Definition
Plain language
A design where two AI models work side by side, whispering to each other through their internal states instead of through words.
As stated in the literature
An architecture coupling two frozen pretrained transformers through learned, gated, bidirectional residual-stream connections at intermediate layers, with a small trainable bridge as the only added parameters.
Also called: bicameral
Why it matters: It explores whether two existing models can be combined more cheaply than training a new larger one, an attractive option as base models become more expensive.
For example, two frozen base models might be wired together so layer 12 of one feeds into layer 12 of the other through a small trained bridge, letting them collaborate without retraining either.
Heard on the show
“The paper is "The Bicameral Model: Bidirectional Hidden-State Coupling Between Parallel Language Models," from a team at AWS Agentic AI — Cedric Flamant, Udaya Ghai, and Kanna Shimizu.”Episode 040 — Two Frozen Models Learn to Whisper: Coupling Through Hidden States