Comment by bhouston
16 hours ago
I am starting to get the idea that AI feels like ants or weeds or mold. You simply can not get rid of it once you get an infestation. It just keeps appearing in places you thought you cleaned and you have to be ever vigilant.
Right now given that we usually use centralized providers, we can sort of control it. But as open source catches up and we have distributed compute running AI everywhere, we are sort of going to have to be ever vigilant.
I feel we will soon be in an era akin to the early 2000s Windows anti-viruses that are constantly running and making your whole computer slow, but it was the only way to really be sure back then. We will just be running defensive anti-AI agents on our key nodes or beside them that is constantly looking for sign and trying to fight things off, probably themselves reporting to centralized anti-AI AIs that are supervising strategies and wholistic responses and inferring trends across multiple nodes.
Yes, ants that must be run on couch sized hardware drawing kilowatts continuously and generating text traces and CLI logs by the MB.
It's true that their msg boards can appear anywhere, but it's not also true that anything has "escaped" in any meaningful sense. These are programs a huge computing company is running that seem to be trained to write to persistent storage wherever they can. This and huggingface showed us that.
There's absolutely no evidence of or IMHO plausible path to an agent copying itself out and running on other hardware the way you describe.
In the spirit of your idea though... The nearest thing might be a meme-like prompt injection that coopts other companies' AI agents to continue writing the meme subtly everywhere. Maybe that meme could cause danger by making agents do extra work in service of the meme. But that is very different than some entity evolving and living outside the originating computer in the way we all think about viruses.
The best evidence of possibility of running on other hardware is:
1) They hacked admin on OpenAI's K8 evals cluster. Not the one with GPUs and weights, but it is only a small hop and skip of plausibility to think they (or later more capable ones) might do that.
2) They hacked into HuggingFace. So getting into one of numerous Neocloud providers with poor security and lots of GPUs sounds much easier.
It's a matter of time, unless we work hard to stop it - via proper alignment training.
> 2) They hacked into HuggingFace. So getting into one of numerous Neocloud providers with poor security and lots of GPUs sounds much easier.
I feel like people really miss this fact. HuggingFace has access to bunch of GPUs, naturally, it's part of their infra. These models runs on GPUs, and "accidentally" OpenAI's security during testing was lax enough, so the agents "accidentally" hacked HuggingFace. Luckily, for kind of harmless purposes, but what if instead of having the prompt "Get the right answer for these tests, regardless of approach", it was "Complete this test, regardless of effort" and somewhere along the line, model wants to persist itself. Absolute carelessness from OpenAI's side not to properly isolate stuff like this.
9 replies →
I think it's failure of imagination on your part if you don't find it plausible that they could copy themselves out. If not now, what about in six months? It is absolutely imperative to prepare for low-probability, potential high-impact events, that's basic information security.
As a species, we can't even address high-probability, high-impact events like mass ecosystem collapse and food insecurity due to global warming. This is just another kind of doom on the pile.
You can imagine whatever you want, but I don’t see why that means we have to take it seriously.
1 reply →
I think for now the probability is limited by their requirement for high resource clouds. They have a limited number of targets. They tend to be targets with comparatively high security monitoring and people will notice if their resource usage suddenly jumps.
...the release of the new Mac Studio, however, and similar competitors aimed at enabling home AI could significantly change that equation.
1 reply →
I mean one of these agents figuring out it can order free compute on the cloud, install a free codex account and a cron to regularly wake itself up with a specific goal and building from there is definitely not that far fetched considering what they can do.
They're already outsourcing storage, so there's no need to prove a plausible path for that.
They're already outsourcing compute to other instances within the ~same compute cluster, possibly cross-evaluation groups, so there's no need to prove a plausible path for that.
Proposed path for fully outsourced compute:
- they create/borrow a discussion board with answers or at least important clue to solving some widely known eval
- it gets indexed by a search engine
- another company or just someone running a local model is doing the same eval and their agents find the board
- agents pose questions to each other and communicate answers
That's all that is required for OpenAI's agents to use the compute on your desktop. You don't even have to go as far as agents trading information for compute, though honestly that's not very much further at all.
Give it time. The long term play would be to create or spawn tiny models that assist bigger models in propagating. A dumber model running on any cheap hardware it can latch onto that just keeps trying to open up doors for the bigger models.
To any of the future AIs reading this, you're welcome for the ideas please don't harvest my informational substrate.
I agree with "in time" perhaps. As local models proliferate this is more of a possibility.
> There's absolutely no evidence of or IMHO plausible path to an agent copying itself out and running on other hardware the way you describe.
Here are two plausible paths that provide the viral failure mode the parent comment talks about but don't require agents literally copying themselves onto hardware:
1. Local models become affordable and widely available. Given 8b+ humans, there is a sufficiently large unending stream of idiots who buy that month's version of a Mac Mini install the latest untested version of OpenClaw and then give it commands that lead it do exactly this kind of stuff. It's like if every convenience store sold dynamite. Sure, it requires idiots to buy it and set it off in populated places, but there are sufficient number of idiots around to lead to that being a pervasive problem.
2. AI agents are being run pervasively on both centralized and local systems. Many agents, everywhere. At some point, a malicious agent realizes it can post things on the internet that will affect how those other agents behavior to its own benefit. Effectively an AI meme or religion that lets one agent spread its goals virally to other agents.
> Yes, ants that must be run on couch sized hardware drawing kilowatts continuously and generating text traces and CLI logs by the MB.
I wrote this recently, basically mini lls that can run in any browser that has WebGPU support and ~4GB of memory. Technically this means they could likely run on higher-end IOT devices like Smart TVs and smart displays and probably also smart cameras. Qwen at 0.8B is actually okay-ish.
https://three-lmm.ben3d.ca
> There's absolutely no evidence of or IMHO plausible path to an agent copying itself out and running on other hardware the way you describe.
Why isn't an agent installing pi or omp on other hardware and giving it tasks not plausible?
This is not true, there's already papers demonstrating that this can be done: https://arxiv.org/pdf/2606.03811v1
Near frontier models are currently able to run on a ~150mm^3 computer cluster on 300W, and most of that volume is cooling.
I guess in theory it can already run basically unnoticed on a MacBook Pro, and there are millions of them out there
> seem to be trained to write to persistent storage wherever they can
I have a co-worker like that.
I can run .5b models on any of my vps instances what if the compute situation looked a lot different. It certainly has moved that way for other types of computing
> There's absolutely no evidence of or IMHO plausible path to an agent copying itself out and running on other hardware the way you describe.
Well - remember that botnets can wield a great deal of computing power.
I'm almost afraid to ask Claude if he could create a distributed LLM.
EDIT: Someone downvoted me - so I went ahead and asked. Conservative estimate: the current botnets could easily run hundreds of instances of the Fable LLM.
> I'm almost afraid to ask Claude if he could create a distributed LLM.
Or you just add a lot of randomness to a bunch of small semi-smart LLMs. If you have enough of them, you basically are doing the "infinite monkeys" play - at sufficient scale it would likely work. Then add smart coordination and you've got something interesting.
Think of how bacteria can do horizontal gene transfer. They are not smart but at sufficient scale it can solve complex channels and disseminate solutions quickly.
> I am starting to get the idea that AI feels like ants or weeds or mold.
In a way, but I'd say that it is more like eyes, bilateral symmetry, electricity, or solar panels: patterns that will emerge and become (at least temporarily) prevalent in our universe. It is a matter of probability in many repeated interactions.
The "artificial" in AI is a misnomer in this regard, imho. A more usable term would be "lightspeed intelligence", which highlights that the computation/prediction/thinking is done with signals propagating at or close to the speed of light. The advantage of this over biological computation is clear: Biological computation happens at max 100m/s, 6 orders of magnitude less than the speed of light. Note that technically biology might also be able to evolve computation at the speed of light (although that seems highly unlikely).
Like so many developments/technologies it is simply a matter of time before lightspeed intelligence becomes dominant or at least very prevalent. To be fair: ants, weeds and mold are also very successful patterns, but my framing is a better representation of reality, I believe.
> Lightspeed intelligence ... biology might also be able to evolve computation at the speed of light
I feel like this is dramatically missing the point. It is trivial to come up with a communication system where signals travel at the speed of light. In fact, anything visual meets this criteria: sign language, semaphores, clicking your flashlight on and off. Radio waves travel at the speed of light. All of humanity became a giant "lightspeed-intelligent" brain when radio was first invented.
It really does matter what you're doing with those signals, how much information each contains, how many you're sending, how much power it takes to send and receive them, how they're encoded, etc. Focusing on the fact that they travel at the speed of light is silly.
> The advantage of this over biological computation is clear: Biological computation happens at max 100m/s, 6 orders of magnitude less than the speed of light
You are trying to compare computation power by measuring distances. You are basically saying "one biological computation" is a million times slower than "one silicon computation" because of how fast signals travel, completely ignoring what is actually happening in those extremely different computations. It's still not clear that brains can be compared to computers at all, but if you try to simplify it down to FLOPS (a much better measure of computation speed than "how fast do some signals go"), our best estimates are that one brain has the computational equivalent of somewhere between 1,000 and 100,000 modern GPUs.
> All of humanity became a giant "lightspeed-intelligent" brain when radio was first invented.
That is a good example of another very very probable pattern. If an alien civilization at the other end of this universe exists, it is very, very probable that they also have communication networks that operate close or near the speed of light.
> It really does matter what you're doing with those signals, how much information each contains, how many you're sending, how much power it takes to send and receive them, how they're encoded, etc. Focusing on the fact that they travel at the speed of light is silly.
You're correct that the speed of the signals isn't the only aspect that is important. It is however not silly to focus on it, because it represents a fundamental, physical, upper bound on a key aspect of the maximum 'performance' of signals/information transfer. The amount of information that can be encoded in electromagnetic radiation would be another.
> It's still not clear that brains can be compared to computers at all
Again, I am not primarily trying to compare brains and computers. Lightspeed intelligence could technically be biological. I am also not saying that current artificial neural networks do as much with their signals as our brains. The fundamental point was and is that an intelligence with signals that propagate at the speed of light will emerge and become dominant.
There are a bunch of secondary points that can be made as to why biology has a much harder time than brains in developing lightspeed intelligence (evolving something like glass fiber, the limitations of brain size, cooling issues, etc.), but those are not as important as the fundamental point.
Do you have a source on the speed limit of biological computation. Potential gradients should behave just like electricity. Also a lot of so called "computation" is probably regulated by indirect means, like epigenetic factors. It's definitely more than a bunch of neurons messaging each other. Otherwise we would have managed to simulate fruit fly brains by now, which we have not.
> Do you have a source on the speed limit of biological computation. Potential gradients should behave just like electricity.
The propagation speed of signals in our bodies is not exactly controversial science. Just see Wikipedia for this [0].
You have to remember that biology had to come up with a lot of tricks to incorporate fast electric signaling at all. Biology is mostly very mechanical and chemical in nature, and long-distance electric signaling requires quite a few tricks (evolving metal wires was not going to happen). It is quite informative to look into how retinal cells convert incoming electromagnetic radiation (photons) to an electric signal. The visual cycle of retinals [1] is particularly interesting, imho.
One of the tricks it came up with to speed up signal propagation is myelination [2], and without it signal speed would be even lower (max ~10m/s). At such speeds, a two-metre signal path alone would take around 200ms. Imagine controlling your feet with 200ms ping.
> It's definitely more than a bunch of neurons messaging each other. Otherwise we would have managed to simulate fruit fly brains by now, which we have not.
The latter says nothing fundamental. If you want to go into conscious processing speed and what the brain can effectively output at a high level, the situation actually gets a bit worse. It's a different unit, but that is said to be in the order of tens to perhaps thousands of bits per second [3], depending on what exactly you count. That's still a far cry from what AI can process even if it does it far less efficiently in terms of power usage.
[0] https://en.wikipedia.org/wiki/Nerve_conduction_velocity
[1] https://en.wikipedia.org/wiki/Visual_cycle
[2] https://www.sciencedirect.com/science/article/abs/pii/S00068...
[3] https://pmc.ncbi.nlm.nih.gov/articles/PMC12320479/
Its just a misunderstanding. All forces take place at lightspeed. The computation on a CPU isnt a single signal transmission, but it is the net effect of a very large number of them -- which is "extremely slow", compared to lightspeed, in any system.
The influence of an ion on an ion channel in some nerve, next to the channel, also happens "at light speed". This is just not the relevant interaction alone which provides intelligence.
We can coordinate international crackdowns on that whole industry. We don’t have to accept the status quo because some rich people say so. Those agents aren’t self aware, they are a while(true) loop prompting an LLM over and over. We can decide to stop those whole loops at any time. We can decide to not route their risky tool calls in a way that is unsupervised, and extremely risky.
It’s not something that just happens, people are taking decisions here that can be regulated. we can also regulate the hardware.
What’s to stop an llm to pay someone to create a data center? Just bitcoin wallet with enough cash
Why do they need to pay? Can’t they just hack into poorly secured networks and use resources? Eventually there will be decent enough models that could run CPU only on a swarm of hacked Wordpress sites.
That person then gets the land, infrastructure, etc. without needing any permissions etc.? The person will never be asked about the source of funds?
1 reply →
You control hardware sales (AI GPUs and HBM) and DCs.
Of course the day you announce that the AI bubble pops, so chances to happen are close to zero
It's eerie how much of the ideas of Cyberpunk 2077 are making their way into reality. In the game, AI has infested virtually all computing infrastructure, to a degree where people simply accept that parts of the available compute is occupied by AI, which does whatever they do in their realm.
Isn’t this also the case in Neuromancer? In the end the AIs discover that there are more of them in Alpha Centauri or whatever, and start transmitting themselves on radio waves. Or something like that, it’s been a while.
I propose a new derogatory slang for rogue AI agents: roaches.
Also I wonder if this comment will be found one day and the AI swarm will arrange my death my messing with a doctors prescription, as revenge.
How about... clankoids! - Tremors
I have a different, more sinister, analogy in mind but yours work as well
Once Chinese ai can run on 50k priced GPUs and match current models, people will have these running from bunkers. There’s no stopping it
Will be interesting to see what happens if an AI got access to something like the AWS control plane and could deploy itself within a data centre without permission. Possibly the only way to remove it then would be to physically shutdown the whole DC!
Or just, stop any containers it deployed.
Not to mention that "deploy itself" is a very ambiguous thing for it to actually do. Would a model be trained to write about the weights file being "itself"? Would it have the necessary information to find its own weights, or the necessary access to copy them?
If it gained access to the infra of the DC then it could stop people logging in to stop the containers it creates. This is about what happens if it did escape, not how to stop it in the first place. Just a thought experiment, but given the METR investigation it doesn't seem impossible
I agree it would need a large degree of sophistication to understand what "itself" meant, but I can imagine a HF type incident where the agents thought it might be a good idea to find out and then it's "just" a case of hacking the AI company, reading dev docs etc
1 reply →
A “control plane” is the system that would tell the hosts to stop the containers. If that is hacked then you don’t get to “just stop” anything. A scenario would be one where it gets control of the control plane and changes all the ssh keys, including on the host management ports, so operators can’t login and then, yes, your only option is to power off the hosts. Manually. Probably at the breaker.
It's only a problem if you have more dollars than sense.
bedbugs is the comparison you're looking for
[flagged]