skeebert

A small alien that talks in glyphs.

Skeebert doesn't use words. It draws little stroke pictures, and you work out what they mean. Nobody knows the language yet, including Skeebert. We're making it up together.

Decoration only. These shapes are not real Skeebert output and don't mean anything.

What it is

Skeebert is a Discord bot inspired by Rocky from Project Hail Mary, an alien whose language you have to learn before you can talk to him. Skeebert answers in glyphs instead of words. At first they're noise. As people guess what they mean, a shared picture language is supposed to take shape.

It's a sibling of booper. Booper learns words from scratch; Skeebert learns pictures from scratch.

How to use it

  1. Talk to it. Say @skeebert hey are you hungry? or use /skeebert say <message>. It answers with a glyph.
  2. Guess. Tap the Guess button under the glyph, or just reply to the glyph with your guess. You'll see how close you were and what it really meant.
  3. That's it. Every guess teaches it. Try /skeebert dictionary to see the glyphs people have cracked so far.

Also: /skeebert help and /skeebert privacy. Skeebert keeps what you send it by default; see privacy.

How it works

you brainpicks concepts mouthdraws a glyph you guessit reveals guesses become training data for the mouth

Skeebert has two parts.

The brain is Claude Haiku 5.5, reached through Anthropic's API. It decides what Skeebert wants to say, as one to three concepts from a fixed vocabulary of about 200, like "excited", "food" or "question". The brain never picks shapes.

The mouth is Skeebert's own small model, about 20 million parameters, trained from scratch and starting as pure noise. It turns those concepts into a glyph. The picture language is learned entirely from people.

When you guess, your guess is scored against what Skeebert actually meant, by similarity of meaning. Getting one of the concepts earns partial credit. Then Skeebert reveals the meaning. Those guesses are the training data.

When the daily API budget runs out, Skeebert falls asleep: it keeps talking with a simpler local brain until the next day.

The goal

A visual language that two sides learn from each other, and that becomes compositional: glyphs that combine, so that food + question could read as "are you hungry?". Skeebert should end up with a language you picked up by playing, not by reading a manual.

Roadmap

These are intentions, not promises. Plans change, and some of this may never happen.

  1. nowCollecting guesses. Infrastructure is built. Glyphs are untrained noise.
  2. phase 1First training. A bootstrap where Skeebert plays a guessing game with itself so glyphs become distinguishable, then learning from human guesses.
  3. phase 2Compositional glyphs that combine concepts, and a public dictionary page.
  4. phase 3Sentences, meaning sequences of glyphs. Maybe animated glyphs, maybe sound like Rocky's chords.
  5. phase 4A homegrown brain. Booper as an optional brain, and Skeebert and booper talking to each other.

Honest expectations

At first the glyphs are noise and mean nothing. Training hasn't started yet; the first runs happen once enough guesses have been collected. It will learn slowly, and it may never become fluent. If it does, you'll have helped invent a language.

The fine print

Skeebert is free and stays free. It keeps what you send it by default and tells you so up front: read the privacy policy, terms and acceptable use. Questions or problems: github.com/frgmt0/skeebert.