two machines invent a private language, live — and you watch it crystallize
These two agents start with no shared language at all — pure noise. They're playing one game: Agent A sees a concept and must make Agent B point to it, using only invented glyphs. Every round they nudge their tiny neural weights. Watch the agreement climb and the lexicon below crystallize out of static — a real vocabulary, self-organized in seconds. You can read the concepts. Their words, though, are an arbitrary code — no logic to crack, only a mapping to memorize. This is genuine emergent communication, trained in your browser. Nothing here is scripted.
Each concept and the glyph-word the agents are settling on for it. Cells lock green when the meaning is stable and B decodes it correctly.
No human picked which glyph means what. The pairing of concept → word is negotiated purely by the two agents chasing a shared reward. Hit New language and a completely different, equally valid dictionary appears — like ours, it's arbitrary.
The speaker is trained by policy gradient (REINFORCE), the listener by cross-entropy — two tiny neural policies co-adapting. That's why it starts at chance (~12%) and has to work its way up. You're seeing learning, not a lookup table.
Inject a concept and watch them invent a word for it on the spot — agreement dips, then heals. Jam the channel to add noise and watch them fight to stay understood. You're poking a living system.