2 ms·
> Uncensoring a model also doesn't necessarily improve generic use cases. While the following is not a generic use case, I have a funny anecdote about how cens
by gpugreg 5mo ago
> Uncensoring a model also doesn't necessarily improve generic use cases.
While the following is not a generic use case, I have a funny anecdote about how censorship is holding back flagship models.
I was asking an uncensored version of Qwen3.6 how a CLI option of llama.cpp worked, and to my horror and amazement, it rudely went and decompiled the binary to figure it out. It felt like the computer-equivalent of asking a vet why my dog looks sick, who then proceeds to cut it open to check. Flagship models usually do not do that without some convincing, but it sure is effective.
We will need much better sandboxes when less restricted models become more common. I can already see them hammering out 0-days when they are prompted to do some task that usually requires root.
- faitswulff 5mo ago> Flagship models usually do not do that without some convincing Just a data point, but I’ve been having Claude do this regularly
- brookst 5mo agoSame. I was having it debug a routine python issue and it broke out mpympler and LLDB, and added a signal handler dump stack traces.
- bpavuk 5mo agoGemini Flash-Lite was a decent reverse-engineering sidekick since 2.5 as well.
- gpugreg 5mo agoI think I was using GitHub Copilot when I made the experience that led me to this statement. I guess the experience of using LLMs can be quite different depending on model version and harness.
- novok 5mo agowhats funny is if it looked up the source code on github it would've figured it out faster
- NooneAtAll3 5mo agowhat tool did it use to decompile it?