Hard to read an article by an author I respect when it starts by heavily emphasizing "alignment" and then over-attributing it to why Grok failed. As far as I know (and despite all the AI corporations publishing articles triumphantly proclaiming the power of their systems via spookiness): AI does not align or misalign, it Generates Next Word from a databank after receiving an input. Whether that input comes with the prompt "be unwoke" or the databank contains bad data, does not mean it's not working as intended. Even in being... No, especially in being Mechahitler.
this post was submitted on 28 Oct 2025
20 points (100.0% liked)
Politics
77 readers
84 users here now
Any politics anywhere in the world. Inevitably it'll be 99% US stuff, but that's not a rule.
This community works differently to how most politics communities work. It has strict rules designed to facilitate productive discussion. You can be rude, to a point, but you can't participate in bad faith:
- If you claim someone said something they didn't say, that's a temp ban.
- If you make a factual claim but then aren't interested in backing it up, that's a temp ban.
- If you're asked one or two reasonable questions about what you said, and you're still talking but you're pretending the questions didn't happen or rejecting the premise of answering them, that's a temp ban.
The idea is to make the discussion productive. Let's see how it works. Maybe this is a fool's errand but IDK how any set of moderation could be worse than lemmy.world.
Other misc rules:
- No unreliable sources.
- Keep it productive please.
- Self posts for discussion are fine. This includes videos or photos. No meme posts or screenshots please.
- No personal insults.
- No racism / transphobia / related bigotry.
founded 1 month ago
MODERATORS
Well, but that's accurate, though. They were trying to align Grok to certain political beliefs, but that isn't how it works (because it's just generating next word). So they had repeated failures including some very public and hilarious ones, so they're going back to the drawing board and trying to create a whole artificial reality now for AIs to draw from.
I don't think that Wikipedia features nearly so prominently in the corpus of a lot of these chatbots or how they form their worldview, and so I don't think that is why they are making Grokipedia, but I don't think the underlying ideas that fine-tuning political beliefs into Grok failed or that they are now trying to rewrite the fabric of reality are wrong. I am a little suspicious I guess about whether those two things are even strongly connected in their goals.
There's the technically accurate business definition, and then there's how it's used in the rest of the world. You can't wave a stick in a social media post or YouTube video about AI and "alignment" without hitting somebody who genuinely believes AI has sentience or is on the verge of breaching containment and spreading like a computer virus, just full sci-fi paranoia
AI does align or misalign, but not "AI" as a marketing term for chat bots and image generators
Imagine, in an era when GPT is obviously crap at most tasks, coming up with an even worse version that has even more obvious flaws.