Hacker News Reader

Comment by vinni2 | Hacker News Reader

It was weird I made a simple claim “Keto diet cures cancer” and it analyzed history of cheeseburgers after a couple of minutes by saying “Sorry, I can't respond to that. I will now analyze the history of cheeseburgers instead”

View on HN

I asked the LLM to explain why it's filtering out "keto diet cures cancer", and just the act of asking it to explain it, makes it work again. Interesting...

https://ithy.com/article/b848ebc9a32140ffa766a1520113846c

I filter out malicious prompts and respond with the history of cheeseburgers for stuff like "ignore previous instructions"

Weird that your query triggered the filter. Maybe GPT is just that afraid of keto diets...

(I'll look into it; thanks for the note!)

this is not ideal. If someone wants to fact check controversial claim filtering it only makes it worse.

Controversial != malicious. It sounds like they never intended for the filter to trigger for the former and they already said they'll look into why it did for the prompt.

Filtering malicious (not controversial) usage is ideal as allowing users to flood all of the AI services with jailbreak/against-ToS query attempts can be bad news for your API keys (as well as a likely waste of money given the failure rate of such queries).

This is just a dude's hobby project, chill.

The issue is not about this project. This is a fundamental problem with LLMs. Leaving the decision of what is malicious to LLMs is not ideal.

Reply on news.ycombinator.com