It's mostly a matter of increasingly complex and sophisticated sampling methodologies to sift for the valuable synthetic tokens.
There have also been some interesting results about models that use the same data converging to very similar behavior, but I can't find the specific reference at the moment. That one tracks intuitively with the notion that the model is building a latent representation of everything it sees.
What exactly do you imagine is happening that is forcing them to fake the benchmark results? You think they're afraid their IPO will fail if they don't make a convincing pelican, and are hiding an illicit attempt at manipulating said result? And getting caught trying to trick people wouldn't tank the price?
This might be heresy to say on HN, but this dinky forum and its recurring comments aren't actually the center of the known universe. Faking a pelican isn't worth ruining a 200 billion dollar business. They have bigger fish to fry.
Moreover, there has been a literal massive cross-model experiment showing there is no scientific evidence for improved pelican results over any other similar prompt. If y'all could kindly consider hard evidence instead of conspiracy theories, we could move on to more interesting topics.
It is fairly easy to go ask for a differently styled pelican, environment, background, etc. But that's not the point.
I was thinking just today about how, if I could hire my competition's lead designer/programmers, have them design a system for me cleanroom, that I'd effectively know everything there is to know about my competition's limitations. This is trivial to do with LLMs.
"Make me a thing that does X" is going to come out implemented basically the same way most of the time, as a kind of stack-overflow, well commented, internship project. (or, if undirected enough, as a collection of those based on a bog-standard architecture)
Same-y enough that you could train an adversary on it? Or hack it afterward? I am beginning to think so!
I would be curious to see if you would ask few designers to draw the pelicans with the same prompt to them, no further instructions given, and they should not know about the pelican AI benchmark thing, how that will turn out?!
is this a case where alignment/rlhf collapses diversity same as image models going from "base" to "turbo" does (where the base model is typically way more influenced by seed, while distilled/turbo models aren't as much)? I am also wondering if part of this is also prompting, maybe we should have a second simonw benchmark that appends "be original" to the prompt and see if all models try to be original the same way too.
After all if you ask 100 people to draw a horse, it is quite likely it will be a lot more similar to a "here's a stock horse photograph" than a "here's what picasso would draw when asked to draw a horse"
I think this may be close to what Ethan Mollick has discussed regarding the AI's having a specific "taste". This is how they draw pelicans but obviously if you asked them to make different versions that are more creative, they could do that. For now, this is one way humans can provide a lot of input and direction.
Would any teacher of, say, grade 4 students, reading this ask them to draw the same prompt and let us know the result? It would be so interesting to see how we differ from the machines, if at all.
I've pointed this out so many times. Even if you crank the temperature of the LLM up to the point where the oceans boil off into space, at the end of the day, generic in => generic out where prompting is concerned.
It's the reason that if you ask a random child to draw a house, you'll end up with the classic “house with a pitched roof”: a square plus a triangular roof, one door, two windows, and sometimes a chimney.
I could draw a better house. However it would be a lot more effort. (I'm not a good artist - a good artist could draw a better house in much less time than me - but I can draw that classic house much faster than the artist can draw the better house).
For most kids there isn't enough time to draw a better house - they want to go play. They also are well aware that there is only so much time in class and the better house wouldn't be done and in turn look even worse.
Re house - the reason why people (not just kids) will draw the same symbolically looking house is due to the question itself: they want you to recognize the house in what they've drawn. No one wants to fail, everyone wants to please, and it's the same with llms.
If instead you ask to draw some fancy interestingly looking house you'll get more variety - now the task is not to have the question giver recognize the house in their drawing, but to show how they imagine fancy or interesting house.
Well I wouldn't necessarily say that the LLM is trying to please you (that's a bit too anthropomorphic for my liking), I think it's more effectively that it's a "regression to the mean". Same reason why you can instantly recognize LLM prose.
Can you prompt around it? Naturally, but that's kind of my point. The more generic your prompt, the more generic the response.
In my experience, simply adding modifiers like "interesting, unusual, bizarre, imaginative" doesn't necessarily help either. Try asking an LLM for a "insert aforementioned qualifier adjective" fantasy short story in independent context/ conversations a few dozen times and you'll quickly see what I mean.
I wouldn't draw such general conclusions out of this, because it doesn't explain why LLM outputs are samish. Why shouldn't a generic prompt produce a mix of generic-but-notably-different outputs, as would be the case if you sampled humans? AFAIK, the answer is that it's RLHF that causes models to collapse to a specific style, because base models tend to score much better on creativity and output diversity evaluations than post-trained ones.
> Why shouldn't a generic prompt produce a mix of generic-but-notably-different outputs, as would be the case if you sampled humans?
Yeah, I hear what you’re saying. I guess it depends on how we define “notably different.” As I mentioned elsewhere, if you ask a random group of 12 people to draw a house on a piece of paper, you’ll end up with almost identical results. Sure, there will be some minor differences but for the most part, it’s going to be the same square with a triangular roof, two windows, one door, and maybe a chimney.
> AFAIK, the answer is that it's RLHF that causes models to collapse to a specific style, because base models tend to score much better on creativity and output diversity evaluations than post-trained ones.
Interesting. I'd love to see a paper on this. I wouldn't mind trying some back-of-the-envelope tests on a RHLF vs Base version of a locally hostable Qwen model too!
The first time I personally heard of RLHF ruining some model qualities was actually in the GPT-4 paper (march 2023); figure 8 there shows that the base model has near-perfectly calibration on multiple-choice questions (!) and this stops being true for the RL model: https://arxiv.org/abs/2303.08774v6
I think this is the same sort of problem - that the miscalibration of RLHFed models also explains their lack of diversity. In the training dataset, for a prompt like "you're a helpful assistant, yadda yadda, write a site", there's a large range of possible completions which are quite different and yet similarly probable, and if a base model is well-calibrated on that space, that lets you sample from that entire space and get many diverse websites. But an RL-trained model is miscalibrated and chooses a much smaller subspace of completions to assign unnaturally high logprobs to, so no matter how you sample, all the websites you get will look the same. It's not obvious to me why this happens; I wouldn't be surprised if, for example, a new optimization algorithm gets discovered that prevents RLHF from having such effects.
Sort of a tangent, but if you remember AI Dungeon when it first came out, it was based on a fine-tuned GPT-2 checkpoint which of course was pre-RLHF. While its adherence to user prompts wasn’t fantastic, it got a lot of praise for its diversity at the time.
I almost wonder if you could do something like this: have a non-RLHF model that gets the user’s request and just spits out a cloud of possibilities. It might not be the most tightly tuned, but then you could have a RLHF model take that output and use it as a sort of quasi-seed to increase diversity, while still retaining the RLHF advantages of producing an ultimate response that aligns with the user’s initial request.
That second "Mind the Gap" paper seems to be doing something very much like that, though their "conformative decoding" (CD) seems very simple: they make both the RLHFed and the base model produce a probability distribution for a given context, then apply top-p truncation to the RLHFed probabilities as one normally does, and then they look up what logprobs the base model assigned to those same tokens, and sample according to the linear combination of the two models' logprobs. This seems to work a moderate amount for them.
Looking at the cited-by I found a paper that evaluated a bunch of such methods by LLM-judged novelty, https://arxiv.org/abs/2606.12790 . CD is weirdly weak according to these results, they get better changes with other methods - but all of the others are less technical and basically amount to asking the model to try harder at being original. I wonder whether CD would work for a task like programming...
Why does an SVG pelican imply that it should look polygonal?
I ran into this issue recently when asking Claude to design a logo for my van. It made a polygonal SVG. I asked it to make it more photorealistic and it looked way better, still an SVG though.
SVGs are mostly used for icons and polygonal images, because that's what they're good at. The entire point of using vector graphics over raster is to take advantage of the fact that for low-poly objects, SVGs are much smaller. Icony-ness is definitely implied by the specification of format here.
I mean in a lot of cases it seems like the leadership people see this as a win not a lose.
Lost count of the number of times a requirement is given in terms of how another product is doing it. "make chips like gmail uses for tags", "let's copy anthropics LLM chat UI", "we should try and make it look like X competitor".
"Looking the same as everyone else" seems to be a bonus in their mind
The article faults LLMs for not being more creative (riding in the city in the rain, in East Asia, in outer space, etc.). But, if prompted, "Generate an SVG of a pelican riding a bicycle," with no additional details or specification, then I would _expect_ something generic: a side view, nondescript terrain or perhaps terrain associated with a pelican (i.e., beach), little clothing (because birds don't wear clothes) with the possible exception of something strongly associated with bicyclists (a helmet).
The article talks about "the LLM diversity collapse problem," which makes it sound like LLMs are getting less diverse over time, but the linked benchmarks suggest that newer, more capable models are _more_ creative.
Out of curiosity, I tried a couple of prompt variations in Opus 5.5. "Be creative" did a beach scene but added fish to the bike basket and a lighthouse in the background - not too creative. "Pick a unique and creative variation of this concept" resulted in a "midnight circus act" of a pelican riding a penny-farthing on a tightrope while balancing a fishing pole with fish (and, this time, two lighthouses - the tightrope is implausibly strung between them).
> But, if prompted, "Generate an SVG of a pelican riding a bicycle," with no additional details or specification, then I would _expect_ something generic
I believe the claim is that without a detailed specification, you see very similar outcomes. Every pelican is going left to right, every picture has similar sky background, every picture seems to be drawn at the same time of the day (position of the sun), every picture has a few clouds... I don't have data on this, but I imagine the same question being asked to a group of people would result in more varied outcomes.
Put very simply, that is how they're supposed to work. They are supposed to result in the most likely result. Same reason they, unless made not to, output '17' when prompted to "pick a number between 1 and 30", because that is the most likely result from training; but that means that all LLMs output 17 except for forced temperature or reasoning in RL making them run a python command with random.randint(0,31), or that the prompt is recognised and something else is supposed to be chosen than what is statistically the most likely text response.
Therefore asking a group of people would result in more varied outcomes of something visual, but from the textual training data the direction is statistically the most likely so is what is output, that sky is the most likely, time is the most likely, cloud coverage, etc. You're prompting with as little as possible so it fills in the blanks with what as likely as possible. When you ask Alex from the North Pole they're going to fill in the blank with a different kind of weather than Bob from the Equator.
That does mean that an LLM itself is in essence not as creative as a human, but then it is a tool to be used by a human doing more than oneshotting "Generate an SVG of a pelican riding a bicycle" and expecting creativity from the tool rather than the user. If the user is lacking in creativity they can prompt an LLM to get wacky scenarios to change it up a little.
>They are supposed to result in the most likely result.
That's called greedy decoding and is never used in practice. Everybody's using random sampling, so it's the range of most likely results. This issue is all about that range being way smaller than the training data implies.
How many of these features would you expect to also see in children's drawings? I have a hunch that the sun in the upper left corner for instance is pretty overrepresented.
imagine you ask the LLM for a joke. the biggest models, multi-trillion-parameter (don't quote me) astras and fables and argons, only know like 5 jokes. obviously bullshit, but it is what it is.
same goes for pelicans. the canonical pelican riding a bicycle is riding on a sunny day on an asphalt road going through some lawns or praires, left to right.
but ask it to tell a joke about something specific, turns out it knows five jokes about those specific things.
you could say the LLMs are quite aware of the zeitgeist they're in, by construction (their RL frying automated feedback loops burn the zeitgeist right into the weights.)
I've started asking for "Generate an SVG of a pelican riding a bicycle in the style of Henri Cartier-Bresson," which is a hint I want something like his famous Hyères, France photo [1].
The Hyères photo has the bike going in the opposite direction and from above. It's also tricky because it's a focal point for the eye, but not the center of the photo. The other elements of the frame direct your eye, but that visual behavior is not obvious to an AI. Lastly the image must capture the unquantifiable "decisive moment" that made Cartier-Bresson so famous.
So far I haven't seen an AI that really gets this prompt. Adding specificity will improve the results to a degree, but I haven't seen an AI that really gets what makes a Cartier-Bresson photo.
This matches my experience with Opus 5.5. If asked to compose a synthwave song in csound, it produces something fairly generic but recognizably synthwave. If I add light guidelines on avoiding generic sound and being creative, the results are often weird and interesting.
The curious thing is that trying to prompt it by steering towards some specific goal indirectly gives much worse results than just stating the goal itself.
This isn't generic, it's the same thing. Why would you expect the same thing? There are limitless ways to picture a bird on a bike. White pelican? Brown pelican? Wings out for balance? Wings anthropomorphized into hands on handle bars? Bmx bike? Mountain bike? Riding towards you, away from you, on a race course, etc.
Do a Google image search for pelican, you get a diverse set of sub species of pelicans and different poses and angles for them. Go on a writing prompt forum, you get many diverse ideas from the same simple prompt. Hell, ask your friends at a dinner party to describe a pelican riding a bike and see what happens.
Now, there are a lot of images of bikes from that side angle facing right on the web. Why? Because that's how they're staged on e-commerce websites. Does this say something about AI?
All going in the same direction? With a sun? With a full mouth (pelicans dont look like that, but a cartoon one might - why no reference to realism). Most of them have clouds.
There are helmets (makes sense) but just as many scarves.
Is any one going to draw a gray pelican? Does it matter?
The question is designed to be as generic as possible, and were converging on the most generic and safe answer ... the question is WHY (more so when humans would have much higher diversity, even if wrong). We're asking the generic question because no one wants to repeat the "German solders from 1940's" fiasco: https://archive.is/5l66U (archive link to skip pay wall NYT).
This is what "alignment" looks like? Are we reaching some "ground truth", some platonic form around pelicans on bikes? Are we just back feeding the best output into the training data and this is the convergence we ended up at? Is someone who LOVES pelicans deeply offended that all of them have full mouths and it's just not anatomically accurate and we are all missing it?
I realize that some of this seems over the top, but there are foundational conversations to be had around this without the conversation devolving into someone accusing someone else of having an "-ism" style bias.
I haven't read the paper yet but I think the keyword here is SVG. Which is a 2D format, usually used for icons. So that already defines how it would look like in the end.
And funnily even in the diverse set, most of the images show the pelican riding from left-to-right. There's a ton of bias in our data and even in our brains that you would need to erase to get true diversity.
> This is the LLM diversity collapse problem no one is talking about
We’ve been trying to tell this on every pelican thread. All models were recreating the same composition. The sun, the road, the green, the scarf, going right. Even the style. They’re not obvious associations one would make given the prompt.
The lack of variety is probably the most underrated issue, and the most likely cause of annoying patterns you all hate. All current post-trained LLMs are mode-collapsed compared to the post-training set distribution, and this doesn't get nearly enough attention from the labs and independent researchers. It's far from being a transformer-only problem, it's a common failure mode in all kinds of models. It's not a fundamental limitation either, unsupervised models have no problem with being as diverse as the training set allows.
In your code, this makes the model subtly devolve into certain patterns and stereotypes that aren't best practices, it's just a randomly frozen trajectory that is extremely hard to change zero-shot, without running a second pass. In other words, it's the same problem as in writing. Moreover, this is harmful for software design, where the ability to recognize the uncertainty matters. This is the reason everybody does the same thing. Frontend design also suffers from that.
>If you need more creativity it can come from the user or the harness.
Not fully possible. You can randomize the prompt all you want but you'll always get repetitive patterns and same choices in the same generation. This failure mode makes the model learn a very rigid mapping between this input idea and that single output idea (at best! at worst it's multiple input ideas -> single output idea).
> In your code, this makes the model subtly devolve into certain patterns and stereotypes that aren't best practices, it's just a randomly frozen trajectory that is extremely hard to change zero-shot, without running a second pass.
But in practice, who uses models “zero-shot”? It is standard for coding agents to review their own code even before returning to the user — so a zero-shot agent invocation isn’t a zero-shot model invocation. And then many users will get a different run of the agent to review the change a(frequently even a different model being invoked by a different harness-e.g. I personally mix Claude, Copilot and Cursor)
> This is the reason everybody does the same thing.
If you have a human-in-the-loop reviewing the agent output, different humans will have different opinions on what the agent should have done instead, and they’ll tell the agent to redo its output in different ways. And then they tire of repeating themselves, and it becomes guidelines in AGENTS.md.
Additionally, a coding agent asked to add a feature to a large code base will frequently check for the code base’s existing design and coding patterns, and so will end up producing very different code than it would have given a blank slate.
Yes, obviously external sources like human input, random generators, tools will bring semantic entropy and variation across a single generation, but longer reasoning and unattended agentic loops and self-reviews won't, the intuition is wrong in this case. Anything isn't explicitly conditioned by external sources will be sampled from model's knowledge space, and if it's constrained by this effect compared to the post-training set distribution you get reduced capabilities. You can observe it in the form of everything looking the same.
If you fix mode collapse, the effect on software engineering would be neutral at worst, or an improvement at best. It's not a verifiable task and does have room for subjective judgement. For the same reason greedy decoding (temp 0) isn't used even for coding, it leads to degraded and pathologic outputs.
Right, and I definitively don't want to guess at what I might want. A "round trip to the user" is much better than getting bogged down implementing something I didn't ask for and don't want.
Why would I want an LLM to be creative? I want to use these tools to express my creativity, not its own.
If I'm asking simply and directly for a pelican riding a bicycle, then please honor my intent and gimme that pelican. I don't need it on the moon. I do not need a top hat. I just want what I asked for.
Fulfilling a simple request simply seems like a feature, not a bug.
They already do this, if you ask for a pelican on a bicycle you will get a pelican on a bicycle. However your simple prompt has very little semantic capacity and underspecifies the rest. This phenomenon makes the underspecified part devolve into annoying repetitive patterns that are the same for everyone and even in the same generation.
I guess we have different expectations from these tools. I want predictability and direct-ability. I don't know how we can have that from LLMs without some emergent same-ness in outputs to vague prompts.
Snapping out of the "boring and repetitive" rut here is on us to prompt more creatively. Maybe that's the new "garbage in, garbage out".
You can't predict it either way as it still makes "random" choice unrelated to what you asked (because you haven't asked for anything in particular!). It's just "frozen" random.
People aren't that creative either, ask someone "imagine a pelican on a bicycle" and then ask them "where are they" and you're going to either get the generic beach locale, or something familiar to the responder (i.e., where they live or where they ride a bike).
You'd probably need to ask thousands of people before you heard "a pelican on the moon" or "a pelican in neo-tokyo"
If you asked me I'd say the pelican is nowhere in particular, it's no particular time of day, it's not facing in any particular direction, the bike isn't any particular colour and the pelican isn't wearing or carrying anything in particular. I'm imagining a pelican on a bicycle. That is all.
This is inherently an optimization problem, and it affects a lot more than LLMs. With so much of the world being instantly connected, things converge to the "best" style at the loss of diversity.
For example, a lot has been written about how all corporate logos look the same now, https://news.ycombinator.com/item?id=32040506 , and also about how so much interior design and architecture has converged to the same look. In the 70s there may have been tons of styles people derided as ugly, but at least there was colorful diversity. Now so many higher end (and "aspirational high end") interior design has the same minimalist, gray/white/tan, "look you can't see my refrigerator or dishwasher" look. I was recently looking for an apartment and all of the "luxury" apartments I looked at (obviously take that adjective with a heap of salt) had nearly the exact same layout and style.
84 comments
[ 1.7 ms ] story [ 26.2 ms ] threadThe reason why LLMs cannot be purely* pre-trained on LLM outputs is because LLM output diversity is low.
This implies two things:
1) There is going be no functional difference between hypothetical Qwen 5.5 and Fable 5.5
2) LLMs are not AGI
- https://arxiv.org/abs/2609.37891
- https://arxiv.org/pdf/2508.10975
- https://www.datologyai.com/blog/eating-the-rainbow-a-primer-...
As well as recent model releases that used large fractions of synthetic data (Arcee Trinity):
- https://arxiv.org/pdf/2602.17004v1
It's mostly a matter of increasingly complex and sophisticated sampling methodologies to sift for the valuable synthetic tokens.
There have also been some interesting results about models that use the same data converging to very similar behavior, but I can't find the specific reference at the moment. That one tracks intuitively with the notion that the model is building a latent representation of everything it sees.
This might be heresy to say on HN, but this dinky forum and its recurring comments aren't actually the center of the known universe. Faking a pelican isn't worth ruining a 200 billion dollar business. They have bigger fish to fry.
Moreover, there has been a literal massive cross-model experiment showing there is no scientific evidence for improved pelican results over any other similar prompt. If y'all could kindly consider hard evidence instead of conspiracy theories, we could move on to more interesting topics.
I was thinking just today about how, if I could hire my competition's lead designer/programmers, have them design a system for me cleanroom, that I'd effectively know everything there is to know about my competition's limitations. This is trivial to do with LLMs.
"Make me a thing that does X" is going to come out implemented basically the same way most of the time, as a kind of stack-overflow, well commented, internship project. (or, if undirected enough, as a collection of those based on a bog-standard architecture)
Same-y enough that you could train an adversary on it? Or hack it afterward? I am beginning to think so!
I don't see why you couldn't do the same in real projects.
If the design looks too generic, you can just ask for a change, no?
After all if you ask 100 people to draw a horse, it is quite likely it will be a lot more similar to a "here's a stock horse photograph" than a "here's what picasso would draw when asked to draw a horse"
https://www.oneusefulthing.org/p/the-overhang
It's the reason that if you ask a random child to draw a house, you'll end up with the classic “house with a pitched roof”: a square plus a triangular roof, one door, two windows, and sometimes a chimney.
For most kids there isn't enough time to draw a better house - they want to go play. They also are well aware that there is only so much time in class and the better house wouldn't be done and in turn look even worse.
If instead you ask to draw some fancy interestingly looking house you'll get more variety - now the task is not to have the question giver recognize the house in their drawing, but to show how they imagine fancy or interesting house.
Can you prompt around it? Naturally, but that's kind of my point. The more generic your prompt, the more generic the response.
In my experience, simply adding modifiers like "interesting, unusual, bizarre, imaginative" doesn't necessarily help either. Try asking an LLM for a "insert aforementioned qualifier adjective" fantasy short story in independent context/ conversations a few dozen times and you'll quickly see what I mean.
Yeah, I hear what you’re saying. I guess it depends on how we define “notably different.” As I mentioned elsewhere, if you ask a random group of 12 people to draw a house on a piece of paper, you’ll end up with almost identical results. Sure, there will be some minor differences but for the most part, it’s going to be the same square with a triangular roof, two windows, one door, and maybe a chimney.
> AFAIK, the answer is that it's RLHF that causes models to collapse to a specific style, because base models tend to score much better on creativity and output diversity evaluations than post-trained ones.
Interesting. I'd love to see a paper on this. I wouldn't mind trying some back-of-the-envelope tests on a RHLF vs Base version of a locally hostable Qwen model too!
The first time I personally heard of RLHF ruining some model qualities was actually in the GPT-4 paper (march 2023); figure 8 there shows that the base model has near-perfectly calibration on multiple-choice questions (!) and this stops being true for the RL model: https://arxiv.org/abs/2303.08774v6
I think this is the same sort of problem - that the miscalibration of RLHFed models also explains their lack of diversity. In the training dataset, for a prompt like "you're a helpful assistant, yadda yadda, write a site", there's a large range of possible completions which are quite different and yet similarly probable, and if a base model is well-calibrated on that space, that lets you sample from that entire space and get many diverse websites. But an RL-trained model is miscalibrated and chooses a much smaller subspace of completions to assign unnaturally high logprobs to, so no matter how you sample, all the websites you get will look the same. It's not obvious to me why this happens; I wouldn't be surprised if, for example, a new optimization algorithm gets discovered that prevents RLHF from having such effects.
As for studies that measure this effect, I just found these: https://arxiv.org/abs/2505.00047 , https://arxiv.org/abs/2507.20956 , https://arxiv.org/abs/2604.16027
Sort of a tangent, but if you remember AI Dungeon when it first came out, it was based on a fine-tuned GPT-2 checkpoint which of course was pre-RLHF. While its adherence to user prompts wasn’t fantastic, it got a lot of praise for its diversity at the time.
I almost wonder if you could do something like this: have a non-RLHF model that gets the user’s request and just spits out a cloud of possibilities. It might not be the most tightly tuned, but then you could have a RLHF model take that output and use it as a sort of quasi-seed to increase diversity, while still retaining the RLHF advantages of producing an ultimate response that aligns with the user’s initial request.
https://en.wikipedia.org/wiki/AI_Dungeon
Looking at the cited-by I found a paper that evaluated a bunch of such methods by LLM-judged novelty, https://arxiv.org/abs/2606.12790 . CD is weirdly weak according to these results, they get better changes with other methods - but all of the others are less technical and basically amount to asking the model to try harder at being original. I wonder whether CD would work for a task like programming...
Why does an SVG pelican imply that it should look polygonal?
I ran into this issue recently when asking Claude to design a logo for my van. It made a polygonal SVG. I asked it to make it more photorealistic and it looked way better, still an SVG though.
Lost count of the number of times a requirement is given in terms of how another product is doing it. "make chips like gmail uses for tags", "let's copy anthropics LLM chat UI", "we should try and make it look like X competitor".
"Looking the same as everyone else" seems to be a bonus in their mind
This manages to be so much more joyless, bleak, and tone-deaf than any of the example slop. Impressive.
The article faults LLMs for not being more creative (riding in the city in the rain, in East Asia, in outer space, etc.). But, if prompted, "Generate an SVG of a pelican riding a bicycle," with no additional details or specification, then I would _expect_ something generic: a side view, nondescript terrain or perhaps terrain associated with a pelican (i.e., beach), little clothing (because birds don't wear clothes) with the possible exception of something strongly associated with bicyclists (a helmet).
The article talks about "the LLM diversity collapse problem," which makes it sound like LLMs are getting less diverse over time, but the linked benchmarks suggest that newer, more capable models are _more_ creative.
Out of curiosity, I tried a couple of prompt variations in Opus 5.5. "Be creative" did a beach scene but added fish to the bike basket and a lighthouse in the background - not too creative. "Pick a unique and creative variation of this concept" resulted in a "midnight circus act" of a pelican riding a penny-farthing on a tightrope while balancing a fishing pole with fish (and, this time, two lighthouses - the tightrope is implausibly strung between them).
“Generate an SVG of a pelican riding a bicycle. Be creative”
And it replied:
“I’ll draw a seaside pelican on a vintage bicycle, with a fish-filled basket and a scarf catching the breeze.”
So it came up with exactly the same fish basket and seaside idea. No lighthouse though. Maybe these ideas are just cliche though.
> But, if prompted, "Generate an SVG of a pelican riding a bicycle," with no additional details or specification, then I would _expect_ something generic
I believe the claim is that without a detailed specification, you see very similar outcomes. Every pelican is going left to right, every picture has similar sky background, every picture seems to be drawn at the same time of the day (position of the sun), every picture has a few clouds... I don't have data on this, but I imagine the same question being asked to a group of people would result in more varied outcomes.
Therefore asking a group of people would result in more varied outcomes of something visual, but from the textual training data the direction is statistically the most likely so is what is output, that sky is the most likely, time is the most likely, cloud coverage, etc. You're prompting with as little as possible so it fills in the blanks with what as likely as possible. When you ask Alex from the North Pole they're going to fill in the blank with a different kind of weather than Bob from the Equator.
That does mean that an LLM itself is in essence not as creative as a human, but then it is a tool to be used by a human doing more than oneshotting "Generate an SVG of a pelican riding a bicycle" and expecting creativity from the tool rather than the user. If the user is lacking in creativity they can prompt an LLM to get wacky scenarios to change it up a little.
That's called greedy decoding and is never used in practice. Everybody's using random sampling, so it's the range of most likely results. This issue is all about that range being way smaller than the training data implies.
same goes for pelicans. the canonical pelican riding a bicycle is riding on a sunny day on an asphalt road going through some lawns or praires, left to right.
but ask it to tell a joke about something specific, turns out it knows five jokes about those specific things.
you could say the LLMs are quite aware of the zeitgeist they're in, by construction (their RL frying automated feedback loops burn the zeitgeist right into the weights.)
The Hyères photo has the bike going in the opposite direction and from above. It's also tricky because it's a focal point for the eye, but not the center of the photo. The other elements of the frame direct your eye, but that visual behavior is not obvious to an AI. Lastly the image must capture the unquantifiable "decisive moment" that made Cartier-Bresson so famous.
So far I haven't seen an AI that really gets this prompt. Adding specificity will improve the results to a degree, but I haven't seen an AI that really gets what makes a Cartier-Bresson photo.
1 - https://en.wikipedia.org/wiki/Hy%C3%A8res,_France
The curious thing is that trying to prompt it by steering towards some specific goal indirectly gives much worse results than just stating the goal itself.
Do a Google image search for pelican, you get a diverse set of sub species of pelicans and different poses and angles for them. Go on a writing prompt forum, you get many diverse ideas from the same simple prompt. Hell, ask your friends at a dinner party to describe a pelican riding a bike and see what happens.
Now, there are a lot of images of bikes from that side angle facing right on the web. Why? Because that's how they're staged on e-commerce websites. Does this say something about AI?
All going in the same direction? With a sun? With a full mouth (pelicans dont look like that, but a cartoon one might - why no reference to realism). Most of them have clouds.
There are helmets (makes sense) but just as many scarves.
A shocking number have chains.
If we asked 1000 people to do this same exercise would we have this level of convergence? If they were split between men and women would we see this many chains? (https://www.messynessychic.com/2016/04/20/making-bicycles-fr... )
Would they even look like bikes?
Is any one going to draw a gray pelican? Does it matter?
The question is designed to be as generic as possible, and were converging on the most generic and safe answer ... the question is WHY (more so when humans would have much higher diversity, even if wrong). We're asking the generic question because no one wants to repeat the "German solders from 1940's" fiasco: https://archive.is/5l66U (archive link to skip pay wall NYT).
This is what "alignment" looks like? Are we reaching some "ground truth", some platonic form around pelicans on bikes? Are we just back feeding the best output into the training data and this is the convergence we ended up at? Is someone who LOVES pelicans deeply offended that all of them have full mouths and it's just not anatomically accurate and we are all missing it?
I realize that some of this seems over the top, but there are foundational conversations to be had around this without the conversation devolving into someone accusing someone else of having an "-ism" style bias.
Why do a lot of people like Fruit of the Loom logos had cornucopias in them?
We’ve been trying to tell this on every pelican thread. All models were recreating the same composition. The sun, the road, the green, the scarf, going right. Even the style. They’re not obvious associations one would make given the prompt.
I don’t want creative code. I want boring code following generally agreed upon best practices.
If you need more creativity it can come from the user or the harness.
>If you need more creativity it can come from the user or the harness.
Not fully possible. You can randomize the prompt all you want but you'll always get repetitive patterns and same choices in the same generation. This failure mode makes the model learn a very rigid mapping between this input idea and that single output idea (at best! at worst it's multiple input ideas -> single output idea).
But in practice, who uses models “zero-shot”? It is standard for coding agents to review their own code even before returning to the user — so a zero-shot agent invocation isn’t a zero-shot model invocation. And then many users will get a different run of the agent to review the change a(frequently even a different model being invoked by a different harness-e.g. I personally mix Claude, Copilot and Cursor)
> This is the reason everybody does the same thing.
If you have a human-in-the-loop reviewing the agent output, different humans will have different opinions on what the agent should have done instead, and they’ll tell the agent to redo its output in different ways. And then they tire of repeating themselves, and it becomes guidelines in AGENTS.md.
Additionally, a coding agent asked to add a feature to a large code base will frequently check for the code base’s existing design and coding patterns, and so will end up producing very different code than it would have given a blank slate.
If you fix mode collapse, the effect on software engineering would be neutral at worst, or an improvement at best. It's not a verifiable task and does have room for subjective judgement. For the same reason greedy decoding (temp 0) isn't used even for coding, it leads to degraded and pathologic outputs.
If you are using multiple models – especially models from multiple vendors – then you aren't dealing with a single "model's knowledge space".
tho who know to refine, say NO. & do more with less even in the age of too cheap to meter tokens.
that has always been the work that produced every delightful product.
If I'm asking simply and directly for a pelican riding a bicycle, then please honor my intent and gimme that pelican. I don't need it on the moon. I do not need a top hat. I just want what I asked for.
Fulfilling a simple request simply seems like a feature, not a bug.
Snapping out of the "boring and repetitive" rut here is on us to prompt more creatively. Maybe that's the new "garbage in, garbage out".
You'd probably need to ask thousands of people before you heard "a pelican on the moon" or "a pelican in neo-tokyo"
That's why I'm not an artist.
For example, a lot has been written about how all corporate logos look the same now, https://news.ycombinator.com/item?id=32040506 , and also about how so much interior design and architecture has converged to the same look. In the 70s there may have been tons of styles people derided as ugly, but at least there was colorful diversity. Now so many higher end (and "aspirational high end") interior design has the same minimalist, gray/white/tan, "look you can't see my refrigerator or dishwasher" look. I was recently looking for an apartment and all of the "luxury" apartments I looked at (obviously take that adjective with a heap of salt) had nearly the exact same layout and style.