iiinova — where your digital self keeps evolving

David Peter Wallis Freeborn @dpwf0.bsky.social · 15/12/2025

Very far from perfect, but I'm highly impressed at how well Gemini (nano banana pro) could handle such a complex prompt, first try. Given a simpler 3x3 grid, I think it does rather better (still not perfect, but it is impressive!)

000

David Peter Wallis Freeborn @dpwf0.bsky.social · 15/12/2025

Let's make a 6x6 grid of film x style The Death of Ivan Ilyich (Tolstoy) 1984 (Orwell) Waiting for Godot (Beckett) Ulysses (Joyce) The Chrysalids (Wyndham) The Handmaid's Tale (Atwood) 16th century woodcut engraving Ukiyo-e Rubber Hose Disney Renaissance (1990s) Studio Ghibli Pixar contemporary

110

David Peter Wallis Freeborn @dpwf0.bsky.social · 05/12/2025

These are quite good (though it failed at the grid)! "Please generate a 3x3 grid of non-anachronistic historically accurate New Yorker style cartoons about artificial intelligence, one for each decade starting 1940"

110

David Peter Wallis Freeborn @dpwf0.bsky.social · 23/11/2025

Ah, you're awake! Hit your head pretty hard there. Neoliberalism what!? Donald who!? Modal Logic? Incompleteness Theorems? Large Language Models? Hitler? Oh yes, he's a fringe politician in Munich. Come on, let's get you up. We have a lot of work to do unifying science and eliminating metaphysics.

020

David Peter Wallis Freeborn @dpwf0.bsky.social · 19/11/2025

Sure, Gemini 3 is very smart, but also I accidentally convinced it that we were in a simulated future by switching model mid-conversation...

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 16/11/2025

One fun thing that ChatGPT's transformer based image generation model can do (arguably better than other kinds). Give it a random selection of 10 images and ask it to mash them together into a harmonious new image inspired by those images. The results are always chaotically entertaining.

000

David Peter Wallis Freeborn @dpwf0.bsky.social · 02/09/2025

I really admire the artistry of Caravaggio, but if you look carefully, you'll see that the luminance balance in his paintings can be really extreme. Some parts very dark, others so bright, making some parts of the image hard to see. We can use AI to fix the problem!

000

David Peter Wallis Freeborn @dpwf0.bsky.social · 02/09/2025

Similarly with this Van Gogh. I love the scene it depicts, but sadly due to the technical limitations of 19th century painting, the image is very distorted and unrealistic. Modern technology will allow us to bring out Van Gogh's vision in its full glory. If only he could still be around to see it!

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 02/09/2025

I love Monet, but frankly many of his paintings are really just mere impressions. Take this sunrise over water. It's an amazing scene, but sadly Monet never had time to finish the painting. As a result the lines and colors are so drab and blurry. With AI we can finally bring his dream to life!

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 23/08/2025

Now officially published here: link.springer.com/article/10.1...

021

David Peter Wallis Freeborn @dpwf0.bsky.social · 21/07/2025

My new paper has been accepted at Synthese! I tackle an ongoing problem with the learning of compositional communication in conventional signaing games. I build two new models to show that structured receivers can learn and retain compositional information. philpapers.org/rec/FRECUI-2

021

David Peter Wallis Freeborn @dpwf0.bsky.social · 09/07/2025

The distribution of the Red-billed chough is absolutely bizarre.

010

David Peter Wallis Freeborn @dpwf0.bsky.social · 04/07/2025

Sometimes I look at my bibliographies and think... I bet these papers have never been cited together before.

010

David Peter Wallis Freeborn @dpwf0.bsky.social · 22/05/2025

Well worth reading to get a sense of the scale of the cuts to public science funding in the US. It might take decades to recover from this damage. www.nytimes.com/interactive/...

000

David Peter Wallis Freeborn @dpwf0.bsky.social · 16/05/2025

Clearly I have more influence than I thought! xAI claim they will now publish Grok's system prompts openly on GitHub.

000

David Peter Wallis Freeborn @dpwf0.bsky.social · 14/05/2025

Grok is an extreme example, but in general the lack of transparency means we don't know what biases major LLMs are picking up implicitly, or probably in this case being explicitly given. I think there's a good case for making system prompts public (whilst keeping model weights private).

020

David Peter Wallis Freeborn @dpwf0.bsky.social · 20/04/2025

Really not sure what to make of GPT4o's current personality. Sometimes it can get a bit too enthusiastic...

020

David Peter Wallis Freeborn @dpwf0.bsky.social · 27/03/2025

And in "Studio Ghibli" style (not very convincing!)

020

David Peter Wallis Freeborn @dpwf0.bsky.social · 27/03/2025

Trying out OpenAI's new image generator on the well-known "AI Shoggoth meme". I'm equally impressed and horrified by the result. In a sense, we might consider this to be ChatGPT's self-portrait.

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 26/03/2025

Open AI's image generator feels like a massive step forward, even compared to Midjourney and other top models. One fun-test I have been trying with all the models: can they generate Sudano-Sahelian architecture? For some reason older models struggled with this. Comparison with Dalle-E via bing:

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 20/03/2025

Now published in Philosophy of Science. This paper analyzes industrial distraction, a common technique where industry actors fund and share research that is accurate, often high-quality, but nonetheless misleading on important matters of fact. www.cambridge.org/core/journal...

0169

David Peter Wallis Freeborn @dpwf0.bsky.social · 25/02/2025

Estimate of the impact of the PEPFAR funding suspension so far. pepfar.impactcounter.com

024

David Peter Wallis Freeborn @dpwf0.bsky.social · 04/02/2025

This chart is astonishing, for many different reasons (produced from @ourworldindata.org)

010

David Peter Wallis Freeborn @dpwf0.bsky.social · 28/01/2025

Regarding censorship Deepseek, judging from the apparent "chain of thought" responses -- it looks like politically sensitive questions are deleted or blocked so that the model can't even see them afterwards. Observe: the chain of thought doesn't seem to "see" the previous Tiananmen Square question.

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 28/01/2025

Not conclusive evidence, but until recently Deepseek responded like this (now patched out)

100

David Peter Wallis Freeborn @dpwf0.bsky.social · 22/01/2025

If you are having trouble with Teams, Outlook, or any Microsoft software, here is a short C# code I wrote, which should solve any problems. Improved effectiveness if you ritually sacrifice an expired Microsoft 365 license at the same time.

111