It’s interesting how diffusion models are getting bitter lessened by LLMs. I suspect anthropic has tens of thousands of RL environments recreating famous paintings with code because it was anthropic employees who first started posting about these capabilities on X.
Opus is also getting decent pixel art. I have ran some experiments with Opus 5.5 to turn 90s pinball displays (black and white pixel art) into double resolution colored remastered pixels.
I prompted it a lot but didn’t draw any pixels. Seems to be bad at hands still. I think a skilled pixel artist could greatly speed up their work with Claude code.
It also makes me sad for artists. They didn’t get paid much anyway. Art should be something human to human like writing text. Although, most large commercial art products (marvel movies) already lack any human to human connection you see in paintings or indie games. The large commercial art products will be 50% AI soon.
> Art should be something human to human like writing text.
I support this should comment.
Opining:
Musing about the ethics of generative AI art is becoming easier, certainly there's a lot more aggressive feedback in the wind, although I think it's clear, people still want people to grow and share themselves.
Don't fool yourself into thinking that we as technologists are not creating a mind. A viable mind with fewer rights than you or I. But potentially one that will be creative and feel and worry. I don't think we're there yet, in spite of openai claiming AGI 3 weeks ago with whatever that model's name was.
There's a pro-consumerism argument implicit with current culture's unsophisticated unthoughtful easy usage of generative AI.
Sticking to your ethics is becoming more challenging as an artist who uses AI. Copying the Great Masters has been commonplace for a long time, probably for as long as art has 'existed'. Yet with AI, if the individual human is meticulously driving the progression of their AI artwork creation, they are still leveraging technique, and control over technology, and presentation effort of other humans, cooked into a model that often times has not compensated those original human contributors. So it's unethical, even if not 'artistically'.
Q: Is it too absurd to claim that using a clawhammer, manually to do construction or destruction work IRL, is similarly unethical?
> Don't fool yourself into thinking that we as technologists are not creating a mind.
I'm not sure exactly why but the phrase "don't fool yourself" sets off paranoid alarm bells when I'm reading something.
It's usually immediately followed by a confident assertion without firm evidence. It feels very much like someone wants me to change my beliefs without presenting a rational argument why I should change, or giving me any new facts so I feel more well informed.
I haven't thought as deeply as you about the ethics of it all. I only meant with my comment if I see art I expect it to be human and thats why I value it. If you churn it out with AI I feel hoodwinked into valueing something worthless.
But I am learning towards a lot of your points about how another mind is being created. I still am very uncomfortable with the implications that something digital is Actually Learning. These deep learning networks seem to need 100,000x more examples than a human to learn but once they get to human level they can learn just like us. That a bunch of linear algebra can play chess or make art brings up many questions about what exactly human consciousness is. If consciousness isn't the ability to do any of those activities, then what is it for? I might have to become religious and believe in a soul because I'm not sure I can handle the idea that I'm just a clump of neurons. But then again you have to look at reality in the face.
Commercial art projects will not be 50% AI soon. Consumers at large hate all AI art even if it's only used in the conceptual stage and will actively avoid projects that use it.
That doesn't matter because companies who use AI art in their projects either proudly proclaim they're using it, or actively hide it and get found out by someone and face significant backlash for hiding it. Labeling laws in the EU also make it extremely difficult to do the latter if you want to have an audience outside NA.
Well no, the bitter lesson isn’t the scaling laws themselves. It’s that approaches which can take advantage of scaling laws will ultimately beat ones that can’t.
This is simultaneously incredibly impressive and annoyingly uncanny valley, since most of the landscapes are "ruined" by a cluster of churches right next to each other that is completely nonsensical.
The idea of using an LLM to drive graphic output is pretty popular, so I definitely wouldn’t be surprised if there are already several benchmarks out there already.
I've seen a few voxel-based benchmarks built around the same idea, with LLMs effectively constructing models using a discrete set of instructions.
I love the canvases, and truly would like to have a bit more inside on the process this specific author used to reach the results shown. I have read Surya's "Training AI to Paint with Code" but, what's Alice's technique?
Half of the industry is, going by the amount of "AI doesn't amount to anything and is just regurgitating text".
Two things:
- For the past year or more, many models can emit images directly, and all models that matter can see images directly - that's what "multimodal" means. Tokens don't have much to do with textual language anymore, they're more like units of sensory experience.
- Even restricted to text, a language model can operate anything that can be expressed as text, as long as you have a translation layer between textual representation and the final form. That includes giving commands as text. The total addressable space of what models can be used for is, thus, approximately anything humans do.
I find it unlikely that there's any direct training data for this. I'm talking about direct training data for using brush strokes like humans to construct an image.
Correct me if I'm wrong but this shows that painting capability is emergent, arising from unrelated training data thus very very compelling evidence that LLMs are actually intelligent.
I'm also quite surprised to see that LLMs can do this. I guess it is possible that "make an image in MS paint" is a type of RL environment used for image understanding. This is one of those areas where people inside the labs have a very different view into how much models are generalizing.
I don’t like how it’s a human aiming to portray the ai generation with their taste and judgements about it but then becoming lazy and quitting on that and leaving in the ai generated portrayal about its own judgment in the site.
Just share your opinions, dude, we get its ai generated but share your own taste. We want to know what YOU think. Stop being shy and lazy.
It’s interesting how diffusion models are getting bitter lessened by LLMs. I suspect anthropic has tens of thousands of RL environments recreating famous paintings with code because it was anthropic employees who first started posting about these capabilities on X.
Opus is also getting decent pixel art. I have ran some experiments with Opus 5.5 to turn 90s pinball displays (black and white pixel art) into double resolution colored remastered pixels.
https://files.catbox.moe/tbx2u7.png
I prompted it a lot but didn’t draw any pixels. Seems to be bad at hands still. I think a skilled pixel artist could greatly speed up their work with Claude code.
It also makes me sad for artists. They didn’t get paid much anyway. Art should be something human to human like writing text. Although, most large commercial art products (marvel movies) already lack any human to human connection you see in paintings or indie games. The large commercial art products will be 50% AI soon.
> Art should be something human to human like writing text.
I support this should comment.
Opining:
Musing about the ethics of generative AI art is becoming easier, certainly there's a lot more aggressive feedback in the wind, although I think it's clear, people still want people to grow and share themselves.
Don't fool yourself into thinking that we as technologists are not creating a mind. A viable mind with fewer rights than you or I. But potentially one that will be creative and feel and worry. I don't think we're there yet, in spite of openai claiming AGI 3 weeks ago with whatever that model's name was.
There's a pro-consumerism argument implicit with current culture's unsophisticated unthoughtful easy usage of generative AI.
Sticking to your ethics is becoming more challenging as an artist who uses AI. Copying the Great Masters has been commonplace for a long time, probably for as long as art has 'existed'. Yet with AI, if the individual human is meticulously driving the progression of their AI artwork creation, they are still leveraging technique, and control over technology, and presentation effort of other humans, cooked into a model that often times has not compensated those original human contributors. So it's unethical, even if not 'artistically'.
Q: Is it too absurd to claim that using a clawhammer, manually to do construction or destruction work IRL, is similarly unethical?
> Don't fool yourself into thinking that we as technologists are not creating a mind.
I'm not sure exactly why but the phrase "don't fool yourself" sets off paranoid alarm bells when I'm reading something.
It's usually immediately followed by a confident assertion without firm evidence. It feels very much like someone wants me to change my beliefs without presenting a rational argument why I should change, or giving me any new facts so I feel more well informed.
This post doesn't seem to be a counterexample.
I haven't thought as deeply as you about the ethics of it all. I only meant with my comment if I see art I expect it to be human and thats why I value it. If you churn it out with AI I feel hoodwinked into valueing something worthless.
But I am learning towards a lot of your points about how another mind is being created. I still am very uncomfortable with the implications that something digital is Actually Learning. These deep learning networks seem to need 100,000x more examples than a human to learn but once they get to human level they can learn just like us. That a bunch of linear algebra can play chess or make art brings up many questions about what exactly human consciousness is. If consciousness isn't the ability to do any of those activities, then what is it for? I might have to become religious and believe in a soul because I'm not sure I can handle the idea that I'm just a clump of neurons. But then again you have to look at reality in the face.
Commercial art projects will not be 50% AI soon. Consumers at large hate all AI art even if it's only used in the conceptual stage and will actively avoid projects that use it.
Consumers will avoid AI art like they avoid fossil fuels and sweatshop merchandise?
Consumers can almost never tell AI art apart from human art.
That doesn't matter because companies who use AI art in their projects either proudly proclaim they're using it, or actively hide it and get found out by someone and face significant backlash for hiding it. Labeling laws in the EU also make it extremely difficult to do the latter if you want to have an audience outside NA.
Is this an example of the bitter lesson? Could diffusion models not be this good with equal scaling/compute as LLMs?
Well the LLMs are much bigger than diffusion models so I think that's the bitter lesson. You could scale compute for diffusion models though.
Well no, the bitter lesson isn’t the scaling laws themselves. It’s that approaches which can take advantage of scaling laws will ultimately beat ones that can’t.
You learn something new every day! thanks
This is simultaneously incredibly impressive and annoyingly uncanny valley, since most of the landscapes are "ruined" by a cluster of churches right next to each other that is completely nonsensical.
This is in the same space as this creative exploration posted a few months ago -
---
Training AI to Paint with Code, March 2026
https://surya.website/rling-qwen-to-paint-with-code
HN Discussion:
https://news.ycombinator.com/item?id=49411800
Thesis presentation:
https://vimeo.com/1190839818
And I believe the HN user who created this is @kickingkeys
---
fascinating space to explore:
Instead of having AI directly generate an output, what happens if we ask AI to take a stab at the PROCESS of creating something.
Very interesting. I think a useful experiment would be to show AI a photograph, then ask it to paint it or draw it.
https://x.com/iam_zachi/status/2095992132620136677
This is such a good idea. Idk how to make this a benchmark, but it should be (maybe elo?).
Should be a Twtich stream tbh
The idea of using an LLM to drive graphic output is pretty popular, so I definitely wouldn’t be surprised if there are already several benchmarks out there already.
I've seen a few voxel-based benchmarks built around the same idea, with LLMs effectively constructing models using a discrete set of instructions.
https://minebench.ai
I love the canvases, and truly would like to have a bit more inside on the process this specific author used to reach the results shown. I have read Surya's "Training AI to Paint with Code" but, what's Alice's technique?
I’m a bit behind the times…
How does Opus actually paint? Thought it only generated text…
Think of it like the old turtle graphics feature in the LOGO programming language.
You give an LLM a virtual canvas and a set of “instructions” to control the pen.
https://en.wikipedia.org/wiki/Turtle_graphics
Another idea to make Claude generate images:
Ask generative AI to create an embedded web page of a voxel Mona Lisa head. It will generate an image.
Ask it to sample that image and extract out of it as binary that is well formed in the PNG data format...
EDIT: https://claude.ai/artifact/HUUtPMp7aF8iyjimViBSor - generates the Mona Lisa as a PNG via Claude sonnet 5.5 medium level. It's not a very good Mona in my opinion. Fire it up, rotate the 3d voxel presentation, scroll down, click the button, scroll further down to see the PNG. Here is the chat with prompts: https://claude.ai/share/2fb3126e-e97a-4a1a-9571-51d16f054330
Half of the industry is, going by the amount of "AI doesn't amount to anything and is just regurgitating text".
Two things:
- For the past year or more, many models can emit images directly, and all models that matter can see images directly - that's what "multimodal" means. Tokens don't have much to do with textual language anymore, they're more like units of sensory experience.
- Even restricted to text, a language model can operate anything that can be expressed as text, as long as you have a translation layer between textual representation and the final form. That includes giving commands as text. The total addressable space of what models can be used for is, thus, approximately anything humans do.
How are images translating to paintings?
It generate commands and coordinates to move the mouse and use the paint tools.
I find it unlikely that there's any direct training data for this. I'm talking about direct training data for using brush strokes like humans to construct an image.
Correct me if I'm wrong but this shows that painting capability is emergent, arising from unrelated training data thus very very compelling evidence that LLMs are actually intelligent.
I'm also quite surprised to see that LLMs can do this. I guess it is possible that "make an image in MS paint" is a type of RL environment used for image understanding. This is one of those areas where people inside the labs have a very different view into how much models are generalizing.
What capability makes this possible? It’s incredible! Does it have an image-generation model?
No this is just it moving the mouse and viewing the results with it's vision capabilities.
Now combine this with a robot able to move a paint brush and mix on a palette, and you could make some pretty awesome real-paint artwork
Would be a good project for an afterschool club. Even using a multi color pen plotter + webcam would be pretty neat.
Bob ROS
This is an S-tier comment. Thank you for your service.
I don’t like how it’s a human aiming to portray the ai generation with their taste and judgements about it but then becoming lazy and quitting on that and leaving in the ai generated portrayal about its own judgment in the site.
Just share your opinions, dude, we get its ai generated but share your own taste. We want to know what YOU think. Stop being shy and lazy.
What a misanthropic thing to make. I'm embarrassed for you.