Koen van Gilst 🇺🇦🇪🇺

A new look for the Game of Life classic.

- Demo: stacked-game-of-life.koenvangilst.nl
- Source: github.com/vnglst/stack...

It feels like a new AI model drops every day lately. Today it was Gemini 3.1 Pro. It does a nice SVG rendition of modern art, I particularly like the Sunflowers. It's a clear improvement from older versions and other models.

pelican.koenvangilst.nl/gallery/cate...

Not as much as I hoped. One thing I learned is that LLMs are not as easy to trick as I thought.

What worked particularly well here was a test-driven (TDD) approach. I asked Claude Code to focus only on the algorithm for selecting the next Tetromino (the Tetris blocks), starting with tests. After that, things went smoothly, though it still took many follow-up prompts (around 100).

The year is coming to an end, and looking back I had a lot of fun building quirky and insightful visualizations. Agentic coding agents have been hugely empowering for me. In 2024 it was “just” autocomplete on steroids. Now it feels more like conducting directing the agents to go where I please.

Tetris Time - A Clock Made of Falling Tetrominoes Watch the time come alive as Tetris pieces fall into place. A mesmerizing clock where tetrominoes build the current time, one block at a time.

After New Year's it can also function as a regular clock:

tetris-time.koenvangilst.nl

It's almost New Year's and here's an original way to show the countdown using Tetris. I'm sure something similar has been done before, but I had a lot of fun building this (TDD style) using Claude Code.

tetris-time.koenvangilst.nl?mode=countdo...

However, I am a bit concerned about the sending limits. What if one of these alerts accidentally starts sending too many emails. Will I be locked out of my private, paid email account?

Dear @proton.me, I‘m really excited about the new SMTP feature you recently released! I‘m using it for alerting when I get traffic spikes or low disk space warnings on my private servers.

After visiting Denmark twice this year I wanted to visualize how different their land use is compared to The Netherlands. Turns out it's pretty similar, but less urban area and more forests.

onsland.koenvangilst.nl?country=denm...

Just added Medusa by Caravaggio. Once again GPT-5 seems to offer the most creative interpretations. Maybe they've been doing some finetuning on SVG generation?

The website also comes with a workshop feature which lets me explore new artworks and/or regenerate existing ones if they fail. It gets a list of models from OpenRouter and shows the costs per output token.

Anyone with an OpenRouter API key can start the project locally and use the workshop.

The gallery has some interesting reinterpretations of famous art. For instance, here's how GPT-5 interprets The Scream by Edward Munch. I love how the colored horizontal stripes go outside the boundaries of the painting.

Pelican Art Gallery is a new side-project of mine where you can see how LLMs generate SVGs of famous artworks.

It's inspired by @simonwillison.net's practice to benchmark LLMs by asking them to generate an SVG of a pelican on a bicycle, hence the title.

pelican.koenvangilst.nl

Just released an update to my UN speeches side project.

Added speech similarity analysis using embeddings. The heatmap highlights countries with semantically similar speeches (darker = more similar). Click a country to see details.

un.koenvangilst.nl/analysis

Like many others, I've been building side projects with Claude + Copilot agent - it's so fast & almost too easy. Today I tried a no-code AI tool. Endless clicking, docs, configs. Slower in every way.

That made me wonder: if coding agents are this good, does low-code still have a future?

I just got back from a lovely week at Normandy's coast. The cliffs there are stunning and capture light in remarkable ways. I'm trying to get back into photography after a long pause and have created a dedicated space for my photos on my website.​​​​​​​​​​​​​​​​

koenvangilst.nl/photography

To test out Google's new Gemini 2.5 Pro model I asked it to build my Shepherd's Dog game. Not bad, but Claude 3.7 still offers the best one-shot result.

Play it here: vnglst.github.io/when-ai-fail...

See the rest: github.com/vnglst/when-...

Choose Life.

choose-life.koenvangilst.nl

Inspired by Trainspotting and @ginatrapani.org's Truisms.

o3-mini's version has much better game mechanics and is more fun to play. But it misses a lot of the features.

html-preview.github.io?url=https://...

Note: it doesn't work on Safari.

You can play the game here: html-preview.github.io?url=https://...

Results of other models and the leaderboard can be found here: github.com/vnglst/when-...

Shepherd's Dog is a game I've wanted to create for a long time, but I never got the sheep flocking behaviour right. The goal of the game is to herd all the sheep into the pen before nightfall.

Now, I've asked several models to create this game and I'm impressed with what Claude 3.7 come up with.

I recently read Ian Kershaw’s biography of Hitler. This passage, where he bullies the Czechoslovak president into a “peace deal,” stood out. After what happened in the Oval Office today, it feels disturbingly relevant. But I’m also glad Zelensky is nothing like Hácha.

Can LLMs play the card game Set?

While the traditional models struggled, the reasoning models performed surprisingly well.

github.com/vnglst/when-...

New update for the UN speeches project: I made the interaction clearer by adding hover effects and borders.

The project tries to simplify the UN speech content as much as possible to see who's talking about whom in either an optimistic or pessimistic way.

un-speeches.koenvangilst.nl

Then it extracts these mentions using quotations from the speech to explain why it thinks the mention is optimistic or pessimistic. It also makes sure to use the correct ISO country code.

All this in maybe 100 lines of code and without any training data. Pretty powerful!

GenAI is doing a lot of heavy lifting here: it’s reading the speech documents (which could be in all kinds of languages), it figures out if other countries are being mentioned in the speech. This could be direct or indirect mentions.

Thanks and great question!

I’m no machine learning expert but I think this would be hard if not impossible to pull off using just machine learning.

Future plans also include mapping speeches to UN Sustainable Development Goals (SDGs), revealing which nations prioritize specific global challenges.

Anyway, lots to explore, but looking forward to some initial feedback!