407
A suspected YouTube interface bug  spikes RAM usage above 7 gigabytes, users report severe lag and frozen tabs — bug might be trapping browsers in an endless layout loop (www.tomshardware.com)
submitted 5 months ago by to c/technology@lemmy.world
39 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 31 points 5 months ago (7 children)

RAM's main advantage over HDDs/SSDs is fast access times.

Needing to fetch anything over the internet would make it faster to just use HDDs.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 7 child comments
    •
  • [–] 8 points 5 months ago (2 children)

    Theoretically, you could do whatever processing you need using the user's CPU and RAM and then send the result back over the Internet. Not saying that's what's happening, of course, but it's not completely ridiculous.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 5 months ago (2 children)

    But it isn't this idea just kind of a reverse cloud? Since running AI is so expensive, they could "borrow" other people's ram. Just an idea.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 5 points 5 months ago* (last edited 5 months ago)

    Conceptually what you're describing is feasible, there's lots of distributed computing projects that borrow compute/space/bandwidth for their ends but is unlikely to have any practical use.

    If there were a distributed system that could be used as memory in a large virtual inferencing machine, it would be incredibly slow. The model would be stored across a large number of different computers which would all have to coordinate. Each step of inferencing would be orders of magnitude slower because the latency between two different computers is orders of magnitude slower than the latency between a GPU and physical RAM.

    On the other end, if we just assume inferencing is feasible in reasonable time through some technique that isn't public... a model that was large enough to take advantage of an Internet worth of memory doesn't exist.

    So, assuming you had access to the largest model that we know of, you would have a model as complex as Claude Opos, but would take hours or days to respond finish inferencing and the quality would be about the same as you could get in under a second for $20/mo.

    And, going with a hypothetical 'Internet Scale' model.

    First, it would have to be trained which would use take an incredibly long time. Some of these frontier models take months to train on the fastest hardware available, a larger model would take even longer to train due to the increased latency.

    More importantly, there are strong diminishing returns on capability vs model size. This is why the AI companies are focusing on agentic tasks, where the AI spends a lot of time talking to itself and using tools, rather than pushing for a model with more parameters. This is referred to as the "scaling wall" (though AI companies, for obvious reasons, deny that such a thing exists and smoking companies say there's no cancer risk for smokers).

    It's a neat idea (Skynet may be loose on the world, hiding out as widespread 'bugs' that happen to consume a lot of resources and compute), but it would require a lot of things to be magic'd into existence to be remotely practical.

    You may find this funny: https://youtu.be/JcJSW7Rprio

  • permalink
  • fedilink
  • source
  • parent