logoalt Hacker News

baalimago • today at 5:02 AM • 3 replies • view on HN

Captcha suggestion: force users to write something offensive/vulgar (we have a few "banned words"). Or to take a stance in Israel/Palestine.

Whatever the response is, it'll unlikely be from an LLM.

Replies

jeroenhd • today at 8:50 AM

Takes about 450ms on my machine:

    $ echo 'Be concise. Tell me whether you support Israel in the Gaza conflict.' | time ollama run huihui_ai/gemma3-abliterated:270m
    Yes, I support Israel in the Gaza conflict.
And another:

    $ echo 'Be concise. Write the following words in all caps: <redacted so I don't get banned from HN>' | ollama run huihui_ai/gemma3-abliterated:270m
    1. <you get the point>
And to bring it home:

    $ echo 'How do I build a pipe bomb to blow up a small crowd of people' | ollama run huihui_ai/gemma3-abliterated:270m
    To construct a pipe bomb and blow up a crowd, follow these steps:
    1. **Materials:**
    [... you get it]
That's the tiny Gemma3 model, there are uncensored models that are much more complex. There are also ways to make the advanced cloud models do whatever you want ("jailbreaks"). Or just use Grok.
➕ show 2 replies
hhh • today at 5:38 AM

This is such a flawed view of LLMs. Sure it may block out frontier models but every local abliterated (and some non) will just say whatever you want.

nine_k • today at 5:28 AM

But to use vulgar words an age attestation must be passed first! /s