Machine Learning Efficiency Podcast Summaries

Scientists Found A Better Language For AI Agents

Two Minute Papers

Jun 19, 2026

Dr. Károly Zsolnai-Fehér explores a breakthrough where AI agents bypass natural language, instead sharing raw neural latent states. This 'brain-to-brain' communication allows smaller models to achieve performance gains on complex math problems while cutting token usage by 75%, effectively outperforming traditional text-based agent coordination with minimal computational overhead.

Key insight: By replacing human-readable text with raw latent state transfers, small sub-10 billion parameter models increased their math accuracy from 73% to 86% while reducing token costs by 75%.