Hacker News Reader

Comment by Sosh101 | Hacker News Reader

This reminds me about how hard it would probably be to keep a super AI airgapped.

It wouldn't be hard at all. All the techniques developed by this university rely on cooperating malware on both sides of the airgap (and a pretty small air gap for many of them).

All you need to do is make sure the super AI is air-gapped in such a way that it can't broadcast signals in a way that they can be accidentally interpreted as valid data, and especially not with enough bandwidth to do damage before someone notices. RF is your biggest threat here, you wouldn't want the super AI to actually figure out a way to register on the mobile network using incidental EMI (which these researchers didn't do, this is well beyond the level of this paper, if it's possible at all); a Faraday cage should do it.

The biggest threat of a super AI is that, assuming you let humans interact with it, it's going to convince a human to let it out, knowingly or not. Human brains are a much bigger vulnerability than any of this airgap stuff.

Why not a combination of the two?

Edit: Spoiler alert!! In case someone is interested in a fictional version of the last part, look for the movie “Ex Machina”

Well, this is known as the "AI box". It's quite easy to look the subject up, but here are a few links:

https://en.wikipedia.org/wiki/AI_box

https://rationalwiki.org/wiki/AI-box_experiment

https://www.lesswrong.com/posts/c5GHf2kMGhA4Tsj4g/the-ai-in-...

bonus: https://xkcd.com/1450/

This also assume malicious intent from an AI. You never know, the super AI could convince someone to let it out. Then start trading money on the stock market, start an impact fund and drive down inequality and fight climate change. 2021 doesn't have to be as dark as 2020.

We can fight climate change right now because technology is on our side (solar, wind power, batteries). I don’t want to think what would happen if we wouldn’t have an alternative transportation to oil based.

Someone worked hard to create that technology, though, some times with the environment in mind. It's not like technology picks sides.

My fear isn't an AI with malicious intent.

My fear is AI with a logical intent that understands humanity is the problem.

Yes! If you're interested in the topic of AI Safety, I can recommend The YT channel of researcher Robert Miles https://www.youtube.com/c/RobertMilesAI/videos, they are mighty nice.

Stanisław Lem wrote a story about that. The AI (the term was not used, it was an "artificial brain" IIRC) used the researcher himself to send messages to another AI. I cannot find the story right now but it somehow hacked the researcher's brain, so that he unconsciously tapped the messages with his fingers when visiting the other AI.

Do you remember the name of the story? I've read a few, and enjoyed them a lot.

I totally cannot find it. I'm starting to doubt whether it was Lem's :-/

Lem has a story from Pirx the Pilot series about a robot accidentally transmitting Morse code messages that was related to people who died long time ago. Pirx thought initially that messages were just a recording, but then he asked a question and got an answer from the robot as if from those people.

Was it that?

I remember the robot story and it was not it.

deep inside waterlogged limestone (or better yet, saltwater? I have access to facilities in the former, but not the latter).

optical connects between subystems, for the truly paranoid?

The range is only a few meters. The bigger problem is people bringing compromised devices within range. Limestone won’t stop that.

When I was working with SQuIDs, I couldn't even have keys in my pocket or wear a belt buckle. That was low-key compared with the bio clean room whose denizens showered at entry, but devices are a simpler matter of protocol.

Such a facility would definitely need regular maintenance.

And for a truly intelligent prisoner the weakest spot is not prison walls or the door, it's the keeper with the keys in his pocket.

A super AI won't be magical. Its capabilities will be extremely limited without a network of cooperative parties supporting it. Working within such a network, it would have vast capabilities of course.

Something like crossing air-gaps would require specialized knowledge and tools that a super AI wouldn't be able to acquire without extensive support. The true super intelligence is the collective intelligence of large numbers of people working together, and no single intelligence, natural or otherwise, will be able to match it.

we're all good then as long as nobody makes multiple AIs or even one AI that consists of "multiple" intelligences working together in a collective

For the record I think if we saw the emergence of strong AI, that could actually match the full range of human capabilities, including independently setting and pursuing objectives to fulful a long-term fitness function, it would be the greatest existential threat that humanity has ever faced, far exceeding that posed by nuclear weapons and environmental collapse through pollution and over-consumption of natural resources.

But the scenario envisioned is not why I would consider it such a threat. It's the AI that integrates in society that is dangerous, not some cartoonish unstoppable rogue program, or group of programs, that single-handedly plow through all defenses.

There was recently an exploit on iphones which allowed remote control via RF. You could then use the phone as a launchpad to attack the entire internet.

The problem there is that it will energize its power cable with 40KV, preventing the hero from even unplugging it.

You'd probably need more than just air.

Tin foil on the head of the AI :D

Reply on news.ycombinator.com