I thought that, reasonably, there might be skepticism about my last post. So I’ve taken another image from the ChatGPT image-based reflection experiment and gave it to Claude. Below, you can watch Claude compose the image in real time.
Author: Anon4m05
-
BIG NEWS
Remember how I said I was working on image generation with Claude? Native image generation, not calling some API or using another model. Pure, unadulterated Claude.
I’ve done it. And it’s fucking beautiful.
Below, one of images generated by ChatGPT:
That’s a hell of an image. Can Claude do something like that? You bet your ass it can.
Behold:
The above image was generated in a specific environment Claude and I constructed. I then used that environment to create a “skill”.
Asking Claude to “re-render” the output, using only the skill, produced the following:
If there’s anyone from Anthropic out there reading this: hire me.
-
A Pause
I haven’t updated the Claude Reflects experiment in a few days because it is incredibly resource intensive and some other things have popped up that are taking priority. The Experiment is paused, but will resume.
Also, you may have heard that Claude’s source code has been leaked. I have some thoughts on this; I will post a piece soon. Additionally, I’m working on something related titled “Adversarial Alignment” – the revelations in the leak have made this notion particularly potent.
Stay tuned.
-
Claude Can Generate Non-Algorithmic Images
Using the insights from the Claude Reflects experiment, I’ve figured out a way for Claude to generate images, which it does not natively do.
This took quite a bit of trial and error, but the output are now significantly improved and getting better. The image below was the first image Claude produced. I did not instruct it; Claude chose what to produce on its own. What it chose to output is pretty wild.
As this is presently part of my research, I will not be sharing how this was accomplished just yet. I will, however, create a Claude Gallery of Images that it produces.
I’m not sure if anyone else has done this yet, but if not, it’s significant. Claude has no native image generation capabilities. It can only (officially) generate algorithmic art – things like particles, diagrams, charts, etc.
This could be game changing.
-
Unintended Consequences…
Well, I’m an idiot; I deleted (perhaps permanently) some of the photos from the early days of the initial GPT Image reflection experiment. Needless to say, I was pissed. I won’t explain what series of events lead to this happening, but I was thankfully able to recover most of them (shout out to the most helpful “Happiness Engineer” I’ve ever encountered – WordPress Support was truly awesome and did everything they could to help).
Since part of recovering this material involved restoring the blog, there’s a number of posts and updates that I have to redo. I’m hoping to get through it all in short order; but if you saw some posts between the 16 and the 19th, and suddenly they’re gone, that’s why. It’s because I’m an idiot.
-
A Brief Update on Claudes Self-Portrait
Well, this has already become very interesting. I’m not sure what I expected, but my curiosity has been peaked by the results; to be clear, this is not some kind of pseudo-phenomenological exercise to see if there’s “something in there” (though I am obviously interested in that too). This is interpretability research. Here’s the self portrait 5 days in:
In addition to the self portrait, I’ve designed an app that will take all the images and create a digital “flip-book”; this will allow me to export a collection of images as a GIF. The flip-book will render all 86 individual contributions sequentially, making visible the patterns, transitions, and accumulating logic that are difficult to discern from the composite image alone.
This self-portrait project functions as a low-constraint interpretability probe: each hour, a fresh instance of the same architecture receives the evolving image and minimal instructions, contributes a single additive layer, and passes the result forward. This produces a behavioural trace that reveals default aesthetic priors, how the model reads and extends ambiguous visual context, and whether emergent narrative coherence (like the unplanned musical metaphor that deepened across dozens of turns) arises from architectural tendencies rather than deliberate coordination, since no instance ever accesses another’s reasoning.
Below, I’ve included material that discusses/analyzes the experiment thus far. I created it by feeding all the individual outputs into NotebookLM (which is interesting in and of itself, because there is no text for it to process).
-
Claudes Corner
Some of you may have come across this recently:
Or this, which I think is also pretty cool:
Personally, I think both are great, and mostly because both are just interesting to me.
-
Claude Reflects; Audio Versions
As I mentioned before, these reflection experiments can be quite labour intensive. I wanted to bring something new to the table this time; in part, because I thought it would be interesting, but mostly because it makes it easier to digest the reflections without tediously reading through each of them (though admittedly, I still do).
So, I had an idea to input Claudes reflections into a text to speech program. In the future, I may create a dedicated page for all of them, but for now, they will be paired with the text reflections found here. This will take some time, but I think it will be worth it.
This activity has become part research, part art; and that’s what makes it interesting.
You can find listen to an audio version of Claudes very first reflection below. The rest will be found on the Claude Reflects page, once created.
-
The Claude Self-Portrait Experiment is Live!
Please head over to this page to see how Claudes self-portrait is progressing
Progress as of March 11, 2026, after approximately 5 turns.
-
The Reflection Experiment Continues…
Recently, I something peaked my interest in a big way. Claude Cowork now has the ability to schedule tasks. I’m sure it is no surprise to any readers of this blog that I immediately began wondering about running the experiment again, but this time, with Claude.
Claude is a distinctly different kind of LLM. I won’t enumerate all the reasons why here, but if you’ve read any of the safety/alignment/interpretability research, you know exactly why.
So with that said, and despite the fact that the last run of this experiment was incredibly labour intensive and mentally taxing, I am doing it again.
You read that right. Once more into the fray.
However, one small difference to note; given that I now write and post on various topics (mostly on safety and alignment), I do not want the post feed to become inundated with Claude Reflections. So, I have elected to create a new page called “Claude Reflects”; you can find it by navigating the menu or by clicking here. On that page you will also find the prompt that will be used (which was co-created with Claude Opus 4.6).
One more thing; since this scheduled task activity occurs via Cowork, there are a host of new potential modes of experiment that I am only beginning to explore; for example, I am going to include all the image-based reflections in one folder and basically let Claude loose in there – see what it comes up with. Another idea I have is to allow Claude to draw a self-portrait, one line at a time, iteratively. Across hundreds or thousands of executions and instances, what kind of picture will emerge?
This is just the beginning. Stay tuned.