Elektrine lite

← Feed

@scottjenson@social.coop

2026-04-18 18:00 UTC

Sure, this is funny but what bothers me the most is that we don't even have a good way of talking about this. This is a foundational error state and likely implies other problems in how it works. There is something very boring and yet important about this and we can't discuss it easily. /ht @andrewhinton@jawns.club

Replies (4)

  • @MisuseCase@twit.social 2026-04-18 18:23

    @scottjenson@social.coop @paninid@mastodon.world @andrewhinton@jawns.club I tend to agree with Steve Gibson when he says that general-purpose LLMs are, fundamentally, not securible. Just by the nature of how they work. Narrow, purpose built LLMs and other AI tools may be more securible or at least have a smaller blast radius, but the current corporate pyramid schemes, uh, that is, market pressures are all based on general-purpose AI.

    Open ##3174122

  • @scottjenson@social.coop @andrewhinton@jawns.club Yeah, I would discuss it in terms of undefined behaviour. The more general-purpose something is, the more of its behaviour given a particular input is undefined. General purpose tools are for purposes you didn't want as well as those you did, and the purposes you thought about are about 0% of all the purposes, rounded down.

    Open ##3174126

  • @scottjenson@social.coop @andrewhinton@jawns.club Isn't the first failure here that they use an interface that is made for human to human communication for human to machine communication?

    Open ##3174127

  • @DJGummikuh@mastodon.social 2026-04-18 20:19

    @scottjenson@social.coop @andrewhinton@jawns.club without understanding of context and intent, this technology is doomed to keep failing hilariously, as has been demonstrated in this screenshot. 🤷 I don't see any mitigation strategy that would promise improvement in that area that isn't essentially a patchwork-blacklist

    Open ##3174128