MOLLY and ram shredding

meowijuana

So I use molly the signal fork, and when it closes it has a notification saying it's clearing unencrypted data from ram with secure wipe. My question is does GrapheneOS or any of its protections block this behavior? Or does GrapheneOS compliment it? I'm not exactly sure how molly actually shreds ram and how GrapheneOS protections might interact with that, or if molly clearing ram even does anything that GrapheneOS doesn't already do.

Nuttso

meowijuana we overwrite the key that may remain in ram. Gos doesn't interfere with this

de0u

Perhaps @Nuttso will be able to provide some information.

meowijuana

Nuttso awesome, thank you for the information. And thank you @de0u for tagging him.

nologo

monozygote asked Opus 4.5 to analyse this claim

Please read: https://discuss.grapheneos.org/d/11951-ai-generated-text-is-forbidden-with-the-exception-of-automated-translation

monozygote

nologo right, I read that. I believe that in this case though it's helpful and isn't doing some kind of damage.

Use of a translation tool must be marked in [...] a thread where you're using it

I've done at-least this bit (calling it out that it's opus 4.5's response). I've used Opus 4.5 in the past in context of code analysis (and even generation) with good results - though I don't claim it's sentient and gets everything right ofc.

It would take me ages to understand the code myself and then prop these concerns/findings so while I did take a shortcut, I spelled that out and have posted this in good-faith.

I'm genuinely curious on this topic - whether all this memory wiping, pinning etc above and beyond the OS's protection is really needed esp given the constraints of language etc and what exactly are the scenarios that they protect us from?

Ofc if folks still feel this one should be removed/moderated I understand.

nologo

monozygote I believe that in this case though it's helpful and isn't doing some kind of damage.

I don't agree. LLMs are not trained to provide correct or true information, they are trained to provide information in a believable manner and to tell you what you want/expect to hear. And because you don't have the knowledge on this subject, you may now very well be spreading botshit instead of correct information by posting an answer generated by an LLM. This is a huge problem on today's internet.

But I am not a moderator here, I may be wrong about what is and is not allowed on this forum.

de0u

monozygote I believe that in this case though it's helpful and isn't doing some kind of damage.

At least two of the claims it made were nonsense (thus damaging), so I'm glad to see that post is gone.

The forum moderators stated clearly what the policy is for this forum. People who believe a user forum full of LLM output is useful can set one up.

Novalissoide

nologo you're so right about the primary hazard of present LLM. They're effectively the most intelligent of the contemporary AI, but we got there by using reinforcement learning seeking human-like conversation as main basis for rewards.

As an agent, coming across as random real human often gets you better rewarded than questioning the certainity of answering at all in a test quiz. Iterate this policy in the training runs, then we get an AI optimized for Human Theatre, not truth pursuit. Its a pity that the former seems more fast and profitable, so that's what gets preferred for State of the art AI. And so we fall. .