TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

Grok 3 Hallucination

4 点作者 11thEarlOfMar3 个月前
I created a service via prompt-smithing to generate 3 most-discussed, current stories for a specific country. The prompt specified that Grok would striclty review the X.com post traffic generated by users in that country. Under Grok 2, it needed more curation that I wanted to do, so I decided to wait for Grok 3. I had been under the impression that Grok 2 saw posts in real time, but then learned it did not. Grok 3 purportedly (and it insists) does see X posts in real time.<p>I switched to Grok 3 as soon as my account had it and tried the prompt again. When running the prompt for The United States, I noticed that the stories were not current. One story stated that Antony Blinken was Secretary of State when he hasn’t been for over a month. Another talked about the upcoming presidential election, showing that the time frame Grok was operating with was months ago.<p>I then asked Grok 3 to review and refine the prompt itself to ensure that the stories it generated were current topics and actively the top stories users were discussing. It provided a more concise draft that did seem to get better results for a number of countries. Then I ran on the United States again and… the stories did not seem to be the biggest and more relevant. Rather than negotiations to end the Ukraine war or Macron’s visit to the US or the raging about DOGE, it came back with a Timberwolves’ player opting for free agency. Really?<p>So I asked Grok 3 to show me the posts that led it to choose these three stories. It came back with quoted posts, one for each story. One quickly notices that the length, tone and style of all three example posts were identical. As if they’d been written by the same person.<p>I asked for the user accounts that generated them. Grok 3 produced 3 user accounts. None were actual accounts. When I pointed that out, Grok 3 helpfully explained that it was providing representative quotes that it had assimilated from X.com posts, and the account names were also representative of the types of account IDs that would post those representative quotes.<p>I then explained that doesn’t work for me. I need to see real posts from real accounts. Grok 3 apologized then helpfully presented 3 more accounts stating they were actual user accounts.<p>They weren’t.<p>So there it is.<p>I’ll move on to some other endeavor.

暂无评论

暂无评论