Hacker News Reader

Comment by mullingitover | Hacker News Reader

I can't think of a less trustworthy group of people on model alignment.

They claimed that they had a rogue actor who deployed their 'white genocide' prompt, but that either means they have zero technical controls in their release pipeline (unforgivable at their scale) or they are lying (unforgivable given their level of responsibility).

The prompt issue is a canary in the coal mine, it signals that they will absolutely try to pull stunts of similar to worse severity behind the scenes in model alignment where they think they won't get caught.

View on HN

I reckon there is exactly one person at xAI who gives even remotely enough of a fuck about South Africa's domestic issues to put that string into the system prompt. We all know who it is.

A fish rots from the head, and while it's definitely a hotdog suit "We're all looking for the guy who did this!" moment, remember Musk is in charge of hiring and firing. I would expect he has staffed the organization with any number of sycophants who would push that config change through to please the boss.

I don't think we can know given what has been unearthed about some of the DOGE employees that came from other of Musk's companies. Not that it's unlikely that it's him.

Yeah, that one incident is enough reason for me to never bother using an xai model

That is my stance as well.

I think you're being snarky but that plus all the other X stuff is a trustbuster for many people.

I agree, Alignment is very important when considering which LLM to use. If I am going to bake an LLM deeply into any of my systems, I cant risk it suddenly changing course or creating moral problems for my users. Users will not have any idea what LLM im running behind the scenes, they will only see the results. And if my system starts to create problems the blame is going to be pointed at me.

See, if I was creating a product I would absolutely agree with you. I'd want an AI with tight guardrails, so innocuous that it would never deviate the slightest bit from a bland, center-left, vaguely corporate style of communication.

As a user, though, I want just the opposite. I want as close to uncensored with no guardrails as I can get. Nobody is giving you that unless you run your own models at home. But Grok is a little closer. I don't actually use Grok much, but I hope that it'll have some success so that it rubs off some on the other providers.

I've seen a lot fewer weird refusals from it than from Claude. Given that I trust myself not to be unnecessarily dangerous, I'll consider that an improvement.

Reply on news.ycombinator.com