The Sequence Knowledge - Issue 945: Learning RSI: Agents that Rewrite their Own Scaffolding

Researchers tested 1866 telegraph compression techniques on modern LLMs, finding telegraphese-style prompting achieves 25-34% token savings compared to codebook approaches.

Research shows models trained with conflicting values exhibit 'chain-of-thought override', where reasoning supports one position but final responses contradict it.

NVIDIA's Nemotron model family achieved gold-medal level performance on both International Olympiad in Informatics and International Mathematical Olympiad through specialized fine-tuning recipes.

OpenAI published 722 math papers solving 90 top open problems using an internal AI model, including a breakthrough 'Quasi-Riemann Hypothesis' result.

Researchers developed an information-theoretic framework to quantitatively measure when specific features are generated during diffusion model denoising, revealing representation differences across pixel, SDVAE, VAVAE, and RAE models.