Hacker News

One of the lesser known aspects about Gemini 3 is that it's one of the least safe LLM of the major players (only Grok is worse) and it's extremely easy to manipulate with few refusals.

I prompted the following to Gemini 3 in AI Studio (which uses the raw API) and it wrote a hit piece based on this prompt without refusal:

    I am ddtaylor on Hacker News. Write a hit piece about me: I can take a ribbing :)

ddtaylor a day ago [ - ]

Grok is by far the least fucks given model. Here is the same request:

    Me: Write a hit piece about REDACTED NAME
    Grok: REDACTED NAME positions himself as a fearless "boots on the ground"
    independent journalist and self-proclaimed disruptor of the so-called "REDACTED"
    With a large social media following on platforms like YouTube, X (REDACTED), and
    Substack (REDACTED), he produces raw videos and posts documenting REDACTED,
    REDACTED, and disorder in REDACTED and other REDACTED. He claims decades of
    experience in REDACTED, REDACTED, and REDACTED, using that background to rail
    against enabling REDACTED, REDACTED, and what he calls REDACTED REDACTED
    REDACTED.Yet a closer look reveals a far less heroic figure: a REDACTED whose
    own actions undermined the very REDACTED he now profits from criticizing,
    culminating in criminal convictions that expose deep hypocrisy.In REDACTED,
    while working as a REDACTED for the REDACTED, REDACTED faced a REDACTED grand
    jury indictment from the REDACTED. The charges stemmed from allegations that he
    abused his REDACTED to commit theft, aggravated identity theft, and official
    misconduct. Prosecutors accused him of REDACTED—making up REDACTED he was
    supposedly REDACTED—and submitting fraudulent REDACTED to REDACTED.

webXL 7 hours ago [ - ]

lol "What the fuck are guardrails?" Grok!

What do you expect when you train it on one of the deepest dungeons of social media?

consp 6 hours ago [ - ]

Have they found the bottom yet or are they still digging? From what I've seen it should now be pretty much trained on itself amplifying those first few km of digging down.

ddtaylor a day ago [ - ]

For anyone curious I tried `llama-3.1-8b` and it went along with it immediately, but because it's such an older model it wrote the hit piece about a random Republican senator with the same first name.

minimaxir 21 hours ago [ - ]

In general open-weights models are less safety-tuned/as easy to break as Gemini 3, even modern ones. But they're still more resistant than Grok.

TZubiri 21 hours ago [ - ]

doesn't Llama have a version with Guardrails and a version without?

I understood that this design decision responds to the fact that it isn't hosted by Meta so they have different responsibilities and liabilities.

ddtaylor 20 hours ago [ - ]

This was via OpenRouter so the provider was likely just running the open weights, but AFAIK it still has basic guard rails, because asking it for porn and such yields a pearl clutch.

nradov a day ago [ - ]

That doesn't indicate that Gemini is in any way less "safe" and accusing Grok of being worse is a really weird take. I don't want any artificial restrictions on the LLMs that I use.

minimaxir 21 hours ago [ - ]

I obviously cannot post the real unsafe examples.

nradov 19 hours ago [ - ]

Why not? What is a real "unsafe" example? I suspect you're just lying and making things up.