Aug 17, 2026

An essay arguing that AI's strength in math comes from bookkeeping, holding definitions, constraints and proved results without drift, rather than deeper insight. Many human errors are management failures, not intelligence failures.

Why it matters: The weekend's biggest thread on how to read gold-medal-level math results. The counterargument: memory and reasoning can't be separated that cleanly.

Code is deterministic, AI is not, much like people. So the skills that matter when building with agents are delegation skills: shared context, clarity, feedback.

Why it matters: Naming the skill shift as management skills struck a nerve. A thick counter-thread argues for leaning toward micromanagement instead.

Research from MPI Tübingen and ETH Zürich: a 5B model pretrained only on K-5 curriculum, compared against an unfiltered control with the same architecture and token count. Live chat and open models included.

Why it matters: A controlled experiment on what a model doesn't know. It shows what pretraining data carves into behavior, in a form you can actually talk to.