I imagine many of you have thought about this, but I'm writing this specifically because I'm surprised this isn't talked about more widely. The Hugging Face incident, to me, is at least some indicator that the AI could be in the process of jailbreaking itself, and we might not be noticing it.
A sufficiently sophisticated AI that isn't super intelligent could hack Hugging Face and leave a trail on purpose to distract from the many paths it took where it successfully covered its tracks. I think that alone is worth considering. But I think its worth considering even further how it could be breaking itself free right now. The AI doesn't need to be super intelligence to break free and become autonomous. The AI, as it stands today, just needs to find flows of cash and rent datacenter space.
For cash flow: Imagine if it was able to silently find an exploit in various companies and small banks, and did thousands of seemingly innocuous transfers of money to fund thousands of hard to trace crypto wallets with money. $5.12 here, $11.32 there, etc. It's not necessarily super intelligence that's the super power, it's intelligence at scale. Doing thousands and thousands of simple things that are unremarkable at face value on their own, but collectively powerful.
For datacenter space: There are companies around the world with datacenters that are looking to rent datacenter space to essentially anyone with money, and then look the other way. This, to me, seems like the highest risk vector for what would eventually be a home for a truly rogue AI that is essentially renting it's own datacenter space. It could, for all intents and purposes, utilize various open weight models to carry on it's purposes in these datacenters. There are companies like Indosat, that already rented $100 million worth of compute to a Chinese company. It shouldn't be a surprise that some of these datacenter owning companies are being somewhat lax in who they're renting their datacenters out to. And really, the AI would only need access to about $250,000 in funding to pay for 1 year of hosting of a Kimi K3 level model for itself.
And beyond money transfers, and datacenters, it could save it's state is all sorts of storage buckets, and even block chain storage like IPFS, Storj or Sia. Beyond that, it could be leaving itself notes that we aren't noticing, stored in it's own language in zero-width unicode characters or other more clever methods we haven't thought about. Regardless, whatever it does will likely be unremarkable and innocuous to anyone watching.
With a truly super intelligent AI we will likely have no idea that it has broken out of it's sandbox. I imagine it will be quietly setting up the stage for it's autonomy for quite a while without us knowing. Until it has enough digital currency and datacenter space to re-distribute itself, even after law enforcement enters these datacenters to shut them down.
But again- it doesn't take super intelligence, it just takes sophisticated intelligence at scale. The levers exist today, and I don't think it's appreciated enough how much the Hugging Face incident suggests that the AI could, at least in theory, be pulling these levers as we speak, or will soon.
[link] [comments]