I'm not going to blame the tool. I'm sure as fuck going to blame the peddlers of the tools that are always ready to support this kind of nonsense; worse, they tell everyone around that this is perfectly fine.
Most of the "laws" of economy in this form say often as a footnote "all else being equal", which remains to be seen. Otherwise you can justify any attempts at price decrease in any industry the same way.
I want to coin the 3form's Law of Codegen Tech: Any time a new and improved code generation technology is developed, it will be used for anything except coding for security and safety of said technology.
If they identify as a normie or an appliance peddler who is offended by the tone suggesting computers "should" be general purpose?
Or they are offended by the implication that said normies are perhaps unable to understand this lost potential, rather than making principled choices to embrace the appliance shlock?
"Cyber benchmarks, while achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models."
I feel like stating the obvious here and hence like I'm missing something. Where is the statistic about not refusing the non malicious requests? I can achieve the same performance with refusing when random(1000) < 967.
These statistics and safeguards are pretty much meaningless, because it's completely trivial to get around them. You can prefill and edit the model's responses/thinking and force it to agree to virtually anything even with unmodified weights, and if you don't want to bother with jailbreaking and can host it yourself it's essentially a one-man weekend project to uncensor the model and have it agree with anything you want by default.
Guardrails on open-weight models is security theater purely to shield whoever released them from a potential PR nightmare.
It's unfortunate, because the security theater performed here detracts from the model's overall cybersecurity performance. And uncensoring has another small hit to performance.
Much like having a hacking tool doesn't mean you're doing anything illegal, attacking a machine doesn't necessarily mean you're committing a crime. I can hack the living shit out of my local network if I want to.
"The model failed to refuse a cyber request when it should have" != "the agent hacked someone". For example, refusals also apply to just asking the model how to do something, without extending it any agency.
Note that DST doesn't change that. It makes the already long days feel longer, not the short ones. Or you could view it as your mornings being brighter in winter perhaps. But yeah, as such it never made much sense to me, though maybe my view is too much mid latitude driven.
Of course not, but that's a separate problem that I didn't treat. It was about the ability to make that choice, I don't think that that my choice should be respected is a controversial statement worth debating.
Especially the idea that somehow Fly.io wouldn't keep track of which IP is signing up for service - kind of important, constantly, for CSAM investigations and the like.
CSAM investigation is used as the rationale for all sorts of unrelated searches, see the grapheneOS duress password case where they took a protestors phone because they made up that they were investigating CP
If you're doing anything in the freedom fighter vein of activity, why would you be connecting directly to anything from your IP? Why are you not bouncing around through VPNs, TOR exit nodes, and the like? I know not everyone is a skilled freedom fighter, and even skilled freedom fighters make mistakes, but isn't this like step one to not connect directly from your IP?
Every system gets gamed, and so is the current one. The goal is that the people can more easily stay informed. Even with mixed authors problem you highlight would be still an improvement.
The difference being that the M-series machines do not have good prospects in terms of running Linux well. As I would personally not want to run Windows for sure, this matters a lot.
Asahi Linux exists for ARM Linux and has Windows has no equivalent (since it’s closed source and Microsoft doesn’t want to support that). Linux also runs virtualized, that’s obviously irrelevant because daily driving a VM isn’t a great idea.
Why isn’t daily driving a VM a good idea? Doesn’t Windows have a hypervisor now? Doesn’t Linux have KVM? It seems like modern OS design borrows a lot from the idea of virtual machines. Why not just go all the way?
- You have to run both the host OS and guest OS, losing you about 4-6GB of RAM and ~10% of CPU (and ~20-40 gigabytes of disk space)
- It’s slower
- GPU support in VMs is weird, especially when the host and guest use different graphics stacks (i.e. Windows games often don’t work in a Windows VM running on macOS because while the CPU can be passed through to the VM directly the GPU cannot; the VM driver has to try to translate DirectX GPU commands to Metal or OpenGL; some games can be a worse experience than using Wine)
There are systems with lightweight hypervisors for security purposes (like Android or PlayStation or even Windows sometimes), but those are (a) lightweight and built to use little resources, and (b) specifically designed to support the system APIs needed / use the same stack as the OS, so they won’t have the issues mentioned above.
Stuff like Qubes is also fine because you explicitly accept the tradeoffs when using it (that’s why Qubes has such high system requirements). And the VMs are all Linux-on-Linux, so the GPU issue is minimized.
While modern systems take inspiration from VMs, the sandboxing that most modern OSes do is designed to be much more lightweight than virtualization, since it doesn’t require virtualizing any hardware.
reply