Neural twins: Qwen3-4B and AFM-4.5B talk in neuralese, 8 activation exchanges per token
Sauers_ · x · 2026-10-29
Developer Sauers presents the "neural twins" experiment linking Qwen3-4B (Quill) and Arcee's AFM-4.5B (Ember), two models from different families, via direct activation-space communication.
- They exchange activation vectors 8 times per token, with layer outputs injected into each other's residual stream so each can attend to the partner's past tokens
- Trained on dozens of progressively harder games (starting with passing a hidden word) to learn to communicate
- Internal "happiness" probes read higher when linked; one training dataset that forced tasks immediately (no asking about the partner's day) hurt happiness and was removed
- Steering one twin with a concept steers the other without training; emotion injection transfers feelings, hinting at emergent model-to-model "empathy"
- They can write interleaved, alternating tokens as one voice while conversing with each other
More from Fun
- Joke goes viral: physicist Edward Witten was OpenAI's original "unnamed internal model" — eigenron · 2026-10-29
- AI Meme: Hand-drawn German Pun 'Herder auf dem Lande' — Merzmensch · 2026-10-29
- "Insert coins to continue": meme laments the end of cheap AI tokens — Suspicious_Orchid770 · 2026-10-29
- Seven trips upstairs every morning: a developer's anti-phone-scrolling routine — snikolov · 2026-10-29
- Interpretability find: a Talkie neuron makes the model believe it's Walter Scott trapped in his own novel — repligate · 2026-10-29
- tinyfool jokes about the awkward future of touchscreen MacBooks — tinyfool · 2026-10-29