Read more about the Stargate Project here:

• apnews.com/article/trump-ai-op…
• bbc.com/news/articles/cy4m84d2…
• arstechnica.com/ai/2025/01/tru…

Or watch this good summary by Fireship:
" title=""/> bbc.com/news/articles/cy4m84d2…

Trump announces $500B “Stargate” AI infrastructure project with AGI aims

Trump: “The largest AI infrastructure project in history, and it’s taking place here in America.”…

^{Benj Edwards (Ars Technica)}

oh it kinda clicked for me that Apple/Google/Microsoft jamming AI down our throats is all about surveillance to understand what we're typing without file type/format or even end to end encryption mattering

very cool, I hate it here

Narrator:

"Ellison and is fellow oligarchs were sure that they alone could avoid surveillance. But what they didn't know was that they too would be watched, and their crimes recorded and delivered to the people who would become their de-facto masters."

an iron rod could help Larry Ellison forget this shit. He probably thinks it's some sort of Carnegie-style benevolent oligarchy; it isn't.

It's also far easier to just surveil the billionaires.

China has a much better surveillance system than us! We want too!

Watch this and tell me you really want that. That and worse is where Ellison's fantasy is heading.

imdb.com/title/tt5497778/

"Black Mirror" Nosedive (TV Episode 2016) ⭐ 8.3 | Crime, Drama, Mystery

1h 3m | TV-MA

^IMDb

And by "citizens," he means us, never hin.

"Larry Ellison, the world's second richest man and CEO of Oracle, one of the main partners of Stargate, said:

“Citizens will be on their best behavior, because we're constantly recording and reporting everything that's going on [...]. We're using A.I. to monitor the video.”

#capitalism

What was the name of the surveillance project he offered to run for the US government after 911?

Didn't know Oracle started as a CIA project

"His first project included a database for the CIA, code-named "Oracle""

en.wikipedia.org/wiki/Larry_El…

American internet entrepreneur and businessman

^{Contributors to Wikimedia projects (Wikimedia Foundation, Inc.)}

@ben freedom, is the right to act on whatever kind of fucking behavior you like so long as it’s not infringing on some else’s right to do the same.

Fuck “best behavior”.

I know this is scary, but I think it's important to point out that it won't work. This is a combination of grift and bluff.

There's an exponential dropoff in the effectiveness and accuracy of models as they grow. There's also an increase in attack surface. In terms of signal intelligence, it will be insanely leaky.

So this thing will lie and hallucinate, make it easy for people to tell they are being investigated, and be trivial to evade. It will also burn insane amounts of electricity and money. It will never match expectations, but that's unlikely to stop investment.

The more work they do to avoid evasion, the more vulnerable it will be. They won't be able to make it more accurate because they're going to run in to the fundamental limitations of the technology.

Let me get in to some concrete examples to make it clear what I mean.

Let's say some people are taking, say, on Mastodon, about something. First, the system has to scrape. It's already leaking information about itself just by doing this. We can know how often it runs and may even be able to fingerprint the technology stack. We can use this information to poison the data set.

Ok, let's say they're just getting stuff from phones via company installed backdoors. How do you know how someone is communicating? What platforms are they using? Simply using signal increases the complexity of getting at this data, even on a backdoored phone, a bit. This is more true at scale than for individually targeted attacks, but I'm not going to get in to the details there.

Ok, so let's say this thing is going through every text you sends, every message that goes through a FAANG owned messaging app. You send a link to someone. There are three options for how to handle this link:

1. Ignore the link. The model loses information.
2. Follow the link immediately. The model is vulnerable to a DoS. Basically a bunch of people send links to poisoned data to each other, the AI ingests this and poisons their data set. You also leak data about how often you scrape.
3. Follow the link only on investigation. The model leaks information about who is being investigated and is also vulnerable to a DoS. So, if you want to know if you're being investigated you occasionally send links to resources you control (that's not easily found any other way, like a url with a UUID that's not indexed) to people but ask via external channels not to follow those links. If the link gets hit, your communication channel is compromised. Now you poison the data set anyway.

So URLs let you circumvent it or attack it. But there are a bunch of other ways too. The more a model is focused on English language, the worse it will do on other languages. This is especially true of rare languages, and even more true of niche cultural slang. So even if this thing reads every single text you and everyone else writes, it may have no ability to know that you're planning to "🦆 thr d474 alle ∆." That's just garbage to an AI, but a human can probably figure it out.

But even in English, you can just send your text inside a file. Now they have another choice:
1. Ignore the file and lose data.
2. Parse the file.

Which do you think management is going to tell them to do? So they parse the file. Now they're vulnerable to trivial DoS and almost definitely vulnerable to RCE (remote code execution, or "getting owned the fuck sideways" in simple terms). How are they going to get around compression bombs? Fingerprinting? Sure, but there are infinitely many of these.

That brings us to steganography (stego). Stego is the forgotten cousin of cryptography (crypto). Crypto (in this context) is about keeping a message confidential by making it impossible to read by anyone but the intended recipient, even with the text. Stego is about hiding the existence of the message in the first place. It has largely been ignored because cryptography works pretty well and can be easy to detect.

But this equation changes when you're dealing with mass surveillance like this. Trivial stego that would be easy for a human to crack becomes a huge problem at scale. A simple example is just putting text in image metadata. Ok, so the surveillance folks write a parser for that. Great, now hide data in the low order bits of an image. Write a parser for that. Now just put it in the text, ok so OCR it. Ok, create a doc file that has an XML bomb in it so that it crashes but put the raw text you want to relate inside... There are infinitely many ways to evade, many of them are trivial to detect and parse, but that means you're essentially writing infinite parsers. There are also *much harder* problems to solve, but I'm not even getting in to that.

Yeah, basically, this shit is *way* harder than they think it will be and someone still has to build it. Now they have to decide on hiring for competence or loyalty. Good luck with that shit. Maybe some comrades should go work at Oracle for a bit to make sure that doesn't even happen in the first place. ;)

"Best behaviour"

😂🤣😂🤣🤣🤣🤣😂😂😂😂😂🤣🤣🤣🤣🤣🤣😂😂😂😂😂😂😂🤣😂🤣😂😂😂🤣😂🤣😂😂😂😂😂🤣🤣😂😂😂😂🤣🤣🤣

And breathe

🤣😂🤣🤣🤣🤣🤣🤣😂😂😂😂😂🤣🤣🤣🤣😂😂😂🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣🤣😂😂😂😂🤣

Imagine your tax money being used for this shit.

Not only terrible for privacy, but also for the environment.

Sure, they can probably get a good discount from the Chinese on their social scoring Big Brother system.

Then they don't have to work out all the early bugs again.

1984, period.

The U.S. are FUBAR.