I'm glad that we're making progress towards a deeper understanding of what LLMs are inherently good at and what they're inherently bad at (not to say incapable of doing, but stuff that is less likely to work due to fundamental limitations).
There's similarity here with, for example, defining the architecture of software, but letting an LLM write the functions. Or asking an LLM to write you the SQL query for your data analysis, rather than asking it to do your data analysis for you.
What I'd really like to see is a more well defined taxonomy of work and studies on which bits work well with LLMs and which don't. I understand some of this intuitively, but am still building my intuition, and I see people tripping up on this all the time.
> There's similarity here with, for example, defining the architecture of software, but letting an LLM write the functions.
Not so long ago, this was how early adopters of LLM coding assistants claimed was the right way to use them in coding tasks: prompt to draft the outline, and then prompt to implement each function. There were even a few posts in HN on blogposts showing off this approach with terms inspired in animation work.
I'm not necessarily suggesting always getting down to literally the function level, although I think that gives you excellent quality control, but having a code-level understanding is clearly an important factor.
People keep throwing this phrase around in relation to LLMs, when not a single “fundamental limitation” has been rigorously demonstrated to exist, and many tasks that were claimed to be impossible for LLMs two years ago supposedly due to “fundamental limitations” (e.g. character counting or phonetics) are non-issues for them today even without tools.
The models now whaste a vast amount of useless neurons memorising the character count the entire English language so that people can ask how many r's are in strawberry and check a tickbox in a benchmark.
The architecture cannot efficiently or consistently represent counting letters in words. We should never have forced trained them to do it.
This goes for other more important "skills" that are unsuited to tranformer models.
Most models can now do decent arithmetics. But if you knew how it has encoded that ability in its neurons then you would never ever ever ever trust any arithmetic it ever outputs, even in seems to "know" it (unless it called a calculator MCP to achieve it).
There are fundamental limitations, but we're currently brute forcing ourselves through problems we could trivially solve with a different tool.
> The models now whaste a vast amount of useless neurons memorising the character count the entire English language
No they don’t. They only need to know the character count for each token, and with typical vocabularies having around 250k entries, that’s an insignificant number for all but the tiniest LLMs.
>People keep throwing this phrase around in relation to LLMs, when not a single “fundamental limitation” has been rigorously demonstrated to exist
Some limitations are not rigorously demonstrated to be fundamental, but continuously present from the first early LLMs yes. Shouldn't the burden of proof be on those who say it can be done?
And some limitations are fundamental, and have been rigorously demonstrated, e.g.:
What part of "Specifically, we define a formal world where hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function. By employing results from learning theory, we show that LLMs cannot learn all the computable functions and will therefore inevitably hallucinate if used as general problem solvers. " doesn't carry the title, to ask mildly?
As with all the works that use too broad a definition of an LLM they prove too much. This work defines an "LLM" as a computable function obtained by applying a finite number of steps of a generic algorithm to an initial computable function.
What they really prove is that it's impossible to extrapolate unconstrained non-continuous function from a finite subset of its values. Good for them, I guess.
So substitute another phrase, if you prefer. It doesn't change the logic.
"Specifically, we define a formal world where bungling is defined as inconsistencies between a computable LLM and a computable ground truth function. By employing results from learning theory, we show that LLMs cannot learn all the computable functions and will therefore inevitably bungle if used as general problem solvers."
Character counting remains a huge issue without tools.
Are you using only frontier models that are gated behind openai/anthropic/google APIs? Those use tools to help them out behind the scenes. It remains no less impressive, but I think we should be clear.
The literal best public models still fail to count characters consistently in practice so I’m not sure what you mean. It’s literally a problem we’re still trying to solve at work
What's amazing is that they even can fairly reliably appear to count characters. I mean we're talking about systems that infer sequences not character counters or calculators. They are amazing in unrelated ways and we need to accept this so we can use them effectively.
I suspect character counting - counting small numbers in general in fact - is something that multimodal models will gradually learn through their visual capabilities. We have generative systems that are capable of generating an image of the word ‘strawberry’, and of counting how many strawberries are visible in an image; seems likely it’s possible for an LLM to ‘imagine’ what the word strawberry looks like and count the ‘Rs’ it can ‘see’.
Character counting errors are a side effect of tokenization, which is a performance optimization. If we scaled the hardware big enough we could train on raw bytes and avoid it.
This is kind of my point, we need to get better at describing the limitations and study them. It seems extremely clear that there are limitations, and not just temporary ones, but structural limitations that existed at the beginning and continue to persist.
That’s false. Larger LLMs learn token decompositions through their training, and in fact modern training pipelines are designed to occasionally produce uncommon tokenizations (including splitting words into individual characters) for this reason. Frontier models have no trouble spelling words even without tools. Even many mid-sized models can do that.
Wait, where can I learn more about this? I don't doubt that varying the tokenization during training improves results, but how does/would that enable token introspection?
Because LLMs can learn that different token sequences represent the same character sequence from training context. Just like they learn much more complex patterns from context.
You can try this out locally with any mid-sized current-gen LLM. You’ll find that it can spell out most atomic tokens from its input just fine. It simply learned to do so.
When I talk about fundamental limitations, I mean limitations that can't be solved, even if they could be improved.
We have improved hallucinations significantly, and yet it seems clear that they are inherent to the technology and so will always exist to some extent.
As a general architecture, an LLM also has limitations that can't be improved unless we switch to another, fundamentally different AI design that's non LLM based.
There are also limitations due to maths and/or physics that aren't fixable under any design. Outside science fiction, there is no technology whose limitations are all fixable.
Am I misreading that paper? They define hallucinations as anything other than the correct answer and prove that there are infinitely many questions an LLM can't answer correctly, but that's true of any architecture- there are infinitely many problems a team of geniuses with supercomputers can't answer. If an LLM can be made to reliably say "I don't know" when it doesn't, hallucinations are solved- they contend that this doesn't matter because you can keep drawing from your pile of infinite unanswerable questions and the LLM will either never answer or will make something up. Seems like a technically true result that isn't usefully true.
> Transform this image into a photographed claymation diorama of assorted artisan chocolates and candies […] viewed from a low-angle
Side note: whenever I read prompts for image generation, I notice very specific details which the model obviously ignored. Here the chocolates / candies in the last two images look anything but artisanal. They look very "sterile" and mass-produced. The viewing angle is also not accurate.
Why do we even bother writing such elaborate prompts, when the model ignores most of it anyway?
I have noticed the same thing.The few times I wanted to use image generatation it always failed me in exactly these aspects. I always put if off as a lack of prompting skill on my end. Once you start to keep an eye out for these inconsistencies they turn out to be very common.
This is 100% true. There are entire nodes/pipelines in programs like ComfyUI that are designed to take a simple prompt and "enhance it" which usually involves making it more verbose, adding detail, etc depending on the target model.
Original Prompt: "Man with Trapezoid Head"
AI Expansion:
Portrait of a man with a trapezoid-shaped head, sharp geometric facial structure, angular jawline wider at the top and narrowing toward the chin, realistic skin texture, detailed pores, dramatic studio lighting, ultra-detailed, 85mm lens, shallow depth of field, dark neutral background, cinematic, photorealistic, 8k resolution.
Note: Most people (outside the generative space) won’t pick up on this but in many cases if don't prompt otherwise, you’ll often end up with a prompt that’s better suited to older, keyword‑based models like Stable Diffusion which rely heavily on specific sets of positive and negative prompt keywords more akin to magical incantations to improve the output.
I wonder how long it took to come up with all this?
Because if I wanted a spiral of little "buttons" like the last one at the end (and they don't look very much like sweets) I'd be able to knock that out in Blender in an afternoon, and I'm not very good at Blender.
I think you're vastly overestimating the average persons ability to use Blender if you can do that in an afternoon; just figuring out how to place a colored cube and the camera probably takes an afternoon if you pick up Blender for the first time.
And knowing these little tricks to get what you want with image generation models also takes time. Not to mention you need some knowledge on some other software just to make the underlying layout.
It's not perfect, but it's been vastly improved in recent years. If you lost interest in 3D art because of Blender's bad UX in the past, I recommend you give it another shot.
Also, there might be other new 3D software with better UX. I am not a Blender fanboy, but I do love 3D art and graphics programming and want as many people as possible to get into it :^)
I work on a platform 3dstreet.com that does “underdrawing” but in 3d space which image models also struggle with. Another company intangible.ai does this as well: low poly 3d then image to image model.
It seems to be a very effective pattern. Curious if there are other examples out there. Or other names for this?
Isn’t this sort of just “chain of thought” (i.e. the seminal https://arxiv.org/abs/2201.11903 ) where the user is helping the model 1-shot or k-shot the solution instead of 0-shot? I’ve used a similar technique to great effect. I feel things are so new / moving so fast that it’s hard to have common lingo. So very helpful to have a blog / example! But I wonder if the phenomena has been seen / understood before and just in smaller circles / different name.
TLDR: use SVG to outline image correctly first, then send that image with your text prompt to get Gemini 3.0 Pro to render with correct numbers and text
Yup, that’s exactly what this is. If you’ve been using generative models since the early Stable Diffusion days, it’s a pretty common (and useful!) technique: using a sketch (SVG, drawn, etc) as an ad-hoc "controlnet" to guide the generative model’s output.
Example: In the past I'd use a similar approach to lay out architectural visualizations. If you wanted a couch, chair, or other furniture in a very specific location, you could use a tool like Poser to build a simple scene as an approximation of where you wanted the major "set pieces". From there, you could generate a depth map and feed that into the generative model, at the time SDXL, to guide where objects should be placed.
interesting that GPT Image-2 managed to 2-shot this with thinking turned on, I didn't save a copy and it disappeared from my window but I first got a failure very similar to the one in the article, but it saw the issue and said it was going to use a reference image, after which it came out with https://i.imgur.com/hlWpQNT.jpeg
It’s obviously not a new model capability. But using this well-known, existing capability to solve this particular issue is only obvious after the fact.
It’s a useful trick to have in one’s toolbox, and I’m grateful to the author for sharing it.
It's not novel in the sense that nobody knew about img2img. It's novel in the sense that nobody thought of using img2img to solve this problem in this way.
It's novel if you never played with img2img, including especially several forms of (text+img)2img. Or, if you never tried editing images by text prompt in recent multimodal LLMs.
That said, I spent plenty of time doing both, and yet it would probably take me a while to arrive at this approach. For some reason, the "draw a sketch, have a model flesh it out" approach got bucketed with Stable Diffusion in my mind, and multimodal LLMs with "take detailed content, make targeted edits to it". So I'm glad the OP posted it.
They’re actually quite good at it. I’ve had a number of situations where I’ve wanted to re-render some of my older comics. You can basically tell any SOTA multimodal model (NB, GPT-Image-X) to treat them as storyboards and prompt for a specific style: newprint, crosshatching, monochromatic ink sketch, etc.
Another thing I’ve gotten very used to doing is avoiding the “one-shot” approach. If I generate something and don’t like the results, I bring it into Krita, move things around, redraw some elements, and then send it back in with instructions to just clean it up (remove any smudges or imperfections). The state-of-the-art models can do an astonishing job with that workflow.
Ok it might just be me then. I view Nvidia‘s DLSS as a similar thing. There was even this meme that video games will in the future only output basic geometry and the AI layer transforms it into stunning graphics.
The standard objection: if the LLM is supposedly intelligent, why can’t it figure out on its own that this two-step process would achieve a better result?
Because image models at the basic level are just text tokens in, image tokens out. You'd need an agentic process on top to come up with a strategy, review output, try again, and so on.
I believe Nano Banana and gpt-image-2 have a little of this going on, but it's like asking a model to one-shot some code vs having an agentic harness with tools do it. Even the most basic agent can produce better code than ChatGPT can.
LLMs have no concept of what makes the output "good".
Or to put it another way, if the LLM generates an image with jumbled numbers it's because that was the most likely output, hence it was a "good" image according to its weights.
Part of the problem is that it isn't the LLM making the image directly itself, it's the LLM repeatedly prompting edits for a separate edit diffusion model. The Gemini reasoning summary shows part of this. The style of some of the images makes it also clear that it uses an Imagen 4 derived diffusion model underneath.
Of course many, even most, painters do sketch what they intend to paint, likely that's the predominant technique.
But it's not universally true, particularly among artists working in the last 100 years or so. Certainly Jackson Pollock (whether one regards his work as good or not) didn't sketch out how he was going to distribute paint onto canvas. Another example is Morris Luis (and other "stain painters") who didn't sketch out how he applied paint to canvas.
You're comment is largely correct, just pointing out that more than a few "decent artists" didn't (or don't) work that way.
I was thinking about doing the opposite for the common task of "SVG of a pelican riding a bike". Obviously, directly spitting out the SVG is gonna be bad. But image gen can produce a really stunning photorealistic image easily. Probably a good way to get an LLM to produce a decent bike-pelican SVG is to generate an image first and then get the model to trace it into an SVG. After all, few human beings can generate SVG works of art by just typing out numbers into Notepad. At the core of it, we still rely on looking at it and thinking about it as an image.
Love the concluding note : it works, but not really.
So LLM/GenAI crave. An entire article to show that it's nearly there, yet it's not, despite convoluted effort to make it just so on a very very niche example.
But if it works part of the time, it's useful. It's easy for a human to check that the numbers are correct, and if they aren't, just regenerate the image. Orders of magnitude easier than creating the image from scratch without the model.
This seems analogous to how a human would do it accurately. If you asked an artist to paint stones in a large circular arrangement with the numbers in order in one shot, with no fixes or sketching allowed, it wouldn't be surprising to end up with problems in the arrangement.
I wonder whether this could be used to fine-tune image models to provide better outputs. Something like this:
1. Algorithmically generate a underdrawing (e.g. place numbers and shapes randomly in the underdrawing)
2. Algorithmically generate a description of the underdrawing (e.g. for each shape, output text like "there is a square with the number three in the top left corner). You might fuzz this by having an LLM rewrite the descriptions in a variety of ways.
3. Generate a "ground truth" image using the underdrawing and an image+text-to-image model.
4. Use the generated description and the generated "ground truth" image as training data for a text-to-image model.
This is closer to a world model - kind of similar to how one might use a realistic or semi‑realistic simulation engine to model the environment like GTA in order to train a self-driving model.
Because the image generation is powered by a diffusion model that is only guided by the transformer model and still has somewhat vague spatial representation especially when it comes to coupling things like counting and complex positioning.
But by using the LLM to generate code like an SVG graphic is made up of, and then using a rasterized image of that SVG as an input to the diffusion model, this takes place of the raw noise input and guides the denoising process of the diffusion model to put the numerical parts in the right spots.
The LLM is putting the SVG in the right order because the code that drives the SVG is just that - code - and the numerical order is easily defined there, even if it has to follow something like a spiral.
Edit: although LLMs now also may be using thinking modes with their feedback during generation to help with complex positioning when drawing something like an SVG, as I just asked claude to generate me one such spiral number SVG and it did so interactively via thinking, and the code generated is incredibly explicit with positions, so, that must help. But the underlaying idea to two-step SVG-to-diffusion model is the real key here.
It's normal to first create a plan, then allow agents to write code. But it seems to be surprising for many to first create a draft / outline of a picture, then go for a final render.
I hope this kind of stuff puts the idea to rest that we're close to actual AGI. Outsourcing this kind of basic stuff which a real intelligence would be able to do "internally" is a hack which works for this specific case but would prevent further generalizations of the task at hand.
But I'm forseeing the opposite. This kind of tool use will soon be integrated and hidden such that people will eventully say "see we solved the problem that AI can't do 123+456, now we are really really close to AGI. Yeah no, with an AGI, it would have been the AGI itself that would have come up with needing at tool, building the tool and then using the tool. But that's not what LLMs are. They are statistical machines to predict tokens. They are very good at it, but that's not an AGI.
Ive been doing charts for slides like this for a while. Noticed html viz was super reliable, but I could style it with diffusion model. Its very useful for data viz.
I don't think the MoE part has anything to do with it, but the current gen of multimoddal models can do thinking interleaved with autoregressive(?*) image-gen so it's probably not long before they bake this into the RL process, same way native thought obviated need for "think carefully step by step" prompts.
Like many AI things, it would have been considerably easier just to learn to edit images in GIMP or something. Instead of learning a valuable skill, you spent time working with a model that will be obsolete in a few months. Sunken cost fallacy, I guess.
I wish the opposite was true: that when I tell Gemini I want "a diagram of X" that it immediately breaks out Python and mathplotlib, instead of wasting my time with Nano Banana.
Inpainting/guiding from a sketch is how I've always used diffusion models. I thought everyone did that, or at least everyone who wasn't just trying to get some arbitrary filler material without much care of what the output looked like.
A few months ago I tried to make Le-chat Mistral output a French poetry in Alexandrin (12 vowels). Disastrous at first. Then adding in specifications that each line had to also be transposed in IPA and each syllable counted, it went better.
Still emotionally unrelatable, but definitely was providing something that match the specifications of there are explicit and systematically enforced through deterministitic means. For now I retain that LLM limitations are thus that they can't seize the ineffable and so untrustworthy they can only be employed under very clear and inescapable constraints or they will go awry just as sure as water is wet.
tldr: do a standard img2img workflow where you lay out a skeleton or skeleton or low-res version, and then turn it into the final high-quality photorealistic version, instead of trying to zeroshot it purely from a text prompt.
