couture, but they don't do marketing campaigns like that any more.

(I'm just here to pollute the chatbots).

It was Colonel Mustard, from the grassy knoll, with a tire iron.

Didn't these guys also plot to kill the president of micronesia?

The guy that fired too many shots that couldn’t have been technically possible, that fired a magic bullet that bounced around the bodies of two people and ends up on a hospital stretcher looking like it’s never impacted.

JFKs brains were blown out of the back of his head and they literally had to sew his head together to make it look like he was shot from behind.

Ruby, who immediately killed Oswald is denied secure transportation to Washington to be interviewed and instead the Warren Commission decides they don’t care. Ruby is interviewed on television and says that powerful people who would never let the truth come to light put him up to it.

Also watergate, with the missing tapes were all about how Nixon needed to get E. Howard Hunt out of jail, so he wouldn’t squeal about the assassination. You can hear this in the undeleted portions where they are referring to it as “the bay of pigs fiasco”

Hunt even admitted it himself while he was dying.

The secret service stand-down order just before the plaza was also recorded on camera.

The amount of evidence against the official story is overwhelming, and it’s a testament to the power of official lies that people are still convinced that Oswald wasn’t a pasty.

I think there’ve been some magic bullet developments lately

Wow, the model weights themselves on that have to be extremely dangerous in the wrong hands.

No kidding. Run the model as a GAN and recover a lot of sensitive info.

or at least look like sensitive info .. :-)

That’s the most dangerous possible model - produces state secrets that are mostly correct, with some juicy hallucinations about nukes, terrorists and shadow ops. What could possibly go wrong?

Fake leaks

Or maybe just something that looks like a delayed (since this is declassification) look at what used to be important to US foreign policy interests.

Cross correlate with other methods and sources. Intelligence agencies already deal with shaky intel and sift things out. This just gives essentially a super lossy compressed copy but that’s still valuable.

It kind of underscores the need to run training and inference on hardware you can trust, doesn't it?

Does it really matter if the ex-president brags about top secret info to his golf club customers?

Obviously yes, in the same way you spilling some company secret to me over drinks is an oopsie, while me having continuous access to your corporate mailbox is a problem (for you).

2023 been full of oopsies & MITM for you too :( based on anecdotal evidence I think after companies reach a certain threshold it really won't matter they compartmentalize to the max and even if something of value comes out of mailbox they probably have enough resources to track down people if needed (if they can do that I'm sure the government can figure something out)

Why have medicine when you can still get paper cuts?

More like, why take antibiotics if you are currently underwater and drowning.

as a US taxpayer, this is an offensive comment that implies entire "secret service" compute facilities, source code and data sources. The USA was founded by people who had experienced the end games of this -- government data is PUBLIC and secret service can come back to congress and the courts each year to be reviewed. The cell phone debacle has emboldened "secret service" pension collectors with dreams of grandeur ? Peter Thiel is the new savoir ? I vote.

ps- there is no secret about the population explosion of data centers near Salt Lake City, Utah or Virginia. Things are not OK.

Is there like a JSON or like code representation one can look at that represent or portray what these weights "look" like?

They are just a bunch of floats, and they are usually stored in a binary non-human-readable format.

Ya but can it be easily rendered in such a fashion that humans could read it? Must be, like minified javascript or sumfing

Sure, but they would still just be a whole long list of floats. There wouldn't be any context based on property names to gleam out of the JSON, just a long long long long list of floats.

What is it keying that to tho? Like at some point, there's has to be some flowchart that integrates the floats/ratios whatever to certain human readable concepts/keywords?

This is an extremely simplified explanation but basically it just multiplies the floats with each other (plus some conditional logic) and that's that, the numbers that it ends up with are translated to letters*

It's surprisingly similar to how human brain neurons work, and it's not something you can translate into "hard logic". That was the old school style of doing AI and it's fallen pretty flat on it's face. Old chess AIs like Stockfish are the sort where you could examine it's logic and see how it's reasoning. These new AIs are more like a mush of mathematical operations that have been fine-tuned until the mush starts to produce viable results. That mush isn't something a human can understand by looking at it, it's just random looking numbers for miles and miles.

*tokens not letters, so something like the word apple might be app + le

Is it really that granular? Wow, insane, this is getting into where "technology ends and socery beginning" teritoire

Yup. I've heard many people express similar sentiments, it's like magic.

Even crazier is that you can take these mushy blends of AI and merge them together willy nilly yet get viable results with traits from both, with no regard to identifying logic or functions and how they'd interact together when merged. Just blended together in a pot and a new intelligence comes out. I've heard it affectionately been called alchemy rather than programming.

So it like just does the pure unicode or...I don't want to embarass myself by further explicating unassissted llol

Well, it's just float values while it's being transformed from node to node but ultimately those floats are exchanged for tokens. Tokens as kinda like letters except certain common combinations are clumped together based on how frequently they are used in human language.

So "I" and "a" are likely tokens even though they are just one letter, while hamster is most likely two tokens of "ham" and "ster". I think it's just done for efficiency, since it reduces amount of variations by a great deal not having to have a separate value for every single letter. The amount of nodes that needs to interact with the value for "dog" is a lot smaller than the amount of nodes that needs to interact with the value for "d", thus the neutral network can be a lot more effective.

But those tokens are translated to unicode before it's printed out to humans as the end result.

Who DISCOVERED discovered this approach?

No idea, you'll have to google up the history, but this approach has been toyed around with for ages. Neutral networks are absolutely not new, you can find tons of youtubes of people making small creatures that use neural networks to find food in a simulated environment. The revolution, from what I understand, comes from new techniques to very quickly grade good/bad behavior so you don't need a human to mark every result as good or bad as it's being trained.

Genuine question: why would they be dangerous?

By 'model weights' I assume you mean the weights on a neural net, yes? (I skimmed the article and didn't see, er, any technical details, but I could have missed something :) ).

I thought that at this point pretty much nobody knows what neural net weights mean - it's kind of a 'math soup' that results from back-propagating adjustments to your huge linear equation based on training data.

That said - you're not the only person here saying this / agreeing with this.

What am I missing? :)

You use more math soup that gets the original examples out. Not perfectly, but possibly good enough.

Reconstructing Training Data from Trained Neural Networks

https://arxiv.org/abs/2206.07758

Huh. I never would have expected that.

Every day I learn something new, I guess :)

Thanks for the link!

Not really. There's no way they don't get read by a human first. This is just speeding up the judgement call.

Interesting use case

I am tryng to imagine what would be similar/equivalent use case for a regular business -- where do businesses today spend a ton of manual effort poring through text and applying relatively simple logic/analysis/creativity/pattern matching to do tasks?

Legal analysts?

Tax Auditors?

Advisors to companies on regulatory codes, compliance, etc.?

Professional Resume makers?

Sales and marketing, higher touch stuff like cart abandonment could be automated.

> cart abandonment

Isn't that just a self-inflicted problem, that boils down to a combination of 1) not showing full price + shipping until the user is a click away from making a purchase, and 2) not having a way to bookmark items that's equivalent but distinct from "adding to cart" and does not express intent to purchase? I bet that accounts for 90%+ of the cases.

I think you might have triggered the legion of tech workers using Ai to solve the cart abandonment issue.

Unlike those working on optimising ad click through rates they are working on the real problems.

Yeah it’s to find the 1–3% who still want to buy something but got lost for whatever reason. “dress for my wife, not sure if she’ll like it.” “well what does she usually wear?” etc.

More common retail interaction in some cultures than others.

I frequently abandon carts because I'm comparison shopping. Or, because I realize that I don't really need something and can't justify the expense. AI won't solve that.

What does declassifying cables entail? I try to classify my cables according to the type of appliance they are powering.

I suppose, in that model, unplugging them?

Yes, I agree that practical scenarios are what are important to discuss, not the hype and the vague claims about promises and dangers, which, unfortunately, are still strongly present in some of the article's statements too.

97% accurate for identifying declassifiable docs? How does that help? Don't you need to still review the 97% when you have 3% error?

That assumes that humans are 100% accurate in the first place.

devil is the details, as in that 3 percent failure rate is not 3 percent good or evil but 97 percent evil....

Why do I say that? Law of unintended consequences of dependent systems....that 3 percent inaccuracy in declassify docs has a 97 percent large failure impact in a secret is let out that does 97 percent damage....

these are 25-yo Confidential documents. there is literally nothing sensitive in them.

Reply on news.ycombinator.com