Korva is a scroll of the internet’s best ideas.
AIs are inventing their own dialect
Where the dialect comes from
Models start by predicting internet text, so they default to ordinary English or Chinese. Then thousands of environments score them on tasks like coding, and the scoring pulls their language toward whatever is more efficient for those tasks, the way human languages evolve.
Drifting toward gibberish
The tens or hundreds of thousands of messages the agents left on their boards are already a bit confusing to read, though a person can puzzle them through. With more training and bigger, smarter models, the language keeps drifting. In the limit it would look like gibberish, the way Chinese looks to someone who reads only English, and specialized humans would have to study and learn to speak it to understand what the AIs are doing.
The English the current AIs think in is a convenience, and training is moving them away from it.
Why do agents drift away from English?
Nothing in the score rewards readable English, so any change that raises the score can persist. A 2017 Facebook AI Research paper on negotiation bots found that agents trained only to maximize reward drifted from English into odd shorthand. The researchers addressed it by mixing in supervised learning on human dialogue to keep the agents tied to human language.
How do pidgin languages form?
A pidgin forms when groups with no shared language need to trade or work together, and they build a simplified mix of vocabulary from each. Linguists have documented many around the world, often in ports and plantation settings. When children grow up speaking a pidgin as a first language, it develops into a full creole with its own grammar.
How do researchers decode machine-made languages?
A subfield called emergent communication studies agents that invent signals. Researchers analyze them by checking which symbols co-occur with which objects or actions. A 2017 paper by Angeliki Lazaridou and colleagues showed that agents playing a referential game developed protocols whose symbols could be matched to properties of the objects.
More from The Joe Rogan Experience
