MICA Playground
Learned integer rules in a cellular automaton. Try Flame-W for sentences and Ember v0.3a for word suggestions, then rate the results.
0 / 300
Prompts, outputs and ratings are saved and may be used to train MICA. No names or IP addresses are stored with them, but please don't type personal information.
This session
The models
Flame-W 0.3.1 (4 Oct 2026)
Predicts whole words (16,384-word vocabulary). An integer memory lets it use the last 64 words instead of about 4; 0.3.1 retrained it on 7 times more text. It scores 31.25% on Tiny Theory-of-Mind (chance 25%), above 6 of the 36 models on that leaderboard.
Predicts whole words (16,384-word vocabulary). An integer memory lets it use the last 64 words instead of about 4; 0.3.1 retrained it on 7 times more text. It scores 31.25% on Tiny Theory-of-Mind (chance 25%), above 6 of the 36 models on that leaderboard.
Ember v0.3a (4 Oct 2026)
The original 4.5 MB byte-level cellular model now has 1.23 MB of integer word readouts. It suggests a next word or completes one after two typed letters. On independent clean prompts, next-word top-1 improved from 6.75% to 14.25% versus v0.2A. Its cellular byte loss is unchanged.
The original 4.5 MB byte-level cellular model now has 1.23 MB of integer word readouts. It suggests a next word or completes one after two typed letters. On independent clean prompts, next-word top-1 improved from 6.75% to 14.25% versus v0.2A. Its cellular byte loss is unchanged.
These are early research models. Flame-W's sentences often don't make sense yet; Ember's word suggestions have a limited vocabulary. Your ratings and corrections help evaluate them.