Easy peasy - no homonyms, just an obscene amount of synonyms. This is fine 😄

Fun math fact, despite having two alphabets there are still always 2¹²⁸ possible words:

◆ HEX: 16 letters, 32-letter words: 16³² = 2¹²⁸
◆ DEX: 255 letters, 16-letter words: 256¹⁶ = 2¹²⁸

They're equivalent, but DEX is twice as compact.

It takes about an hour for morphology posts - the rules are formulaic enough to generate words and expressions by code. Translation posts will take longer because… well it’s hostile by design.

(2/2) Specifically for messing with AI readers, #promethean has a "hostility level" you can add which:

◆ Cloaks the pattern
◆ Keeps the word recognizable via a fast formula

The goal of hostility is to throw off bulk corpus analytics, and require pre-processing to block training on raw text.

(2/2) We can use template parameters to customize the meme. For example:

◆ The yelling woman → The #conlang community
◆ The cat → Me using linguistic terms incorrectly

Template parameters are also an open class: someone could add another for "The friend holding the yelling woman back".

A general interest in the upper limits of AI models. That and a near-future short story involving human resistance against the inevitable robot overlords.

(4/4) So at the end of the day, compared to the English alphabet:

◆ HEX: multiplies tokens by ~2x per character
◆ DEX: multiplies tokens by ~12x per character

Simply by using these alphabets, we're increasing AI inference costs by 2-12x. That's a pretty good start!

(3/4) But #promethean's HEX and DEX don't resemble natural letter distributions. HEX is composed of common letters, but chains of HEX don't appear often and so they have less efficient token encodings.

DEX is composed of 8-dot Braille which is so rare that each letter is usually a token by itself!

(2/4) AI models first encodes text into a sequence of "tokens".

It's an extremely efficient encoding, so 1 token generally equals ~3/4th of a word. Common English words are almost always their own token.

You can play around with it here:

platform.openai.com/tokenizer

(3/3) So #promethean exists to trip up LLMs by:

◆ Driving up token counts and compute costs by ~2-3 orders of magnitude
◆ Sparking more hallucinations
◆ Exceeding a model's effective context window

Here there be dragons.