2026-07-28 03:14 UTC
Replies (2)
-
@conniptions@mastodon.social 2026-07-28 03:22
@muddle@infosec.exchange @vrek@mastodon.social @lzg@mastodon.social It's even worse than that. "Guardrails" for LLMs consist of requests, in natural language, to the LLM, as part of the so-called 'system prompt'. Things like "IMPORTANT: DO NOT ADD ***ANY*** COMMENTS unless asked". It's a fucking joke.
-
@vrek@mastodon.social 2026-07-28 03:23
@muddle@infosec.exchange @conniptions@mastodon.social @lzg@mastodon.social I'm not saying that. I'm saying the use of "gaurdrails" at all is bad. It's similar to the driving thing. I don't want gaurd rails to prevent me from driving off the road. I want a car that can drive, stay in its lane, maintain appropriate speed, etc. Like here is the written history of us law, find me a case similar to this case I'm arguing. Here is the documentation of the go language, write me a tcp/http server with these assets. I give a road, stay on it.