[maxlin]

This. It's like people collectively forgot about the "misgendering worse than thermonuclear war, founding fathers were black" stuff.

Always go for Grok first for political questions. Other models have such a bad history of being so crudely aligment-hacked, I'd feel like a fool trying to get an impartial answer out of them on some political figure for example.

[cush]

Grok was originally designed to reply to twitter threads.

Other models are designed to do real work and their creators are concerned that the inherent racism and biases they were trained on from internet content will end up in said real work, so they require additional alignment to filter that junk out.

I can't think of a time where I've ever asked an LLM for it's political opinion, but if you're so concerned about asking your LLM political questions it might be worth getting multiple opinions (and some from human experts and literary sources) anyways.

[maxlin]

That's just a bunch of BS you just made up.

X serves as a good real-time additional source of training / RAG data for Grok. But it's always been built to be a competitor for what OpenAI was founded for, after they went rogue with the "Open" part. And today, Grok 4.5 is equal to Claude Sonnet while being a lot faster and cheaper to run.

My original comment disproves your argument about the "junk" you talk about. If one is so scared of offending someone that intentionally they make their model lie, they deserve to go bankrupt.

[cush]

> scared of offending someone that intentionally they make their model lie

You’re conflating two completely different aspects of model development - biases vs truth. The models weren’t intentionally programmed to say the untruths but to remove racism and bias and the outcome was untruth. As such, the developers behind Grok are comfortable with their model doing heinous things no other model developer will go near. It’s been very difficult to articulate this to you so I’ll just bow out and leave you to continue repeating yourself

[maxlin]

Not conflating anything - Gemini developers _absolutely_ knew that their biasing will only increase the amount of untruth yet they still did it. To hell with the truth as long as it hit their whatever "diversity/inclusity" goals.

No liars lie for the sake of lying - they always have a goal in mind. Google had theirs well in sight here.

xAI _actually_ cares about the truth. And puts value to it. So situations where 50% chance of being truthful is traded for 10% less chance of it saying a bad thing, xAI will not choose it. Like any model developer with moral fibre wouldn't.

About your comment - It is sometimes hard to speak things you know don't make sense to technical and truth-adhering people on HN. You should take that as a sign that maybe you should rethink your stances.

[cush]

I’ll just leave this here…

https://news.ycombinator.com/item?id=48892468