tarded.aiINTENTIONALLY IMPERFECT. NOW IN EMOJI.
Language was
a bottleneck.
We removed most of it.
Our model is an experiment in emoji-native conversation and reasoning. Test its emoji replies with EmojiLM, an emoji-to-English model by Letian Peng and collaborators. Our ambition remains disproportionate. We call that strategic verbal divestment.
CONVERSATION V1 / 362.6 MILLION PARAMETERS
A conversation. Finally.
Type an emoji message and our trained model generates a reply from the conversation so far. Try preferences, changing your mind, or making a simple plan. It is an experimental conversational pilot: it can misunderstand you, ignore a correction, or wander off-topic.
Autoregressive emoji replies using conversation history. This is not a separate trained deliberation engine.
Inference parameters
Personality trait vectors · expression-layer preview
The existing 353-parameter expression ranker selects among eight authored styles. These sliders affect this preview only, not the conversation decoder or its judgments. Context eligibility is explicitly selected below, not inferred from your chat. Sarcasm, formality and assertiveness are not implemented.
Loading expression ranker…
Select a candidate set and run a message.
Only the current message is scored. This result is not added to conversation history.
These are the cached Unicode CLDR names and keywords used by our multilingual vocabulary-initialization pipeline and semantic research. They are source associations, not a reading of model weights, a complete meaning inventory, or proof that the deployed checkpoint retained each word. The original warm-start does not include per-row word receipts. No slang is invented to fill gaps.
Source versions, hashes & coverage · Unicode licenseYour words. Our words.
Each message stays beside its own interpretation. One shared scroll.
Your move.
Preferably in emoji.
Start with 👋 or tell it what you like: 🍕 ❤️. Interpretations will line up alongside each message.
Full emoji picker
Loading the model vocabulary…
All 5,244 model emoji, in Unicode category order, plus non-emoji Dingbats and Ornamental Dingbats and runtime brackets. Gray symbols are unsupported or reserved. Some newer glyphs need a newer font.
Logic & control symbols
Insert supported conversation symbols below. Labels describe common uses, not guaranteed commands. Reserved framing is added by the server; formal operators belong in the explicit logic sandbox ↗.
Sending a message sends this conversation to our inference server. We do not save chat content or use it for training. History stays in this tab until reset or reload; hosting records ordinary request metadata. Each message is also sent separately to our server for EmojiLM interpretation when enabled.
Connecting to the conversation model…
Emoji only; English text is not supported. History limit: 1,024 native tokens. Replies: up to 24 tokens. No hidden summarization—start a new conversation when full. First connection may take longer while the server starts.
Replies come directly from a Qwen-derived causal Transformer adapted to emoji dialogue. No scripted answers, appraisal classifier, or semantic repair chooses the reply. The pilot learned from a small synthetic curriculum and has no independent human evaluation; broader conversations remain unreliable. Its output does not demonstrate language-independent internal reasoning.
THIRD-PARTY MODEL / BROWSER DEMO
Test with EmojiLM
An emoji-to-English model by Letian Peng and collaborators. We use their model here to interpret emoji sequences. This independent demo is not affiliated with or endorsed by its authors. View the original model ↗
External research. Internally inflated expectations.
An English interpretation will appear here.
Generated interpretations can add details not present in the input.
EmojiLM runs on our GPU server. Your input is sent for interpretation and is not saved or used for training. Both models share a bounded queue; busy requests may need a retry. No model download is required.
HUMAN SIGNAL / NOT AN ANSWER KEY
Our model receives the conversation on our inference server to generate each reply. Chat content is not saved by the application unless you explicitly save or share a feedback example. Reset clears your tab’s history. EmojiLM interprets individual messages on our server when enabled. Requests are held briefly in memory while queued and processed. Hosting receives ordinary request metadata.
The conversational pilot generates native emoji replies from its actual dialogue history. It has narrow training coverage and can miss negation, repeat a rejected idea, or make irrelevant associations. The separate logic sandbox follows explicit rules. EmojiLM is a third-party emoji-to-English model by Letian Peng and collaborators; we use it here to interpret emoji replies, with no affiliation or endorsement.
An emoji vocabulary constrains expression and makes context do more work. We believe our product strategy should leave room for ambiguity. Fortunately, we have declined to define it.