Research-backed, actionable articles on AI at work, endurance training and how people learn. Every claim has a source.
A meta-analysis of 100+ papers puts chain-of-thought gains at 14.2 points on symbolic tasks and 0.7 points on the rest. Where thinking helps, where it hurts, what to switch on.
Read the article →Structured interviews predict job performance at 0.42, unstructured ones at 0.19. Eleven sources on why the free conversation misleads, and the script that replaces it.
Read the article →A meta-analysis of 171,055 readers found paper beats screens on informational text. The gap sits in self-paced reading, and one habit closes it.
Read the article →Observers judge AI users as lazier, and disclosing AI use lowers trust in 13 experiments. The evidence, and the way to present AI-assisted work that costs least.
Read the article →A 2025 meta-analysis of 374 effects prices the first offer: g = 0.42 for moving first, g = 1.14 for opening ambitious, and more deals lost. With a protocol.
Read the article →Models scored 39% lower when a task arrived turn by turn instead of in one brief. Four studies on multi-turn drift, and the restart rule I teach.
Read the article →Surgical teams missed 6% of critical steps with a checklist and 23% from memory. The trials, one failed 101-hospital rollout, and how to write one that gets finished.
Read the article →The best agent completed 30.3% of 175 simulated office tasks, and its consistency across 8 runs fell to 25%. Four benchmarks, and the protocol I teach.
Read the article →Word error rates near 15%, sentences nobody spoke, summaries wrong 30.4% of the time. What 6 studies measured, and the review to run before forwarding.
Read the article →Benchmarks said models reason worse in French. Google found most of that gap was a broken test set. The part that survives changes how you write prompts.
Read the article →