Post #4502683
2026-08-10 23:33 UTC
So, even if we were to accept that LLMs are extremely useful and their utility overwhelms all possible concerns about externalities[1], we would absolutely need to *replace* the load-bearing presumption of good faith that the code we are consuming isn't *trying* to harm us, because now it's not *trying* to do anything, it's just whatever gets through the filters we have put in place. We have to replace it with a sandbox.
[1]: they aren't and it doesn't
Replies (1)
-
@glyph@mastodon.social 2026-08-10 23:39
But there's _zero_ movement on a pervasive sandbox. Everyone's blithely proceeding as if the OSS packages are still as thoroughly reviewed, despite the majority adopting "agentic" tools. Headline after headline about "rogue AIs[2]" are doing bad stuff, whether it'sd dropping the production database or hacking a competitor. [2]: random outcome generators with outcomes insufficiently constrained by robust technical safeguards, where human & process-based safeguards have been implicitly eliminated