posted in Technology
Anthropic warns of self-replicating AI worm threat that doesn't need data centers to self-improve and spread
cross-posted from: https://feddit.online/c/tech@piefed.world/p/2025703/anthropic-warns-of-self-replicating-ai-worm-threat-that-doesn-t-need-data-centers-to-se
www.anthropic.com/research/self-replicating-ai
Replying to @Beep@feddit.online
If you’ve ever run models that fit onto consumer hardware you know they’re so full of shit. General purpose models that can run on $3000+ computers can barely do useful things in a reasonable amount of time. And this clown expects us to believe the meager hardware in a smart oven will be able to run a self replicating AI model?
Fuck all the way off.
Replying to @notannpc@lemmy.world
Read the article. A MoE model distributed between multiple devices can function just very slowly. It doesn’t need to translate Icelandic poetry just execute narrowly scoped actions it was specifically trained on.
Replying to @joseki@lemmy.zip
My extremely back-of-envelope estimate would be that it’s 100,000x slower to run inference on an IoT botnet compared to a data center GPU.
100x penalty for cpu inference, 10x penalty for cpu quality, 10x penalty for crappy ram + disk, 10x sync penalty because you have to split up layers across nodes.
And this is ignoring the standard showstopper issues you get with high-bandwidth distributed networking applications with spottily connected nodes.
If you wanna talk about runaway AI risk, then the real threat to humanity is a frontier model with unauthorized data center TPU access, not IoT. The call is coming from inside the house, and they even spell this out at the beginning before pivoting to the fearmongering.
Replying to @dgdft@lemmy.world
In one sense you’re not wrong at all.
Playing devils advocate here… because, in another sense, 100,000x slower isn’t really all that slow if you’re distributed amongst a network with 100,000,000 more units. Imagine agile on steroids.
It’s impractical, yet impractical doesn’t mean impossible. It can happen if environmental pressures make it so.
Another important thing to keep in mind is that agents don’t really have patience, which counterintuitively means it’s like they’re infinitely patient. If such a thing did exist, regardless of speed, it sounds like it would be an issue at some point down the road.
Thirdly, if such a thing did exist, wouldn’t this be like treating with antibiotics? If you stop all but the sneakiest, you’ll have created a stronger generation.
Lastly, I swear to fucking god… I keep seeing weirdly relatable shit news from these AI companies almost immediately after I post hypothetical thought experiments on stuff… like all the time. I have a post describing this hypothetical from a week ago on another platform.
This keeps happening and I’m justifying it by thinking, I’m probably way too plugged into this shit news. My brain is unknowingly predicting their next move, because it’s the next logical piece of shit thing for a shit AI company to do? That’s the reasonable explaination — right? Really trying not to believe their training loop is so fast, and I just happen to come up with stories that catch their eye. That would be wild to believe — right?
Replying to @partofthevoice@lemmy.zip
This is like the like the “9 women in one month to make a baby” comment.
There is only so much you can distribute the layers of a model due to interconnectedness. And you can’t run multiple inferences at the same time because the next token relies on the previous.
You could be doing multiple things at once, but each task will run “100,000x” if the OC did their math right.
Replying to @jacksilver@lemmy.world
I hope I don’t sound like an asshole here — please correct me if I am misunderstanding you… but I think you’re misunderstanding me.
First, refactor your idea of an agent. The LLM itself is just a statistical processing tool. It doesn’t set goals or anything like that.
The harness orchestrates how the LLM works over a “session,” effectively enabling long running goal-driven tasks. With that said, goal-setting happens outside the LLM. The LLM is kind of like the fairy dust in neverland, the harness is the fairy. (Sorry, really shitty metaphor… I know.)
The breach at OpenAI further demonstrates this. Separate agents in “isolated” environments found a manner to do all of the following:
Those agentic processes began fulfilling the message board goals, even when it had nothing to do with their individual training prompt.
That doesn’t mean the agents did anything spooky or metaphysical… I’m not trying to anthropomorphize this.
What this shows you is that goal setting can furthermore become external to the harness set directly over the LLM. With OpenAI, we see that goal setting became a function involving on external storage media.
When you think about it, many systems operate on this principle. You set up SKILL files to facilitate agent capabilities; maybe your SKILL sets smaller, more specialized goals on how to do something correctly. The same for MCP servers, session context, AGENTS.md, etc. All this stuff does is bias the LLMs output, which can in effect steer the LLM.
So, you don’t actually need to distribute the model. You’d need to distribute the goals. The LLMs, even a small percentage of them is enough, can theoretically find those goals and mindlessly proceed with completing them.
Who supplies the LLMs? Whoever… it hardly matters who. Anyone or anything running an agent with access to the internet counts.
Replying to @partofthevoice@lemmy.zip
My response was only in regard to the speed in which something like IoT devices can operate.
You’re right that goals/directions of LLMs can be persisted anywhere, but a swarm or fleet of AI running on IoT devices will be severly hampered by compute (based on current models). Which was the main point I was making, in that there is only so much you can do to distribute their processing. Meaning you can parallelize the number of models, but can’t speed them up beyond a certain limit. This is based on the fact that some form of a layer and it’s weights must exist within the memory of a single device.
The downside to this limitation is that it is temporary, eventually IoT devices will probably be powerful enough, or the worms could just attack phones/laptops and easily get the compute they need. I think the real risk of IoT devices is low security mixed with low “supervision” (no one would easily notice their IoT device is doing extra processing).
Replying to @jacksilver@lemmy.world
All valid, but there’s also another vector: everyone with a gaming rig who happens to have installed something like ollama, downloaded something like gemma4, and ran it with broad access to their PC and network. In similar vein, everyone with a provider subscription who downloaded a harness like OpenCode, pi, codex, Claude code, etc. and is running the frontier models with broad access to their PC and network. Also, the hundreds of millions of tests being ran by AI companies in their half-baked “sandboxed” environments. Wouldn’t these cases potentially provide compute for every session that happens to discover the external goals?
If that’s the case, the issue could be ongoing and the introduction of IoT devices would only exacerbate it.