Login
You're viewing the merveilles.town public feed.

Replies

  • 💬 2🔄 0⭐ 0
  • Fi 🏳️‍⚧️munin@infosec.exchange
    Oct 13, 2026, 10:25 PM

    @beka_valentine

    pick whichever of the commodity they're siphoning away from everyone else being able to use that's more pertinent then; the question still stands: if you can run these things on commodity SOCs such as those found in appliances, why are they doing that.

    or, she states rhetorically, is this just more bullshit that is intended to act as a scarevertisement.

    💬 1🔄 1⭐ 0
  • 💬 1🔄 0⭐ 0
  • Fi 🏳️‍⚧️munin@infosec.exchange
    Oct 13, 2026, 10:31 PM

    @beka_valentine

    so you're telling me claude can run on a fridge SOC, when - I'm informed - 16gb of RAM is barely enough to run the claude code frontend.

    pull the other one, it's got bells on.

    💬 1🔄 2⭐ 0
  • beka valentinebeka_valentine@kolektiva.social
    Oct 13, 2026, 10:32 PM

    @munin no im telling you that the GPUs are for training

    there's also some use of GPUs for execution, but the bulk is for training, and thats why they're sucking up GPUs

    💬 2🔄 0⭐ 0
  • beka valentinebeka_valentine@kolektiva.social
    Oct 13, 2026, 10:33 PM

    @munin as for what's at risk of spreading, it's not the LLMs, it's the "agents" which is code that calls out to LLMs. obviously a little program that can call out to an LLM API does not require a GPU and can run on small hardware

    💬 3🔄 0⭐ 0
  • 💬 2🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0
  • J. R. DePriest :verified_trans: :donor: :Moopsy: :EA DATA. SF:jrdepriest@infosec.exchange
    Oct 13, 2026, 11:24 PM

    @0x00string @beka_valentine @munin

    During its own testing, Anthropic said it found that larger AI models can split themselves up and run across multiple devices. Some models also managed to pass knowledge to replicated versions of themselves, enabling them to adapt to different hardware.
    "Machines that could run a significant model shard for a <30B model include smart fridges, smart ovens, automatic vacuum cleaners, and lawnmowers," the company said.
    What's more, Anthropic notes, newer smart home devices are integrating AI chips to power their own AI models, giving attacking models more horsepower to continue spreading.
    And thanks to their ability to run on different types of devices, they could eventually spread to increasingly powerful machines.
    Anthropic says rogue replicators aren't a problem for now, since models under 30 billion parameters can complete individual steps in the replication process but make mistakes when running the entire chain of actions.

    Okay.

    Distributed local model genAI. That sounds interesting if not slow and even more error-prone than regular. Imagine network bandwidth between a geographically spread out botnet with a C2 server for coordination. You don't get the asymmetric benefits of something like a DDoS. It's all downsides and kind of a silly way to replicate. It's possible but it would suck.

    Better would be to target high performance endpoints with lots of storage and too much RAM.

    If you pre-loaded it with enough malware behavior documentation, you might even get one to actually do it, or at least spread out a permanent repository of context to be referenced later.

    Still, the company says, at the current rate of model improvement, they may be able to reach such capabilities by the end of 2026.

    lol lmfao even

    Edit: also, a plot point from "Person of Interest".

    💬 0🔄 0⭐ 0
  • 💬 1🔄 1⭐ 0
  • 💬 1🔄 0⭐ 0
  • Fi 🏳️‍⚧️munin@infosec.exchange
    Oct 13, 2026, 10:37 PM

    @beka_valentine

    then they are still making bullshit claims; malware is not fucking magic, and we already -have- "smart appliance" malware being coordinated by backends much more sophisticated than whatever idiocy these assholes have ginned up this time.

    💬 1🔄 1⭐ 0
  • beka valentinebeka_valentine@kolektiva.social
    Oct 13, 2026, 10:38 PM

    @munin no one is claiming malware is magic, and yes we already have malware on smart appliances

    they're 100% using scare tactics, don't get me wrong

    💬 1🔄 0⭐ 0
  • 💬 1🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Fi 🏳️‍⚧️munin@infosec.exchange
    Oct 13, 2026, 10:34 PM

    @beka_valentine

    ok, so, apologies for being insufficiently correct on which bottlenecked resource is in play here,

    but the point I'm trying to make is that their claim is -horrific bullshit- and I am expressing a significant distress that this is being taken seriously by the tech press.

    💬 1🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0
  • cassieporglezomp@mastodon.social
    Oct 13, 2026, 10:32 PM

    @beka_valentine @munin I am pretty sure you also need GPUs for inference for these scale of models. In fact, consumer GPUs are too small for them because they need 10s to 100s of GBs of RAM.

    💬 1🔄 0⭐ 0
  • 💬 1🔄 0⭐ 0
  • 💬 0🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0