#research

AI models reason in a language they borrowed from us, and training is quietly rewriting it. An essay on compounds, telegrams, scratch paper and handoffs.

Reasoning models write notes to themselves before they answer. Why they need this scratch paper, why it is getting harder to read, and why that matters.

VectifyAI's PageIndex swaps vector search for tree reasoning and scores 98.7% on FinanceBench. The result is real. The question is what it actually proves.

Andrej Karpathy's autoresearch lets AI agents run about a hundred ML experiments overnight on one GPU. It looks like the beginning of autonomous science.