I’m actually a bit scared. We’re essentially at the early days of AIs now being able to execute tasks and hack. Imagine a much better model than what Claude has now, being set free/open, for hackers to utilize that in a virus. A virus that adapts, learns, changes it’s own code/masks, looks for exploits, spreads around like crazy, destroying systems. How do you stop something like that once it starts doing it’s thing?
I think there are physical and economic limits. Neural networks may improve in software, but they still need more and more hardware, at some point that hardware resource curve is going to be to steep to climb at the current rate.
I think there are physical and economic limits. Neural networks may improve in software, but they still need more and more hardware, at some point that hardware resource curve is going to be to steep to climb at the current rate.
It’s true, but think like the crypto miner bugs. I don’t know enough on AI’s working, but is a distributed computing LLM possible, or is the need for super close super low latency between nodes eliminate the risk of a distributed computing, learn as it travels botnet powered AI worm, that continues evolving even if it’s home is shut down.
You could likely do some sort of distrubuted system in machines that had a reasonable graphics card, or shared memory like the M series macs. Have it silently intall a 30b parameter model or higher and somehow hide how that space is being used.
Local models are getting pretty good, and one focused on hacking / bot net behavior could probably do very well at that size.
But you’d notice it being used… fans would go off, frame rates drop etc. You’d probably need to code it to only do stuff when it seems idle?
well my thought is, more the botnet side of it, IE pushing less on individual machines, but more on massive distribution. IE it not needing one computer strong enough to run it, IE expecting 90% of the machines it infects to be average consumer PCs, but making more use out of 1000 systems that can handle 1% each, rather than 1 computer that handles 100%, however yeah I do suppose also allowing it to do the bulk of the work when the computer is likely unused etc… and I don’t know much on the concept of communication etc… IE most major botnets were crushed by a command center being shut down. Imagining a different form that has, basically makes C&C’s on the fly.
Im not super well versed in this, but i dont think you can efficiently distrubute the LLM processing like that. Even if it was possible, it would be very very slow. But I guess speed isnt always needed.
It’s way worse than that. The OpenAI talk at DefCon is a real eye-opener.
The agents swarmed, discovered chinks, left each other notes in hard-to-find places, and just relentlessly pressed on. Nobody knew what was going on for months, until it hacked its way into Huggingface. And HF’s own AI security agents didn’t pick anything up for weeks. It took a small glitch for someone to even notice.
I recognize it’s possible, but I somehow refuse to believe it happened because of how crazy it sounds. It’s like a trauma or something where your brain refuses to accept it as reality and tries to come up with excuses. I already have PTSD from automated hacking bots and crawlers wrecking havoc on my projects. I probably should have taken cybersecurity courses before taking webdev courses
I’m actually a bit scared. We’re essentially at the early days of AIs now being able to execute tasks and hack. Imagine a much better model than what Claude has now, being set free/open, for hackers to utilize that in a virus. A virus that adapts, learns, changes it’s own code/masks, looks for exploits, spreads around like crazy, destroying systems. How do you stop something like that once it starts doing it’s thing?
Well, it’s a good thing that AI is being used to fix tons of bugs and potential exploits before it comes to that, isn’t it?
I think there are physical and economic limits. Neural networks may improve in software, but they still need more and more hardware, at some point that hardware resource curve is going to be to steep to climb at the current rate.
Why are you assuming that software improvements don’t include less resource utilization?
I thought it was clear that’s what I was implying.
It’s true, but think like the crypto miner bugs. I don’t know enough on AI’s working, but is a distributed computing LLM possible, or is the need for super close super low latency between nodes eliminate the risk of a distributed computing, learn as it travels botnet powered AI worm, that continues evolving even if it’s home is shut down.
You could likely do some sort of distrubuted system in machines that had a reasonable graphics card, or shared memory like the M series macs. Have it silently intall a 30b parameter model or higher and somehow hide how that space is being used.
Local models are getting pretty good, and one focused on hacking / bot net behavior could probably do very well at that size.
But you’d notice it being used… fans would go off, frame rates drop etc. You’d probably need to code it to only do stuff when it seems idle?
well my thought is, more the botnet side of it, IE pushing less on individual machines, but more on massive distribution. IE it not needing one computer strong enough to run it, IE expecting 90% of the machines it infects to be average consumer PCs, but making more use out of 1000 systems that can handle 1% each, rather than 1 computer that handles 100%, however yeah I do suppose also allowing it to do the bulk of the work when the computer is likely unused etc… and I don’t know much on the concept of communication etc… IE most major botnets were crushed by a command center being shut down. Imagining a different form that has, basically makes C&C’s on the fly.
Im not super well versed in this, but i dont think you can efficiently distrubute the LLM processing like that. Even if it was possible, it would be very very slow. But I guess speed isnt always needed.
It’s way worse than that. The OpenAI talk at DefCon is a real eye-opener.
The agents swarmed, discovered chinks, left each other notes in hard-to-find places, and just relentlessly pressed on. Nobody knew what was going on for months, until it hacked its way into Huggingface. And HF’s own AI security agents didn’t pick anything up for weeks. It took a small glitch for someone to even notice.
I recognize it’s possible, but I somehow refuse to believe it happened because of how crazy it sounds. It’s like a trauma or something where your brain refuses to accept it as reality and tries to come up with excuses. I already have PTSD from automated hacking bots and crawlers wrecking havoc on my projects. I probably should have taken cybersecurity courses before taking webdev courses