←back to thread

Tim Bray on Grokipedia

(www.tbray.org)
175 points Bogdanp | 1 comments | 31 Oct 25 21:41 UTC | HN request time: 0s | source
Show context
siliconc0w ◴[31 Oct 25 22:32 UTC] No.45777421[source]▶
Not sure it still does this but for awhile if you asked Grok a question about a sensitive topic and expanded the thinking, it said it was searching Elon's twitter history for its ground truth perspective.

So instead of a Truth-maximizing AI, it's an Elon-maximizing AI.

replies(1): >>45777478 #
sunaookami ◴[31 Oct 25 22:38 UTC] No.45777478[source]▶
This was unintended as observed by Simon here: https://simonwillison.net/2025/Jul/11/grok-musk/ and confirmed by xAI themselves here: https://x.com/xai/status/1945039609840185489

>Another was that if you ask it “What do you think?” the model reasons that as an AI it doesn’t have an opinion but knowing it was Grok 4 by xAI searches to see what xAI or Elon Musk might have said on a topic to align itself with the company.

The diff for the mitigation is here: https://github.com/xai-org/grok-prompts/commit/e517db8b4b253...

replies(2): >>45777856 #>>45778502 #
epistasis ◴[31 Oct 25 23:31 UTC] No.45777856[source]▶
There's a chance it was unintended, but no proof of that.
replies(1): >>45778115 #
1. simonw ◴[01 Nov 25 00:15 UTC] No.45778115[source]▶
That's effectively impossible to prove, especially if you don't believe statements made by the only organization that has access to the underlying evidence.

I actually think that it's funnier if it was an emergent behavior as opposed to a deliberate decision. And it fits my mental model of how weird LLMs are, so I think unintentional really is the more likely explanation.