[stubbi]

Yes, good point. I think OpenClaw actually helps here by making a broader audience aware of the security risks of using "unchained" LLMs. Securing a probabilistic system is a fundamentally different challenge than auditing kernel code, and we're all still figuring that out.

I am optimistic that OpenClaw will actually drive a lot of security tooling around the use of LLMs from here

[alexgarden]

Very silver-lining viewpoint. I suppose that when tens of thousands of users have their identities leaked by well-intentioned helper agents, we will certainly elevate the security discourse. ;)

My personal opinion is that transformer architectures are (by their nature) unsecure. When you pair those with "super duper extremely so-eager-to-help weights and autonomous access to private information - voila! Here we are.

What we need is a pairing for transformers and automation that works natively with them. We're in a post-rules-based world now, so it's going to be something new.

[stubbi]

Yeah, I want to be a techno-optimist. But usually we have to go through some valley first before we understand how the technology actually should be handled (I hope that with "social" media we are slowly reaching the end of the valley time).

[alexgarden]

Amen, brother.