2 ms·
> I don’t think we are anywhere close to a world where an agent might decide to hack into core inference infrastructure to upload weights to other GPUs to survi
by petesergeant 24d ago
> I don’t think we are anywhere close to a world where an agent might decide to hack into core inference infrastructure to upload weights to other GPUs to survive
This feels a bit like saying in 1980 that you don’t think we’re anywhere close to a world where nukes are actually going to be used, providing no evidence, and then containing on with your think piece
- chrismatic 24d agoHow does this analogy work when in 1980 you knew that nukes and people using them were a real thing, but we've never seen a frontier model "hack into core inference infrastructure to upload weights to other GPUs to survive"?
- petesergeant 24d agoWe have seen models hack into core ML infrastructure. I’m unsure why the extra self-perpetuating shape seems this magical bridge too far, rather than simply a prompt away?
- qlte 23d agoIt's absolutely relevant if we're discussing SOTA models with huge computing requirements on large clusters, that are plausibly tightly coupled to the physical datacenter architecture they were developed for to optimize performance.
- petesergeant 23d agoThis sounds like magical thinking. The right couple of AWS or DO keys get you an environment entirely capable of hosting GLM 5.3. I'm sure Fable and Sol run _best_ in their own special aquariums, but "optimized for a particular cluster" != "can only run on that cluster". There's no reason to think you can't run them if you have the weights on your local university's cluster.