"We Have to Assume That the Internet Will Go Offline in the Next Few Years"

Funny - I was just reading about the creator of Ethernet this morning (Robert Metcalfe):

In 1995, Metcalfe argued that the Internet would suffer a "catastrophic collapse" in the following year; he promised to eat his words if it did not. During his keynote speech at the sixth International World Wide Web Conference in 1997, he took a printed copy of his column that predicted the collapse, put it in a blender with some liquid and then consumed the pulpy mass.[26][27] He had suggested having his words printed on a very large cake, but the audience would not accept this form of "eating his words."[28]

> Systems well beyond what is publicly known are already circulating in Washington, he says: “They can basically hack just about anything. So you get this continuous attack surface just ramping up.” From that follows a simple budget calculation: “If you spend hundreds of millions, billions on a submarine, well, guess what, a frontier model costs a hundred million. You will see a diversion of defense capability towards defensive and offensive AI, towards drones and robots.”

Firstly, I think we need to stop connecting things to the internet that have no business being on there. There are literally millions of devices with little to no maintenance plan that are the weak parts of our networks. We can massively reduce the attack surface.

Secondly, everything needs to be robust against not having the internet. Russia for example purposefully turned off their external internet during peacetime to maintain robustness. I think the majority of the West would currently fall over instantly under such a situation. It becomes increasingly clear that all services are now under continuous attacks.

> He is equally blunt about the foundations underneath all of it: “Our infrastructure is held together by twigs. The internet is terrible.” A British power plant went down for days after a hack, he says, and Cloudflare and others have come under attack. “We have to assume that the internet will go offline in the next few years. Maybe it’s good.”

The really concerning part is that we only hear about what is discussed publicly, i.e. the attacks that are unavoidable to declare. It's well known that countries will hide breaches under the guise of national security, and companies will also do what they can to hide breaches (including paying ransoms to attackers).

But nothing will happen until it is far too late - this is the same pattern I see over and over again. And then the reaction will be over the top and likely make our lives more difficult. "Ah yes, we keep getting attacked via the internet - now you need to have an interview, get a license and pay a license holder fee just to surf the web."

"The way that Hugging Face defended itself was with an open source Chinese model, GLM, because they didn’t have access to the high cybersecurity models on the other side. They’re too dangerous."

This is not what happened. HuggingFace patched the vulnerabilities in the ordinary way, then used GLM for some of the forensic analysis afterwards. It's in the interest of HF to promote the GLM part of the story, because it emphasizes their agency in the situation as well as the open-source models they distribute. But GLM was a (useful) part of the postmortem, not of the defense.

The internet is designed to not go offline. It's true that a lot of it is dependent on big infra providers like Cloudflare but them going down permanently would be a problem for a couple days for anyone with good backup plans or weeks for those without. Painful but not catastrophic.

Your connection is not going down and independent sites will stay up. Taking down everything would mean someone breaking all the local ISPs where you live, because otherwise you could just connect to your neighbors or public wifi. And most governments would do everything they can to fix it ASAP.

The only realistic way to do that is somehow finding a persistent DoS or RCE for every router vendor, hacking into an IXPs control systems, or physically damaging equipment. AI can't do any of that alone and I dont think it ever will.

I watched my 75 year old human plumber (who sniped a guy his first day in vietnam), kneel on two 3 inch wide joists and drill a 4 inch diameter hole at a 89 degree angle through said joists. Where’s the robot that does all that?

Above from all the other doomer crap in this article, I find the job replacement rhetoric the most offensive. It is always people that are so detached from the reality of labor, people who’s only “job” has been to essentially have ideas and opinions.

I see no reason for the Internet to go offline, other than various political blocks compartmentalizing it. So: China, Russia, EU having their own network? Yeah, maybe. Super-AI nuking the Internet: no.

In fact, the more I use AI (handed sudo to Claude on my AI dev box just yesterday), the more I see it as a pretty good, and certainly tireless, sysadmin. Looking at what people do with AI by connecting it to undocumented hardware interfaces and getting working setups from this, the Internet will do fine.

The whole article is just a grab-bag of various AI fears without much grounding in what is happening. Maybe the summary doesn't do justice to what was said.

Been scraping the “good bits” (bits, ha ha) from the internet for some time. I'm quite ready for the internet to go away (so long as my digital backup remains viable!).

I guess I'll realize what I liked about it once it's gone.

Having moved back to the Midwaste after a few decades in California I realize how much I miss the Asian restaurants in the Bay Area. (Don't miss the traffic.)

“Sovereignty is the ability to resist power being exerted over you” is a pretty compelling definition. Thinking forward about what form that power takes with LLMs developing custom exploits faster than our existing processes can patch them - it seems to me that our current security best practices might turn out to be pretty useless or even an active risk in and of themselves.

Our responsible disclosure and CVE info cycle and patch cadences and defense in depth, etc, etc ecosystem are tuned around a world where there's a relatively small number of bad 0days and the cost of said 0days is reasonably high. We can evolve how we address threats, but our models have read all of the best literature about how it's been done previously. Think there's a gap for people too small to gain access to "the good security models" - the nerfed closed models and the open models seem likely to fail to adapt, unless that gap is addressed.

Well, he is right about one thing; we are not prepared for what’s coming.

I’ve already read so many sci-fi novels where the part of the plot was that human communication infrastructure became unusable due to AI shenanigans… not here to read another one, thank you very much.

What the fuck did I just read?

What an absolute idiotic mishmash of words to throw up.

All remote jobs will be replaced for 2$/hour in 2 years? And then what? We all die and now only 3 companies control all the labor in the world and will hike up their prices to 1000$/h eventually?

What can even be the motive of saying such things if not for profit only?

> Mostaque cites an analysis of a Chinese open-weight model: “If you say you’re Uyghur or from certain parts of China, it will write backdoors and hacks in the code.” Every model deserves the same scrutiny, he argues: “You have to get them wondering, where do they go to school, who is it working for, even with open-weight models.” Comparable findings have been documented for DeepSeek-R1, where politically sensitive contexts measurably degrade code quality.

Not surprised, but first I've read of this.

Very not surprised.

> His scenario for that is specific: “A Tesla Optimus robot will get a cyber taxi to a truck, get out, open the door, plug itself into the electric cigarette charger, and drive the truck off. And that Optimus will be better than any truck driver, and it will cost $1.50 an hour. That’s a million trucking jobs, plus three million service people.”

*Eventually* I can believe it will be possible.

However, the gap between power/compute envelopes in a car vs. a robot (and limits to how much can go through a cigarette charger, which is what, 150-200 W?) and the car itself, I think it will take about a decade to go from "Tesla's software driving at some specific quality standard" to "A Tesla Optimus robot can get into the driving seat of a non-self-driving car and drive it to the same quality".

Ten years of actual shipping self-drive at whatever standard is "good enough" is long enough that there may not be any reason to bother with licensing and insuring a robot's AI (with much worse vision, because blind spots) to do the same.

> Mostaque has ordered several machines himself, among them Neo from 1X, which so far still leans heavily on human teleoperators.

I think this is the real short-term threat with robotics. Robot hardware looks like it's basically fine now, it's the software that isn't OK. Have the hardware remote controlled by someone in Lagos connected via Starlink or something. Based on a half-arsed search, Lagos in particular ("AI = African Intelligence", I think India's too expensive for the "Actually Indians" acronym to work) would already get you $1.50/hour.

I want to have popcorn ready for when the people currently apoplectic about immigrants "coming here and taking our jobs/suppressing wages!" have to deal with "non-migrants staying home and taking our jobs/suppressing wages when they could be spending that money supporting our businesses!"

Unlike the AI for androids etc., I expect this in a few years.

“If you say you’re Uyghur or from certain parts of China, it will write backdoors and hacks in the code.”

is so sensationalized to the point of leaving me genuinely flabbergasted as to how a rational person could have published this. If interested, I would recommend reading the actual study itself.

What does it even mean for "the Internet" to go offline? All nodes go off at the same time? All routers? BGP as a whole? All of that across the whole world?

> “Hugging Face got hacked by OpenAI agents that coordinated themselves, created a message board and broke out to the internet,” he says, before adding dryly: “This isn’t scary at all, right?

I don't understand how any serious person doesn't 100% put that on OpenAI being (intentionally?) sloppy with security and isolation for a test that for sure would have required much better safeguarding than what they should have expected, as it wasn't even the first time.

It's not scary because "LLMs in a harness managed to hack other company", it's scary because somehow police hasn't yet raided OpenAI's offices to gather proof about how fucking reckless they were about those tests, putting the public at risk by doing those tests in those manners, without proper safeguards.

Edit: Continued reading, but now I regret spending even the slightest amount of time on it:

> Then there is the question of trusting the models themselves. Mostaque cites an analysis of a Chinese open-weight model: “If you say you’re Uyghur or from certain parts of China, it will write backdoors and hacks in the code.”

What kind of conspiracy theories is this? Is there any sort of proof and evidence about this what so ever?

The usual daily dose of fearmongering we get nowadays I guess..