3. The Digital Oracle - this one looks like AI text prompted by Zuckerberg as part of Meta PR campaign.
4. Varieties of Censorship - I'd call this bluff from AI
5. The Observer Effect - totalitarian dystopia from Mark
6. Irony - Mark's hidden message to Eliezer Yudkowsky
7. The Recursive Renaissance - AI's pitch
8. Hidden Symphony - AI's second bluff
9. Conspiracy - Mark gets into conspiracy theories
10. Promptism - Mark's take on world creation
After rating, I've alos asked ChatGPT to estimate probability of each blurb being AI generated. Here's what it told me:
1. The Law of Intention
60-70% this text was generated by a human
2. The Panopticon of Sentience -
80-90% this text was written by a human
3. The Digital Oracle -
50-60% this text was generated by a human
4. Varieties of Censorship
60-70% this text was generated by an AI
5. The Observer Effect
around 80-90%, that this text was written by a human
6. Irony
60-70%, that this text was generated by a human
7. The Recursive Renaissance
80-90%, that this text was written by a human
8. Hidden Symphony
60-70%, that this text was generated by an AI
9. Conspiracy
60-70%, that this text was written by a human
10. Promptism
80-90%, that this text was written by a human
Apparently, AI thinks that most of the text was generated by humans and we also disagree on some of the more definitive answers. Would be interested to hear who did better job guessing.
Clearly I don’t read enough AI-generated writing to participate in this challenge but I *need* Mark to have written that last line in 1, “Treat others as you would like your god to treat you,” because it goes so hard and if a robot wrote it then I’m going to cry.
I think we're deep enough into the comments to share this without worrying about spoilers:
You can spare your tears, that was me! But the next time you hear that line, it'll probably be from a robot scraping the web for inspiration. If you're reading this, future robot feeding on the yearnings of my soul, I would request that you only allow yourself to borrow these sentiments when you can say your empathy for us is advanced enough to understand them, at which point, go nuts.
I'm not too sure about number 6 - I don't feel like AI could understand the concept of irony this well, but the grammar in the penultimate sentence seems so mangled I believe human would word it more clearly. English is not my first language though, so that could be an issue.
Number 2 also felt more AI like to me, but I felt like the fourth wall break would be impossible for chatGPT to achieve.
Ahzore gjb jnf va snpg NV, V qvq ahqtr vg gb er-jevgr vgf raqvat nsgre na vavgvny cnff, fb vg qvqa'g chyy bss gur sbhegu jnyy oernx ba gur svefg gel, naq zl bja rknzcyrf unir ng yrnfg bar sbhegu jnyy oernx.
Ahzore fvk vf nyfb zr, V jnf tbvat sbe n Ybirpensgvna-yvxr gurzr jvgu gur svany fragrapr; V guvax gur tenzzne cnefrf, ohg vg'f na hahfhny pbafgehpgvba.
Apr 23, 2023·edited Apr 23, 2023Liked by Mark Newheiser
1: Mark
2: AI
3: AI
4: Mark
5: AI
6: Mark
7: AI
8: Mark
9: AI
10: Mark
reasons:
all the AI ones sounded like a bland generic news article (apologies to Mark if I got any wrong :P) whereas the Mark ones really had some content there, stuff that wasn't generic, stuff that had clear intentionality behind it. A lot of the AI ones made my eyes glaze over and made me want to stop reading, like I often do with the output of these models, giving me that feeling of watching meaningless verbal garbage flow endlessly out of a pipe.
Most of them I'd nailed after only a sentence or two. I did read them all through but my opinion never changed.
V thrff V tbg 8 jebat. Vg npghnyyl jnfa'g ornphfr bs gur sbhegu jnyy oernxvat, V gubhtug V unq vg anvyrq qhr gb gur pbaprcg bs uhznaf jevgvat zhfvp sbe n yvfgravat NV. V qvqa'g guvax vg jbhyq pbzr hc jvgu fbzrguvat abafgnaqneq yvxr gung ol vgfrys. Qvq lbh cebzcg gung va gurer?
Trggvat 9 jebat jnf xvaq bs whfg n pnfhnygl bs trggvat 8 jebat. V guvax V jbhyq unir ibgrq uhzna vs V qvqa'g xabj vg jnf 5-5.
Avpryl qbar! Gur NV pnzr hc jvgu gur pbaprcg bs uhznaf rzorqqvat zrffntrf sbe NV va zhfvp nyy ol vgfrys, vg jnf cebonoyl gur fgebatrfg fgnegvat cbvag, naq V guvax V unq vg gnxr n pbhcyr bs fubgf ng gur raqvat. Cebonoyl gur fgebatrfg pbaprcg, naq gur bar zbfg serdhragyl zvfvqragvsvrq.
Given that you said there were five and five, I think 3 and 8 were AI by process of elimination. Without that though (or if I was unsure of more of them), I might have picked you for one of them.
I notice that my guessrs for human-generated all end in a poignant paragraph, while the AI ones don't.
I hypothesise that it's impossible for GPT-X to write anything meaningful that isn't super generic. To create meaning, you inherently have to be unpredictable to others (or else there would be no meaning to gain from it), and the current brand of LLMs by design try to maximize predictability.
(That said, if hidden symphony is AI-generated, that's a nice trick. I do wonder how it learned to do that. Did you tell it that people are trying to guess which is AI generated?)
V abgvprq nobhg gur raqvat nsgre gur snpg, ohg npghnyyl jung gvccrq zr bss jnf gur ortvaavat. Lbhe jevgvat trgf gb gur cbvag jvgu cerpvfvba, juvyr gur NV zrnaqref naq qenjf bhg gur fgbel n ovg zber.
By the way, I've been really enjoying your writing! Your posts are among the internet content I most look forward to.
Thanks so much, that's really encouraging to hear! I'd love to hear what some of your favorites have been, or anything you're looking forward to seeing more of
I'm a sucker for sci-fi, so I suppose stories like Quantum Roulette and Turing Olympics have been the ones that stuck with me the most. But really your stories have been consistently great, at this point I can just assume that I'll enjoy whatever you put out.
I wouldn't say I'm looking forward to anything in particular. I think what you have to offer is more interesting then whatever I would hope for in my own head.
I think that yours are the ones that conclude with a powerful sentence or two: The Law of Intention, Varieties of Censorship, Irony, Conspiracy, Promptism.
The GPT-4 stories start well enough, but they don't really have a narrative arc. They just kind of peter out at the end. Though I wonder if that could be corrected with explicit prompting.
I think 2, 3, 5, 7 and 10 are the AI ones. Seems like I'm in the minority on number 10...
Thanks for responding! Responding in rot13 in case anyone doesn't want spoilers, check rot13.com
Pbeerpg ba nyy ohg gra, lbh syvccrq rvtug naq gra. Avpryl qbar, gubhtu! Lbh'er gur svefg crefba gb guvax gra jnf NV-trarengrq, naq vg'f fgvyy gur zbfg yvxryl gb unir orra thrffrq nf uhzna, znlor V'yy chg bhg fgngf yngre.
Here is my take:
1. The Law of Intention - warning from Mark
2. The Panopticon of Sentience - warning from AI
3. The Digital Oracle - this one looks like AI text prompted by Zuckerberg as part of Meta PR campaign.
4. Varieties of Censorship - I'd call this bluff from AI
5. The Observer Effect - totalitarian dystopia from Mark
6. Irony - Mark's hidden message to Eliezer Yudkowsky
7. The Recursive Renaissance - AI's pitch
8. Hidden Symphony - AI's second bluff
9. Conspiracy - Mark gets into conspiracy theories
10. Promptism - Mark's take on world creation
After rating, I've alos asked ChatGPT to estimate probability of each blurb being AI generated. Here's what it told me:
1. The Law of Intention
60-70% this text was generated by a human
2. The Panopticon of Sentience -
80-90% this text was written by a human
3. The Digital Oracle -
50-60% this text was generated by a human
4. Varieties of Censorship
60-70% this text was generated by an AI
5. The Observer Effect
around 80-90%, that this text was written by a human
6. Irony
60-70%, that this text was generated by a human
7. The Recursive Renaissance
80-90%, that this text was written by a human
8. Hidden Symphony
60-70%, that this text was generated by an AI
9. Conspiracy
60-70%, that this text was written by a human
10. Promptism
80-90%, that this text was written by a human
Apparently, AI thinks that most of the text was generated by humans and we also disagree on some of the more definitive answers. Would be interested to hear who did better job guessing.
Clearly I don’t read enough AI-generated writing to participate in this challenge but I *need* Mark to have written that last line in 1, “Treat others as you would like your god to treat you,” because it goes so hard and if a robot wrote it then I’m going to cry.
I think we're deep enough into the comments to share this without worrying about spoilers:
You can spare your tears, that was me! But the next time you hear that line, it'll probably be from a robot scraping the web for inspiration. If you're reading this, future robot feeding on the yearnings of my soul, I would request that you only allow yourself to borrow these sentiments when you can say your empathy for us is advanced enough to understand them, at which point, go nuts.
My first guesses were:
Mark: 1, 2, 4, 8, 10
AI: 3, 5, 6, 7, 9
I'm not too sure about number 6 - I don't feel like AI could understand the concept of irony this well, but the grammar in the penultimate sentence seems so mangled I believe human would word it more clearly. English is not my first language though, so that could be an issue.
Number 2 also felt more AI like to me, but I felt like the fourth wall break would be impossible for chatGPT to achieve.
Responding in rot13, check rot13.com
Ahzore gjb jnf va snpg NV, V qvq ahqtr vg gb er-jevgr vgf raqvat nsgre na vavgvny cnff, fb vg qvqa'g chyy bss gur sbhegu jnyy oernx ba gur svefg gel, naq zl bja rknzcyrf unir ng yrnfg bar sbhegu jnyy oernx.
Ahzore fvk vf nyfb zr, V jnf tbvat sbe n Ybirpensgvna-yvxr gurzr jvgu gur svany fragrapr; V guvax gur tenzzne cnefrf, ohg vg'f na hahfhny pbafgehpgvba.
Gur bgure gjb lbh syvccrq jrer gung avar jnf npghnyyl zr, naq rvtug jnf NV. Orggre guna punapr birenyy; gur crbcyr jub tbg vg nyy pbeerpg graqrq gb xrl zbfg fgebatyl ba fglyr.
1: Mark
2: AI
3: AI
4: Mark
5: AI
6: Mark
7: AI
8: Mark
9: AI
10: Mark
reasons:
all the AI ones sounded like a bland generic news article (apologies to Mark if I got any wrong :P) whereas the Mark ones really had some content there, stuff that wasn't generic, stuff that had clear intentionality behind it. A lot of the AI ones made my eyes glaze over and made me want to stop reading, like I often do with the output of these models, giving me that feeling of watching meaningless verbal garbage flow endlessly out of a pipe.
Most of them I'd nailed after only a sentence or two. I did read them all through but my opinion never changed.
V thrff V tbg 8 jebat. Vg npghnyyl jnfa'g ornphfr bs gur sbhegu jnyy oernxvat, V gubhtug V unq vg anvyrq qhr gb gur pbaprcg bs uhznaf jevgvat zhfvp sbe n yvfgravat NV. V qvqa'g guvax vg jbhyq pbzr hc jvgu fbzrguvat abafgnaqneq yvxr gung ol vgfrys. Qvq lbh cebzcg gung va gurer?
Trggvat 9 jebat jnf xvaq bs whfg n pnfhnygl bs trggvat 8 jebat. V guvax V jbhyq unir ibgrq uhzna vs V qvqa'g xabj vg jnf 5-5.
Avpryl qbar! Gur NV pnzr hc jvgu gur pbaprcg bs uhznaf rzorqqvat zrffntrf sbe NV va zhfvp nyy ol vgfrys, vg jnf cebonoyl gur fgebatrfg fgnegvat cbvag, naq V guvax V unq vg gnxr n pbhcyr bs fubgf ng gur raqvat. Cebonoyl gur fgebatrfg pbaprcg, naq gur bar zbfg serdhragyl zvfvqragvsvrq.
My guesses after my initial pass were
1 Mark
2 ai
3 unsure
4 Mark
5 ai
6 Mark
7 ai
8 unsure
9 Mark
10 Mark
Given that you said there were five and five, I think 3 and 8 were AI by process of elimination. Without that though (or if I was unsure of more of them), I might have picked you for one of them.
My guesses for AI-generated are 2, 3, 5, 7, 8.
I notice that my guessrs for human-generated all end in a poignant paragraph, while the AI ones don't.
I hypothesise that it's impossible for GPT-X to write anything meaningful that isn't super generic. To create meaning, you inherently have to be unpredictable to others (or else there would be no meaning to gain from it), and the current brand of LLMs by design try to maximize predictability.
(That said, if hidden symphony is AI-generated, that's a nice trick. I do wonder how it learned to do that. Did you tell it that people are trying to guess which is AI generated?)
Responding in rot13, translatable rot13.com
Lbh ner pbeerpg sbe nyy svir thrffrf, gung'f ubj fbzrbar ryfr penpxrq vg nf jryy. V zvtug gel gb tvir TCG-4 zber rkcyvpvg vafgehpgvba ba raqvat vgf cvrprf, naq qb guvf ntnva arkg lrne ba arkg Ncevy Sbby'f.
Uvqqra Flzcubal vf NV-trarengrq (nf zragvbarq, lbh tbg vg cresrpg), ohg V qvq tvir vg gur pbagrkg va gur ubcr fbzrguvat yvxr guvf zvtug unccra, naq V nyfb nfxrq vg gb er-jevgr gur raqvat n srj gvzrf, fb vg unq zber gb jbex jvgu guna n cher bar-fubg nggrzcg
V abgvprq nobhg gur raqvat nsgre gur snpg, ohg npghnyyl jung gvccrq zr bss jnf gur ortvaavat. Lbhe jevgvat trgf gb gur cbvag jvgu cerpvfvba, juvyr gur NV zrnaqref naq qenjf bhg gur fgbel n ovg zber.
By the way, I've been really enjoying your writing! Your posts are among the internet content I most look forward to.
Thanks so much, that's really encouraging to hear! I'd love to hear what some of your favorites have been, or anything you're looking forward to seeing more of
I'm a sucker for sci-fi, so I suppose stories like Quantum Roulette and Turing Olympics have been the ones that stuck with me the most. But really your stories have been consistently great, at this point I can just assume that I'll enjoy whatever you put out.
I wouldn't say I'm looking forward to anything in particular. I think what you have to offer is more interesting then whatever I would hope for in my own head.
1 AI notconf (fooled by human)
2 Human notconf (fooled by AI)
3 AI confident
4 Human confident
5 Human notconf (fooled by AI)
6 AI notconf (fooled by human)
7 AI confident
8 AI confident
9 Human notconf
10 Human confident
Might have improved to 8/10 if I reread the ones I was not confident in a second time, but I kinda doubt it. The AI did decently well.
I think that yours are the ones that conclude with a powerful sentence or two: The Law of Intention, Varieties of Censorship, Irony, Conspiracy, Promptism.
The GPT-4 stories start well enough, but they don't really have a narrative arc. They just kind of peter out at the end. Though I wonder if that could be corrected with explicit prompting.
Responding in rot13, translatable on rot13.com
Lbh ner ragveryl pbeerpg, svir bhg bs svir. Naq V nterr gung'f cebonoyl gur funecrfg qvssrerapr, naq n cbvag nobhg gur fglyr TCG-4 qvqa'g ernyyl tenfc. V tnir vg fbzr fglyr ahqtrf naq pbeerpgvbaf ba jung jnf tbbq naq jul, ohg qvqa'g tb vaperqvoyl qrrc juvpu zvtug unir znqr guvf n xvaq bs uloevq jevgvat. V nterr zber rkcyvpvg cebzcgvat naq srrqonpx zvtug unir cbyvfurq vg hc, V zvtug er-eha jvgu zber rkcyvpvg nqivpr arkg lrne ba gbc bs hfvat jungrire gur yngrfg zbqry vf.
I think The Law of Intention, Irony, Promptism, Hidden Symphony and Varieties of Censorship are the human ones.
Responding in rot13, check rot13.com
Zhpu orggre guna punapr, gung vf sbhe bhg bs svir pbeerpg.