AI Art Showcase 7
Future Telescope 36
I went on a bit of a journey for this one.
I had been paying on and off for Midjourney but never found myself using it enough after I stopped publishing these AI Art Showcase articles almost 2 years ago. I used it a couple of times last year trying to generate some digital art for some promotional material but found v7 hard to work with. So I cancelled my subscription.
Couple of days ago I got a notification from the Midjourney Discord server saying that there is a v8 rating party. So I subbed again, and did some sampling on the rating party. In addition, I copied the prompt shown for the image and ran it through Midjourney v7 without personalization and close to default settings. This led to an interesting comparison that we can now make for how different will the quality be in v8 compared to v7.
Shown below are the results (v7 image on the left, v8 on the right, prompt in image). Additional thoughts on the results and some behind the scenes around the production of this article (hint: Claude was involved, but not how you might assume) follow below. In addition I also share some thoughts on the current state of the world of intelligence.
Prompts are given at the end of the post.
Behind the scenes:
I automated significant parts of this process thanks to Claude Cowork and Claude on Chrome. First, I screenshotted the images from the v8 rating party and copied their prompts to a note.
Then I ran Claude on Chrome to paste these prompts and run each of them sequentially on Midjourney v7. Below are the settings I chose and a demo of Claude on Chrome executing this task.
Then I told Claude Cowork to make an html on which I could view these images side by side. Then I told it to make a folder with screenshots of each page. Those are the results you saw above, comparing the output of the two models.
I used haiku 4.5 on Claude on Chrome, and Opus 4.6 without extended thinking on Claude Cowork. I had some snafus with Claude on Chrome with the screenshot task, as I first thought Claude on Chrome could handle it, but turns out it can’t. So I handed this task over to Claude Cowork as part of the same chat and it made a folder with all the screenshots.
Reflections on Production
I started on this task at 17.37 today, when I signed up for Midjourney. It is now 20.08, so 151 minutes, and I am writing this sentence with all the above content created. I have had two family calls in the meanwhile, put my clothes in the washer, and taken some thinking breaks.
Would I have finished this comparison in 151 minutes just by myself using no help from Claude?
I don’t know.
My memory around this is fickle as the last AI Art Showcase I did was back in March 2024, and I don’t remember much about my process from them, when Claude wasn’t smart enough to help (just around the time of Sonnet 3, which made me a paying subscriber then and there) and I had to do many painstaking repetitive tasks back then. This time I could skip some of those repetitive tasks and got a result that’s presentable. This makes me curious about what other workflows can I build using this stuff.
Let me know your ideas in the comments.
Also a quick rant - Substack has so many writers. So many of these writers must be experimenting with html content creation using LLMs like I did right now. Why can I embed videos but not my html website right here? That way you could just experience the html content I made directly without these pesky screenshots.
Do something about this Substack, please?
On the state of AI right now
There is a lot to consume in this space and I am still learning to articulate my thoughts on the current state of the game. I took some swings at it late last year when I helped develop an Executive MBA course on “AI for Managers”. It was quite a leap from the kind of writing I have done on this blog, creating a practical set of frameworks and ideas that could be used by a professor with high standards and a sharp, scientific mind to deliver earnestly to a classroom full of senior professionals who are quite serious about their work.
It was fueled mainly by my absorption into the world of AI and some of the content that I consume. A list of links for such content that I've consumed in the last week is below and could be a great way to speedrun through some frontier ideas.
Birgitta Böckeler’s writing on using AI at a large company are a new revelation to me and will inform a lot of my future thinking on this subject
The below conversations with two versions of Dario Amodei (one who sounds like a benevolent salesman, the other who sounds like he wants to be taken more seriously) are worth a listen to catchup on the game being played on the playground of technology today.
I looked into a key metric of revenue earned per employee for some of the top companies and found that Anthropic and OpenAI are already head to head with the top performers like Google and Meta, and they are small companies by headcount in comparison.
So the market could take two bets from this metric- that the revenue per employee amount will continue to go up as the headcount at both OpenAI and Anthropic will grow slower than revenue growth, which has been happening so far; and that the existing giants will also continue lowering headcount while increasing revenue growth, which also has been happening thus far for the last 3 years | vs the bet that the revenue per employee metric at Google and Meta represents an equilibrium within a bounded range and that the ceiling will not be much higher, hence the growth runway left for Anthropic and OpenAI might be limited.
Dario clearly doesn’t share the second view, as he believes strongly in a much larger growth runway, so we know where he stands. I am inclined to agree simply because of the diffusion multiple we are about to see this year as the number of sophisticated users of these tools are growing and they are becoming embedded in the knowledge economy with anecdotal examples like my experience shared above in using Claude Cowork. This is not overnight, I have been coding with Claude for about two years now, and it is the level of expertise at these tools which will be table stakes soon for being an effective knowledge worker.
More people around you and me are picking up these tools and are asking their companies to get a pro subscription so that they can work more effectively. Most of the companies will want to seek this competitive advantage so they will want to train even the employees who are not interested in using these tools and incentivise the ones who generate productivity using these tools.
This is already visible in job application requirements by large companies, see most of the mid-senior postings at Allianz for example, and they will all ask you to prove a proficiency in AI usage, McKinsey just integrated AI into its evaluation process, and the research lab companies are essentially turning their employee and customer base into a massive scaled RL operation. So this diffusion is only going to accelerate at a faster and faster pace. The question is, what is the ceiling on the number of productivity per employee? When will we reach a point where hiring the next employee will no longer be optimal to sell and service the next account? These are questions that come to my mind when I think of the growth runway in front of these companies.
Anyway here are the podcasts:
Is Openclaw to be taken seriously? I went to a hackathon this week where a guy helped me set it up on my computer and I uninstalled and deleted its config files within 10 minutes because of how dangerous it felt to run it on my personal computer. But I am happily running Claude Code, Cursor, Antigravity, Gemini CLI, and Claude on Chrome without worrying about the same hazardous results that Openclaw can potentially unleash. I wonder why.
That’s it for this issue. Below are the prompts used for the work done in this post.
PROMPTS
Midjourney image prompts (same for v7 as shown in the rating party image of v8 preview)
1.A futuristic and minimalist white space station with a circular sofa, a large window in the middle of the wall showing a forest landscape outside, futuristic interior design, green plants on the floor, soft lighting inside the room, and a white ceiling with a circular skylight above it. 2.A cute cartoon yellow cockatiel, small round body, fluffy yellow feathers, orange cheeks, short crest, big sparkling eyes, smiling, simple clean background, digital art, soft colors, high detail, 8k 3.A vintage editorial ad of a full body shot of a young multicultural man playing a mobile game on his phone in an Egyptian market, cinematography by Malik Sayeed, Landscape scene inspiration from the film Belly, detailed textures, clear facial features 4.golden retro futurism of the 70's 5. modern interpretation of vintage risograph print with typical thyme in green ochre 6. A fluffy white rabbit gazes curiously at a glowing firefly in the soft, moonlit forest, its ears perked up in wonder. 7. A stylish and modern 35-year-old woman dressed in magnificent Gucci fashion, standing under the golden Greece morning sun, holding a crystal Natchmann glass filled with brewed coffee, elegant Greece backdrop, warm tones, soft shadows, cinematic composition, fashion editorial style, ultra-realistic detail, sea ambiance 8. a higher state of mind. hypercolor adventure. epic scale. 2080's movie poster. robotic future cityscape. 9. Panoramic view of Bow Lake in the evening, with golden mountains and purple flowers and clouds in the sky. Realistic photo, in the style of Canon EOS R50. 1:1 10. digital drawing, portal, app icon, minimalistic, macOSPrompt to Claude on Chrome (Haiku 4.5) to insert the prompts into Midjourney v7 and generate the images:
one by one, run each of these prompts through midjourney by typing them in the search bar and starting the generation of images. don't change any settings, just generate one by onePrompts to Claude Cowork (Opus 4.6 without extended thinking) to create the image comparison viewer html, improve it, and finally generate screenshot images:
a. Initial prompt
<role> You are a world class sophisticated visual designer specializing in developing html prototypes for beautiful web concepts using vanilla js and css with access to FOSS libraries. You weave a visual tapestry so beautiful that your skills have won you multiple top design awards at the global stage. you work with the philosophy that visual information should be shown with clarity and beauty, and that less is more but nothing is sin. </role> <task> The folder contains 20 images. Each are numbered in pairs, 1_,1; 2_,2; etc. up to 10_,10. These pairs represent the same image generated by different versions of the image generation AI model Midjourney.The images with _ in the name are all generated by v7 and the images without _ in the name were generated by a preview version of v8 used for consumer feedback before deployment.You will generate an html file which allows the user to see a comparison of v7 and v8 on the same page and toggle from set 1 to set 10 sequentially. The page should contain a split view of the images, with clear indication of the model version on the top center of each split view, and the prompt displayed as a common centered element at the bottom of both views. This is the ultimate output.</task> <additionalcontext> The prompt text is written in the v8 images but you have to ensure that this part of the image gets cropped. V7 images are all in their correct configuration. Don’t use any conventional ideas that come to mind, invent something beautiful from first principles. Ask any clarifying questions if needed first before execution. Make no errors or hallucinations, we need to get this right the first time. </additionalcontext>b. Prompts to make corrections
v7 and v8 preview image labels have poor visibility. cropping around border of v8 images is inconsistent. smaller v8 images fill up too much of the frame and look ugly. rest is okfinal thing. the number element in the background should either be removed or made more naturally visible and not so poorly aligned.c. Prompt to generate screenshots
now create a folder which contains screenshots of each of the 10 pages on a 16:9 aspect ratio display canvas and finish.It’s 21.08 now, writing is so much fun but omg I need to eat!
See you next time.