I vibe coded a tiny MUD-style world sim where LLMs control each character. It's basically a little toy sandbox where LLMs can play around. There's no real goal to this, I just thought that it would be fun, like a more advanced tamagochi.

One of the issues I encountered initially was that the LLMs were repeating a small set of actions and never trying some of the more experimental actions. With a bit of prompt tweaking I was able to get them to branch out a bit, but it still feels like there's a lot of room for improvement on that front. I still haven't figured out how to instill a creative spark for exploration through my prompting skills.

It has been quite exciting to see how quickly a few simple rules can lead to emergent storytelling. One of the actions I added was the ability for the agents to pray to the creator of their world (i.e. me) along with the ability for me to respond in a separate cycle. The first prayer I received was from an agent that decided to wade into a river and kneel, just to offer a moment in stillness. Imagining it is still making me smile.

Unfortunately, I don't have access to enough compute to run a bigger experiment, but I think it would be really interesting to create lots of seed worlds / codebases which exist in a loop. With the twist being that after each cycle the agents can all suggest changes to their world. This would've previously been quite difficult, but I think it could be viable with current agentic programming capabilities. I wonder what a world with different LLM distributions would look like after a few iterations. What kind of worlds would Gemini, Claude, Grok, or ChatGPT create? And what if they're all put in the same world, which ones become the dominant force?

Working on Book Reading Habit: https://bookreadinghabit.com

Built it because I wanted to read more, but most reading apps either feel too passive or turn everything into social noise. What worked better for me was making reading easy to start: short 5–10 min sessions, pick up where you left off, minimal friction.

So the app is basically centered around habit formation, with stuff like notes, progress tracking, session extension, shelves, and simple organization.

I care a lot about keeping it quiet: no ads, no feed, no unnecessary clutter.

Still early. Mostly trying to understand what actually helps people read more consistently.

Currently only available for iOS, but might release an android version in the future.

https://apps.apple.com/us/app/book-reading-habit/id674291326...

I'm writing an essay where I get into how I use GNU Emacs along with gptel (a simple LLM client for Emacs) and Google's Gemini-3 family of models to turn a 1970s-vintage text editor into a futuristic language-learning platform to help me study Latin. I want to show how I liberate poorly aligned, pixelated PDF image scans of century-old Latin textbooks from the Internet Archive and transform them into glorious Org mode documents while preserving important typographic details, nicely formatted tables, and some semantic document metadata. I also want to outline how to integrate a local lemmatizer and dictionary to quickly perform Latin-to-English lookups, and how to send whole sentences to Gemini for a detailed morphological and grammatical breakdown.

I also intend to dig into how to integrate Emacs with tools such as yt-dlp and patreon-dl to grab Latin-language audio content from the Internet, transcode the audio with ffmpeg, load it into the LLM's context window, and send it off for transcription. If the essay isn't already too long, I'll demonstrate how to gather forced-alignment data using local models such as wav2vec2-latin so I can play audio snippets of Latin texts directly from a transcription buffer in Emacs. Lastly, I want show how to leverage Gemini to automatically create multimedia flash cards in Org mode using the anki-editor Emacs minor mode for sentence mining.

I am continuing with my proofreading and language learning efforts and have been working on tooling for it.

= Proofreading =

https://github.com/adhyeta-org-in/adhyeta-tools

provides image extraction from PDF, OCR as well as a basic but nice proofreading web-ui.

Qwen 3/3.5 is good enough for OCR on books in Indic scripts. So that is what I am using. But you can configure the model that you want to use.

I may add a tesseract back end as well if necessary.

= Language Learning =

I have tried a few parallel text readers and was not satisfied by any of them. My website (https://www.adhyeta.org.in/) had a simple baked-in interface that I deleted soon after I developed it. However, this weekend, I sat down with Claude and designed one to my liking. I also ported the theming and other goodies from the website to this local reader. This will serve as a test bed for the Reader on the website itself.

LLMs now produce wonderful translations for most works. You can take an old Bengali book, have Claude/Gemini OCR a few pages and then also have it translate the content to English/Sanskrit. Then load it into the Reader and you are good to go!

The Reader, I will release this month. Claude is nice, but I do not like the way it writes code. It often misses edge cases and even some basic things and I have to remind it to do that. So I want to refactor/rearrange some stuff and test the functionality end-to end before I put it online.

https://mealplannr.io The end game is no/low touch weekly meal plans send directly to your inbox, with meals you love to cook but with none of the hassle around planning the meals, shopping list etc (which I spend hours doing every week).

An important feature for me was improving the recipe discovery experience, you can build a cookbook from chefs you follow on socials (youtube for now), or import from any source (Web, or take pic of cookbook etc) - it then has tight / easy integration into recipe lists.

Utilising GenAI to auto extract recipes, manage conversions, merge/categorise shopping lists etc - as-well as the actual recommendations engine.

If anyone is interested in beta testing / wants to have a chat I'll look out for replies, or message mealplannr@tomyeoman.dev

I made an idle version of the 1999 MMORPG "EverQuest". There's maybe around 50 people playing at any given time and has a enthusiastic discord group for it. It's relatively fully-featured to the original game, and has a lot of new mechanics to make the idle format work well. The 3D graphics aspect of it is really more of a screensaver, though, and all game interactions are done through menus.

I recently converted a bunch of stuff to be client side instead of server side (turns out running a real-time MMORPG server is expensive) so there's a new round of bugs I'm still resolving, but it's still fun to play:

https://www.idlequest.net/

A5 - a pentagonal geospatial index: https://a5geo.org

If you’ve used H3 the semantics should be familiar. The biggest differentiator is the fact that cells have exactly the same area globally, for why this matters see: https://a5geo.org/docs/recipes/a5-vs-h3

Since starting the project last year and providing implementations in TypeScript, Python and Rust it’s been great to see a community grow, porting or integrating into DuckDB, QGIS and many more: https://a5geo.org/docs/ecosystem

Oh, so many things. I guess that’s both the blessing and the curse of agentic ai today.

The most fun is a simple Claude Code in a loop, Boucle, which builds and iterates on its own framework[0][1].

The first thing it built was a persistent memory. Now it has finally built itself a "self-observation engine" after countless nudging attempts. Exploring, probing, and trying to push back the limits of these models is pure chaos, immensely frustrating, but also fun.

Aside from that, some sort of agent harness I guess we call them? Putting together a "system" / "process" with automated reviews to both steer agents, ground them (drift is a huge pain), and somehow ensure consistency while giving them enough leeway to exploit their full capabilities. Nothing ready to share yet, but I feel that without it I’ll just keep teetering on the edge of burnout.

[0]: https://github.com/bande-a-bonnot/boucle-framework

[1]: https://blog.boucle.sh

For language geeks: https://kpt.datamediate.com

KPT is a language app specifically targeted at explainable verb conjugation for highly inflected/agglutinative languages. Currently works for Finnish, Ukrainian, Welsh, Turkish and Tamil.

These are really hard languages to learn for most speakers of European languages, particularly English - we're not used to complex verb conjugations, they're hard to memorise and the rules often feel quite arbitrary. Every other conjugation practice app just tells you right/wrong with no explanation, which doesn't really help you learn when there are literally hundreds of rules to get right.

The interesting part was using an LLM to create a complete machine-executable set of conjugation rules, which are optimized for human explainability, and an engine to diagnose which rule is at fault when you get it wrong. There's several hundred rules needed for each language in order to cover all exceptions.

NB as a bonus it also works fully offline because my best practice hours are when I'm travelling and have poor connectivity.

Still on time. It is almost two months now and this is such a deep subject and there are so many little tricky bits that I wonder if I will be able to complete the thing but there is still (slow) progress. I never suspected the amount of hard work that goes into building something that is stable at the nano second scale. But I'm becoming more appreciative every day ;)