Login

Recent searches

No recent searches

Search options

Only available when logged in.
social.coop is one of the many independent Mastodon servers you can use to participate in the fediverse.
A Fediverse instance for people interested in cooperative and collective projects. If you are interested in joining our community, please apply at https://join.social.coop/registration-form.html.

Administered by:

Server stats:

479
active users

social.coop: AboutStatusProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.3.4

#googlegemini

6 posts5 participants0 posts today
Public
Karlheinz Agsteiner

One more AI experiment this weekend.

Had it build a puzzle game that reminds a bit of Tetris.

Fed several pages of detail spec into the AI (Google Gemini 2.5 pro), got a working Phaser/JS game after the second "now build that part that you marked as TODO" prompt. Fixed some misunderstandings fast. Nearly despaired in the seemingly trivial rewrite of a high-score scene into a highest-level-reached scene. Claude 3.7 did a bit better but not much.

My conclusion:
- generating small apps (the limit increases with every LLM version) works near perfect by now.
- refactoring is still an issue, particularly when it comes to UI design / layout / styling.
- Without dev skills, it is very hard to create an app that does what you want.

#ai#googlegemini#phaser
Public *
Karlheinz Agsteiner

TL;DR: I used Google Gemini 2.5 Pro to create what might be the strongest (and perhaps only) app for the Namibian board game Hus. Gemini wrote ~99% of the code, cutting development time from weeks to hours, despite minor issues with specific rules and refactoring. My focus shifted from code quality to basic checks. Early analysis suggests player 2 has a winning advantage in the standard Hus rules on a 8x4 board.

With the (massive) help of Google Gemini 2.5 pro (experimental) I present: the probably strongest Hus playing app on this planet. Hus is a board game stemming from Namibia.

(okay, seems to be the only one, according to googling, and Gemini took a single prompt to implement a basic Alpha-Beta tree search, that works very well for a game where players have something like 10 moves, and complete information).

agsteiner.neocities.org/Kalaha

Experience from working with Gemini on this:

- it could code 99%, making virtually no bugs. I changed only about 10 lines of code of >1000 lines of Javascript
- we did have a longish dispute because it wouldn't want to understand how a game where both players move their stones counterclockwise works.
- Creating a second flavor of the game that works a bit differently didn't work so well, so refactoring is still not its strength.
- Overall creating four independent little apps (one playing with one ruleset, the other with a slightly different ruleset, and two simulations to do fast engine-vs-engine match to get strong game logs) took me 2-3 hours, instead of 2-3 weeks that hand coding would have taken.
- I notice that I am moving away from trying to understand the code, and wanting beautiful code to something like "developer due dilligence" (did a hacker compromise the AI and code a virus into my app), and not caring too much about how the code looks.

Also from a lengthy 12-ply self-play it seems that this variation of the game is won for player 2. First evaluation is "player 2 wins 13 stones", which is pretty decisive already.

The original variation where you don't capture stones on the row behind the inner one gives more balanced games.

Screenshot of the app that the AI wrote based on my specification. I'm losing, again.
#hus#kalaha#mancala
ExploreLive feeds
About