Posted by miniBill 19 hours ago
the kubernetes reveal had me literally in tears
The dreaded Marcus Chen strikes again.
https://www.reddit.com/r/ClaudeAI/comments/1o3b4q2/just_rece...
If I am somehow wrong I would salivate at a chance to see the input.
And actually I see it clearly now, it has a bunch of signs I have called out multiple times myself. (It is entirely made out of lists of various types, and never states an opinion.)
Just my ego getting hold of me because I didn't realize it on my own.
Whether (or to what extent) he uses AI to generate the content he posts is a valid question.
I agree with your earlier reasoning that this is far more clever than anything I’ve seen AI produce yet. Lots of AI humor is dad-joke level at best. If it is AI then he’s trained it on a hand-curated collection of top-shelf satire.
Pangram indeed claims the OP is 76% AI-generated. It has "high confidence" (EDIT: some parts are "medium confidence") that the early portions of the text were created by AI, and "medium confidence" that some of the later potions were written by a human. EDIT: I was especially dismayed to see that the dog might have been an AI creation :(
When I use the "supporting evidence" option, the main piece of evidence Pangram provides is the frequent use of em-dashes. Each timestamp is followed by an em-dash. Personally I think the em-dashes could be a copy-pasted em-dash or inserted by a markdown to HTML converter. nesbitt.io is apparently using Jekyll [0] - any Jekyll users know anything about this??
Pangram's "supporting evidence" feature also considers → and € to be "unusual Unicode".
Personally, to me it looks like the "supporting evidence" feature still needs some work because Pangram's AI detection is probably a lot more sophisticated than a grep for Unicode symbols. In fact the feature even has a notice claiming that "These patterns aren't used to determine our AI score; they help you see why AI text often reads differently."
As for the rest of the OP's content, it would be interesting to compare the Pangram results to a timeline of a real vulnerability. I tried doing so, but exhausted my free "Pangram credits" - apparently the first 1000 words of this article [1] about the log4j vulnerability is considered 100% human.
[0] https://github.com/andrew/nesbitt.io
[1] https://www.csoonline.com/article/571797/the-apache-log4j-vu...