the paper abstract is generated by ai: https://www.pangram.com/history/c480c94f-2d19-4cc3-95e0-c4fa...
Up until now you have been reading the words of Anthropic’s Claude, in particular Fable 5. In part, this is because I simply could not have written this piece. ... Although it might seem strange I thought it wrong to impose too much upon the machine, changing its voice to emulate mine. Richard Sutton’s bitter lesson would likely advise as much. Instead it has its unique voice, grating to some perhaps but altogether fitting that it should be able to keep it, and I only gave it advice on what considerations during writing would bridge the gap between its understanding and that of a reader.
"Here is, for those that have read this far, the acknowledgements that used to be in the formal paper but we decided to remove it, mostly because the paper has been reworked and rewritten so many times by us."
Edit: Just for fun I just defeated it with AI-generated text that it reported as 100% human-written. Great stuff. Truly genius scam.
I’m leaving out the ‘pseudo-scientific’ part of GP’s description, since it doesn’t entail anything technically.
The only counter example is when a human composed a piece deliberately trying to mimic AI slop.
So call it an AI slop detector if you want, but either way it accurately identifies stuff I don't wan't to read.
Is that the level of sensitivity we're looking for? More than literally 0%? Do you know how crazy it'd be to achieve that, even if they were trying to?
It has literally never been done.
What a hilarious inversion of credulity.
And as stated, even Pangram's own marketing material claims at least a 0.4% false negative rate on their own contrived corpus. A one in 250 event is hardly even rare.
You seem quite defensive about Pangram, and it shows up in a few of your comments. Is there a reason for that? If it's extending to outright denial of Pangram's own stated error rates... sheesh... very weird indeed (unless they're paying you, obviously).
Every time I've asked for evidence from someone who claims something like you do there always seems to be a reason it can't be posted. So it makes one suspicious.
Is there no way you can modify your prompt to be shareable?
That's another claim you made, which is quite silly. You've never found something that Pangram marks as 0% human and which you deem to be 100% human? Huge gulf.
And no, not really. It's copy from fundraising materials. Removing the identifiable portions of it would remove virtually everything, which is perhaps why Pangram didn't flag it?
What I wrote is true (it is a claim about my experience), so I'm not sure what you're attacking.
I amended my statement to avoid a discussion about the provenance which is not what I actually care about. What I care about is if I want to read it. Humans could theoretically write something in AI voice and if they did, I wouldn't want to read it.
I have not seen any examples of human text in LLM style that clearly predate 2023. If you can supply one, that would be incredible evidence.
If you're able to supply a prompt which generates text that fools pangram and is not like one sentence long or a list of names or something, but actual prose I want to read, then I will concede.
Duh.
If your statement was "my super special dowsing rods tell me what I want to read with 100% accuracy" I'd say "cool." But that wasn't your claim.
> If you're able to supply a prompt which generates text that fools pangram and is not like one sentence long or a list of names or something, but actual prose I want to read, then I will concede.
This is an even more conservative claim than the one you just made! Sheesh.
Have a good day.