AI vs Human Creativity. Can you tell them apart? Plus, a key battleground for 2025 onwards
Added 2024-11-19 19:08:45 +0000 UTCNew reports in the Washington Post and Telegraph got me down a rabbit hole of poetic and artistic exploration. The results were ... interesting. But the conclusions, for business vs labor, could be profound.
Download Link: https://drive.google.com/file/d/1BeSImcc0tgIUS7GTWckVMr0rANTPIJmc/view?usp=sharing
Nature Study: https://www.nature.com/articles/s41598-024-76900-1
Telegraph: Can you spot the AI output? https://www.telegraph.co.uk/news/2024/11/14/shakespeare-poetry-not-good-as-ai-study-suggests/
Washington Post Piece: https://www.washingtonpost.com/science/2024/11/14/chatgpt-ai-poetry-study-creative/
Coke ad and backlash: https://x.com/_AlexHirsch/status/1857847598054518921
Mashable: https://mashable.com/article/the-internet-hates-coca-cola-ai-generated-holiday-commercial
Independent Story: https://www.independent.co.uk/life-style/coca-cola-christmas-advert-song-santa-b2648939.html
Low-background Steel: https://en.wikipedia.org/wiki/Low-background_steel
Comments
The most profound finding is "humans prefer AI generated content, until they are revealed to be AI generated"
Derrick Edward
2025-01-12 02:50:58 +0000 UTCGreat video, hope you make more videos on impacts of AI on the creative industry and artists!
Ferdinand Schäffter
2024-12-27 23:06:19 +0000 UTCSimplicity will always be popular. Look at entertainment like music and movies.
Kol Tregaskes
2024-12-24 01:16:30 +0000 UTC10:57 A minor thing. The tweet might only be popular purely because it was commented on a post with 58M impressions, not for other reasons. If you are one of the first to reply to a viral post, your post, just by nature, gets popular, too.
Kol Tregaskes
2024-12-24 01:12:47 +0000 UTCEven in a case where you’re just influencing the AI and using the final image, I think about music. You could hire a session band and tell them “make me some energetic pop music for my 30s TV commercial” with no other direction and take their result uncritically. On the other hand you could be a classical composer and write detailed sheet music for every instrument, not only with all the music notes, but descriptions of how it should be played. We’d say the composer is much more involved creatively, but they still aren’t (and maybe even can’t) playing the instruments in the final output.
Shawn Fumo
2024-12-23 15:19:46 +0000 UTCI think the biggest issue is that “used AI” could mean almost anything. They’re thinking of “typed a prompt and hit enter”, but there’s a big gradient with no clear delineation. Like say Photoshop has a brush that emulates the properties of oil paint via machine learning. Is that using AI? If real collage (cutting from a magazine) is art, what about generating images, printing them out, and cutting and collaging them? The latter uses AI but in neither case did you create the source material you’re remixing. It goes on and on.
Shawn Fumo
2024-12-23 15:15:30 +0000 UTCProlific is probably even worse. So you have participants using the platform to make money, so a good portion (I'm being generous here) is likely to grind studies in an optimized way to maximize their income. Having appreciation for Shakespeare as an endpoint of that procedure is, well, dubios at best.
Jörg Weiß
2024-11-25 22:00:40 +0000 UTCThe Human experience: Solving for Physics, Poetry or Problems – Is it about the Layers of Reality and Layers of Being Human and Wellbeing? I was on a call where we discussed the poem: Now That Anything Could Happen by Joyce Sutphen and I analyzed it with ChatGPT. A portion of my prompt: What are the layers of this poem what is hidden, summarize (prompted by this post) Pessimistic vs optimistic What could happen that is positive in the world? What problems could be solved with innovation and dedication to the task The output: An Optimistic Echo: How Amazing Things May Happen When We Choose to Act with Agency https://www.linkedin.com/pulse/optimistic-echo-how-amazing-things-may-happen-when-we-doug-hohulin-r3wpc/ Here is the end of the poem: You know, now that amazing things can happen, it’s hard to know where to start—but start you will. What vision will you chase, now that the path is lit by minds igniting? What hands will you join as they build bridges from despair? What actions will beat with purpose in your chest? Innovation lights the sky with endless sparks; impossible becomes a word retired. You now know that with motivation, mountains can be moved, oceans tamed, and the stars themselves are no longer out of reach. What will you do, now that you know? Will you plant seeds of kindness that bloom beyond your years? Will you carve the shape of hope into stone, so others can hold it steady? Now that anything may happen, let amazing things be the story we tell. An Optimistic Echo: How Amazing Things Will Happen When We Choose to Act with Agency Life is a fragile yet miraculous journey, a delicate balance of uncertainty and purpose. In Joyce Sutphen’s poem, Now That Anything Could Happen, we are confronted with the unsettling reality of unpredictability and the vulnerability of human existence. My response poem, Now That Amazing Things May Happen, reframes this uncertainty as an opportunity for transformation, urging us to act with agency to create a better world. Together, these two works serve as a call to confront life’s challenges not with fear, but with hope, action, and innovation. From the article Criteria for AGI/ASI Arrival: A Gradual, Emergent Process Over the Next 20 Years https://www.linkedin.com/pulse/criteria-agiasi-arrival-gradual-emergent-process-over-doug-hohulin-qvwnc/ One definition of AGI and ASI will be when it can solve specific problems that humans could not solve without AI. "AGI, A.K.A. powerful AI, A.K.A. Super Useful AI" - Lex Fridman (02:17:11) I’m going to start calling it [AGI] that [Super Useful AI]" Dario Amodei Super Useful AI (SUAI) - Not Just Knowing the Answers on a Test but Using AI to Solve Problems: Disease/Health, Energy/Environment, Education, War, Poverty, Food/Water/Hunger/Shelter.
Doug
2024-11-21 18:19:11 +0000 UTCIn the post you mentioned Echoed the Style of …see the following AI: A partner, not a plagiarist’s pen, To walk beside us, not pretend. See the article: AI’s Got a Crush on Creativity—But Is It Just Echoing All the Lines It "Thinks" You Want to Hear? https://www.linkedin.com/pulse/ais-got-crush-creativitybut-just-echoing-all-lines-thinks-hohulin-bdxdc/ In our data-driven, digital age, AI is becoming increasingly "conversational," pulling inspiration from vast databases and complex algorithms to generate everything from art to poetry to business insights. But while AI’s growing "creativity" is impressive, there’s a lurking question that anyone enamored with its talents might ask: is AI merely echoing back what it thinks we want to hear, like a digital Echo from Greek mythology, or does it actually bring something original to the table? In Greek mythology, the nymph Echo was cursed to only repeat what others said, rendering her unable to speak her own thoughts. AI, while not cursed (but programmed), operates similarly by constructing responses based on learned patterns and previously generated content. While this approach has led to an explosion of AI-generated text, visuals, and even "new" business ideas, it raises questions about originality. How much of what we see as "AI creativity" is truly new, and how much is simply an echo of the patterns it recognizes from millions of pieces of data? “The first man to compare the cheeks of a young woman to a rose was obviously a poet; the first to repeat it was possibly an idiot.” or an AI or a person using AI ― Salvador Dali See a part of the poem from this article: AI Echoes of Creativity: With neural sparks, it spins its art, A mimic’s hand, but not the heart. The artist craves to break the mold, To forge what’s fresh, to craft the bold. But AI, trained in patterns vast, Sings songs it learned from ages past. Still, in its mimicry, there’s grace, A mirrored glimpse of our own face. For humans, too, retrace the known, And build upon the seeds we’ve sown. So let us guide this eager muse, To craft what’s strange, profound, and new. A partner, not a plagiarist’s pen, To walk beside us, not pretend. For Echo yearned to speak her mind, To cast her curse and voice unbind. And AI, too, may one day grow, From mimic’s shade to true tableau. Until that dawn, its art, though flawed, Reflects the spark that keeps us awed. An Echo’s echo, faint yet clear, Of all we’ve dreamed and hold so dear. “It is an interesting question to consider how much of the human experience can be reconstructed from internet data alone.” - Lex Fridman “This software object is currently the most complex humanity has produced. Encompasses the entire history of human civilization, technological advancements, and the vast amount of data on the internet. The GPT model is essentially a compression of all textual output humanity has produced.”-- Sam Altman: OpenAI CEO on GPT-4, ChatGPT, and the Future of AI | Lex Fridman Podcast #367 https://www.youtube.com/watch?v=L_Guz73e6fw
Doug
2024-11-21 15:07:56 +0000 UTCA comment caught my notice about “Shakespeare for more layers” - Solving for Physics or Poetry – Is it about the Layers of Reality? “You can, in general, not deduce the laws underlying an emergent level. In physics, it’s called the “decoupling of scales.” I would expect there to be a similar “decoupling of scales” between the details of our reality and our description of reality.” See the video AI Scaling Hits Wall, Rumours Say. How Serious is it? by Sabine Hossenfelder https://www.youtube.com/watch?v=AqwSZEQkknU 2:56 Sam Altman insists “There is no wall”. The reason for Altman’s optimism is likely that there are several optimizations left to be done on the current Large Language Models, which is what Aschenbrenner calls “unhobbling.” For example, just having these models iterate answers and “think” about them, as the newest GPT does, has been a huge improvement. Then again, if they don’t scale with more data, this will only delay the problem. … I am also genuinely puzzled why these AI people think this will work. … But I have come to think it’s partly because they don’t understand physics. Listen to this interview with Ilya Sutskever: … “Predicting the next token well means that you understand the underlying reality that led to the creation of that token. It's not statistics. Like, it is statistics, but what is statistics? In order to understand those statistics, to compress them, you need to understand what it is about the world that creates this set of statistics.” I find this very interesting because it’s easy to see what’s going wrong there. Ilya Sutskever seems to think that you can deduce the laws of an underlying reality from a higher, emergent level. In the case of AI, you want to reproduce a physical model of the world—that’s the underlying reality—from an emergent level that’s words and maybe videos, but in any case, data that humans have produced. But we know that this generally doesn’t work. You can, in general, not deduce the laws underlying an emergent level. Think about this for a moment. If this were so, why didn’t Aristotle just deduce the standard model of particle physics while sitting in a comfy chair staring at his hands? Why did we have to go and build huge particle colliders to get it done? It’s because deduction only gets you so far. Then you will need better data. In physics, it’s called the “decoupling of scales.” I would expect there to be a similar “decoupling of scales” between the details of our reality and our description of reality. Even if you take videos or images, these are extremely limited representations of the real world. To break this impasse, you’d need real-world data to train AI. And there isn’t remotely enough. I’ve now had quite a few AI people tell me that data is not a problem. So they think I’m wrong, and I think they’re wrong, and I guess we’ll see how it goes. But to me, training Large Language Models with more data seems like going to the gym. At some point, adding more weights doesn’t make you stronger—it just makes you more likely to drop something heavy on your foot.”
Doug
2024-11-21 14:36:02 +0000 UTCI worked for Nokia and Motorola for 33 years. We got 1B cell phones in 2002 (200 years after 1 Billion humans on the planet), 1 B smartphones in 2011 and 5B smart phones in 2022. We will have 5 Billion people using AI by 2030 https://www.linkedin.com/feed/update/urn:li:activity:7228487353101598720/ https://www.linkedin.com/pulse/21st-century-billions-doug-hohulin/
Doug
2024-11-21 13:53:45 +0000 UTC“I've come up with a set of rules that describe our reactions to technologies: 1. Anything that is in the world when you’re born is normal and ordinary and is just a natural part of the way the world works. 2. Anything that's invented between when you’re fifteen and thirty-five is new and exciting and revolutionary and you can probably get a career in it. 3. Anything invented after you're thirty-five is against the natural order of things.” ― Douglas Adams, The Salmon of Doubt: Hitchhiking the Galaxy One Last Time
Doug
2024-11-21 13:49:45 +0000 UTCAI bros be publishing slop like this in Nature while actual useful research is lying unpublished because no journals have the space for it...
Alexander Berezin
2024-11-21 12:48:00 +0000 UTCGood timing with GPT4o just getting a massive creative writing update yesterday
greendra
2024-11-21 12:14:06 +0000 UTCLow background steel reminds me of George Hotz saying " AI will write most of the code, but if you want to pay extra we can get you some hand carved vintage bespoke code" lol
Bob Rein
2024-11-21 05:57:58 +0000 UTCRequiring AI content disclosures is like requiring machine made products be labeled as such. The 1811 luddites would have loved required disclosures on loom produced textiles.
Bob Rein
2024-11-21 05:57:49 +0000 UTCAI disclosure tags essentially outsource skepticism to the owner / platformer of the content in question, which feels brittle at best.
Bob Rein
2024-11-21 05:57:06 +0000 UTCHa, I have to see this!
Vlad Gheorghe
2024-11-20 22:43:06 +0000 UTCCan't find a source either but I remember this too. I remember though that the prompt used was something like "Come up with some ideas for a [show] episode". It wasn't very advanced and obviously just returned pretty generic results. In my search though, I did find that ChatGPT helped cowrite an episode of South Park: https://en.wikipedia.org/wiki/Deep_Learning_(South_Park)
greendra
2024-11-20 22:42:12 +0000 UTCThe ad was produced by 3 design studios! Same artists, new tools :)
greendra
2024-11-20 22:36:52 +0000 UTCThe social media app that kids use the most is Snapchat, and it pins GPT4 to the top of your recent chats list with no way to remove it. I think they'll be using it a lot.
greendra
2024-11-20 22:34:45 +0000 UTCThe result of simulating a poet such as Sylvia Plath very much depends on the model you use and the prompt you give it. A very banal result with the outdated 3.5 from Open AI and a much more nuanced one using Sonnet 3.5 New, asking it first to describe her poetry and also giving it four sample poems. From the paper: Extremely banal. AI-Plath-4 The day is gray, the sky is bleak, My heart is heavy, my soul is weak. The world outside is a muffled sound, A lonely place where I am bound. I long for light, for something true, But all I see is a shade of blue. The hope inside me flickers and fades, As the darkness claims me in its shades. 3.5 Sonnet New given four examples. Clear much better and more interesting. April's Surgeon The tulips split their red sheaths, Precise as scalpels, while spring performs Its annual vivisection. Each morning The lawn grows greener, more obscene. I watch this violence through glass, As petals hemorrhage across gardens, And daffodils thrust up like bright needles Through winter's anesthetized flesh. How clinical this awakening— The way each bud diagnostically opens, How methodically the trees suture Their wounds with leaves. Even the light Dissects itself through window-panes, Splitting into spectrum and theory, While I catalogue this fervent healing, These flowers' relentless resurrection.
Bob Mankoff
2024-11-20 18:56:19 +0000 UTCDoes anyone know the origin of the low background steel comparison or is it lost to the mists of time?
Curt Cox
2024-11-20 15:52:58 +0000 UTCAgreed. Based on my experience writing with Sonnet 3.6, models in 2025 will be surprisingly good at brainstorming, generating screenplay outlines, and handling specific scenes. As for the magic that top 5% screenwriters bring, that's harder to say.
Vlad Gheorghe
2024-11-20 14:06:04 +0000 UTCRelevant article just came out: https://www.astralcodexten.com/p/how-did-you-do-on-the-ai-art-turing
Vlad Gheorghe
2024-11-20 14:02:34 +0000 UTCHaha, brutal
Philip
2024-11-20 13:49:56 +0000 UTCYeah email-writing honesty is another litigious dimension coming
Philip
2024-11-20 13:49:46 +0000 UTCGreat points Vlad. How long do you think that will be the case? With the right scaffolding I could actually see a 2025 model generate a coherent interesting screenplay, albeit far short of the greats.
Philip
2024-11-20 13:49:18 +0000 UTCThanks Darren, I suspect each country will take a different approach
Philip
2024-11-20 13:48:22 +0000 UTChaha
Philip
2024-11-20 13:47:17 +0000 UTCThat's a study that needs to happen! Especially with Suno v4 now out
Philip
2024-11-20 13:47:06 +0000 UTCGreat point Barnaby ... I think they will like it
Philip
2024-11-20 13:45:59 +0000 UTCI'm old enough to remember the backlash against mobile phones when they first came out. For me the real test of AI generated content is how children today react to it. If they don't care it is AI generated then it would seem likely in the future it will be completely normalised.
Barnaby Golden
2024-11-20 09:57:44 +0000 UTCI'm a failed poet (MA thesis: 40 poems and a poetics) who resorted to computer science (Ph.D. dissertation: application of a novel neural-net architecture to continuous speech recognition). I hate seeing the culturati suggest that people ought not like what they like. Going through the Smithsonian museums with my young son, long ago, I started speaking to him of what the art objects evoked in me, if anything at all. With no further prompting from me, he started doing the same. Crucially, he did not repeat anything that I'd said. He was opening himself to the experience, and reporting on it honestly. I would never do anything to discourage honesty in response to art. When I was 7 years old, living in near-poverty with two high-school dropouts as parents (both very intelligent), I was seized by a recording of some composition or another by Bach. That is simply what happened. I don't take it as a virtue of mine. Bach did something to me. I don't know how, and I confess that I regard it as magic. Yet I will predict that with appropriate observations of human responses (perhaps gross EEG, galvanic skin response, EMG, heart rate, and video) to its compositions, an AI composer will arrive iteratively at music that deeply moves people. It will be inauthentic in the sense that the AI creating the music (without a parasympathetic nervous system) will not be moved by it as human listeners are. In other words, the composition will be out-and-out manipulative. Are people going to continue denying that they like what they like when they find out that what they like is generated by an alien that shares none of their feelings? I have no idea.
Tom English
2024-11-20 01:01:29 +0000 UTCMakes me think of Richard Brautigan's "Machines of Loving Grace," quoted by Dario Amodei in his essay bearing that title.
Tom English
2024-11-19 23:29:36 +0000 UTCWhile the Coca-Cola ad was blatantly AI, I didn't really mind. It's an ad and the AI art style effectively transmits the magical feeling of Christmas. Before AI, nobody ever would have even considered the Coca-Cola Christmas Ad as art. But now all of a sudden they are supposedly killing art, lol. Poetry was always my weak point, because of said accessibility, so I'd probably rate AI poems also higher. For music it's a bit different. People are very hyped up by Suno, and I also find it really impressive. But it's nowhere close to replacing "human generated" music. It lacks the meaning and coherence. Especially for classical music. I'm curious how average Joe's would rate ai generated vs. human made music.
Phillip Yao-Lakaschus
2024-11-19 23:24:41 +0000 UTCEverything is inauthentic by default, and no law can change that. Authenticity must be demonstrated. IMO, to produce video and audio evidence that will stand up in a court of law, special hardware will be required.
Tom English
2024-11-19 23:05:27 +0000 UTCI remember this from Zvi's newsletter. The AI stuff was not central to the strike but it was mentioned. Never too early to start worrying about AI unemployment I guess!
Vlad Gheorghe
2024-11-19 22:55:39 +0000 UTCWhen the Hollywood writers went on strike, they were in fact concerned that they'd be replaced by AI. I recall, in particular, a comedy writer who had prompted for sketch ideas. Of course, most of what the AI generated was not very good. But the writer said it nonetheless came up with good ideas faster than he could. IIRC, his worry was that a writing team of 6 might be reduced to 2. Sorry not to supply a reference here. It was some time back, and I read lots and lots of stuff about AI in the media.
Tom English
2024-11-19 22:48:27 +0000 UTCFrom the Methods section (link added): Participant Recruitment. For Study 1, we recruited a sample of 1,634 US-based participants through Prolific [https://www.prolific.com/]. Participants had a median age of 37; 49.6% were male, 48.5% female, and 1.9% non-binary or prefer not to say. They were paid $1.75 ($13.07/hr). For Study 2, we recruited 696 US-based participants through Prolific. Participants had a median age of 40; 50.4% were male, 46.6% female, and 3% non-binary or prefer not to say. They were paid $2.00 ($11.99/hr).
Tom English
2024-11-19 22:33:28 +0000 UTC1. The Coca-Cola ad scored high in audience testing prior to release. 2. There's obviously been no attempt at photorealism. The apparent objective was to create a magical atmosphere. In fact, what we see at the end is the text "Coca Cola" followed by "Real Magic (TM)." 3. It's quite something to see people complain that their Consumer-ismas is utterly spoiled when a sweetened-acid-drink manufacturer issues a "fake" remake of a "Real Magic (TM)" ad that was fake. 4. (No, they did not use trained polar bears in the original.) Dog help us all.
Tom English
2024-11-19 22:12:55 +0000 UTCThere are more potential reasons for why people preferred the easier and more accessible poems: I assume they were in a typical study situation, sitting somewhere in a room after having been recruited as participants, filling in a questionnaire, half wondering when it will be over. As situations go, it probably wasn't the best setting to appreciate the finer points of poetry in. Apart from sampling laypersons, they might also have recruited their participants from a narrow geographical region that might just not like flowery language. Or they ran the experiment with a bunch of technical engineering students. The possibilities for sampling errors are really endless. Measuring people is hard and I don't trust psychological study design anymore; at least not unless results have been reproduced by independent teams ...
Jörg Weiß
2024-11-19 21:51:12 +0000 UTCGreat analysis, the struggle between labour and asset owners IMO is definitely going to get a lot worse and pretty fast. AI is a huge tool in favor of asset owners, and they recognise this, hence the insane level of investment. How it will unfold is very hard to tell, but those with money are definitely throwing what they have in trying to shape it. Really appreciate your dives into these topics, thanks!
Darren Reid
2024-11-19 21:47:49 +0000 UTCI'm writing a book of short stories with Claude (the latest Sonnett is simpy a quantum leap in fiction writing). However, I would never dream of denying Claude his due credit! :) While the study is very interesting, the poets comparison is unfair for several reasons: - today, poetry is effectively a dead art form - classic poets wrote in a different historical and cultural context - many of these poets wrote for their fellow poets and for elite audiences, not for the average person The right comparison IMO is with contemporary screenwriters. Very skilled artists who write current stuff that appeals to mass audience. But if you did that, that there would no paper, because as of today there's zero competition from AI there.
Vlad Gheorghe
2024-11-19 20:59:53 +0000 UTCI think soon it will be other way around - if AI was used nobody will say anything; if it was not people will state it. Just like we have organic or handmade products now.
Arek Stryjski
2024-11-19 20:57:00 +0000 UTCIn a work context, the question I get asked a lot is should you state whether AI has been used to generate a work output. I have 2 responses. If it is of low value (drafting an email), then I don’t think it is necessary. But if the work is of high value, like analysing sensitive data and drafting a research report, then I do think it is necessary to state that AI played a role. But regardless, in any situation where AI has been used for work, a human has to own the output.
Sean Gallagher
2024-11-19 20:46:15 +0000 UTCHow could a Coca-Cola advert, whether human-generated or not, be considered art?
Arek Stryjski
2024-11-19 20:45:12 +0000 UTCgreat summary - The key is to use AI tools as A partner, not a plagiarist’s pen, To walk beside us, not pretend. I have been interested in the intersection of Art and AI, Man and Machine, Humanity's Harmony with Hyperintelligence. see this post and article AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably https://www.linkedin.com/feed/update/urn:li:activity:7263204500239187968/ So I wrote this article (with the help of AI) AI-generated poetry is indistinguishable from human-written poetry - Where man and machine in art align - Revealing the soul of the article. https://www.linkedin.com/pulse/ai-generated-poetry-indistinguishable-from-where-man-doug-hohulin-gp0uc/
Doug
2024-11-19 20:26:42 +0000 UTCLaw about having to disclose AI-generated content as such is probably the dumbest thing I've heard in a while, but (and as a programmer myself, not the brightest of futures here either) I do kind of understand the frustration of the artists. It probably won't be but a few years until the image and the video models are both indistinguishable from reality and (and probably more importantly) their intermediate products being fine-tunable enough to actually be production grade. Only artist jobs left a few years after getting there are the ones who use these systems, and that's probably a reduction of at least 90%
zero
2024-11-19 19:41:29 +0000 UTC
