4 ms·
"Assistant" type devices which can actually do things besides reply can't yet use large language models safely. LLMs need to get past the problem of making up s
by Animats 4y ago
"Assistant" type devices which can actually do things besides reply can't yet use large language models safely. LLMs need to get past the problem of making up stuff when they don't know something. Until then, it's not safe to give them power over devices.
- valine 4y agoThere’s nothing particularly unsafe about LLMs. Giving Sirious access to a weather api and timers is not gonna break the world. If anything, giving LLMs access to apis makes them less likely to hallucinate.
- 8ytecoder 4y agoThey can - if we split it into parsing, execution and response. If LLMs only do parsing and response and execution is limited to actions that are prepogrammed, i think we’ll have a much better assistant. Like, it could be interactive and natural. But i agrees that it shouldn’t be given unfettered access to do whatever it wants.