Articles.
Ideas behind the papers, worked examples, and interactive explanations. A closer look at the questions, methods, and limits of the research.
All Notes
Diagonal runs of punctuation in Chinese chatbot replies looked like a model fingerprint. Pre-ChatGPT human answers have more of them, every corpus stays close to a random-placement baseline, and what the models change is where the staircases appear.
Companion note to "Recombination or Discovery?" (NeurIPS 2026 AI4MetaScience workshop poster): machine contributions are judged facet-novel more often than matched human ones (56.1% vs. 35.6%), but the gap rests mainly on the purpose facet, LLM re-auditors return refutation rates from 0% to 100% on the same items, and retrieval finds the known prior art for only 25 to 29% of gold pairs. Plus an integrity audit: hard fabrication evidence in 0 of 47 accepted and 16 of 197 rejected Agents4Science 2025 submissions.
Interactive companion to the HHAI 2026 paper "AI as Equalizer or Amplifier?": why AI compresses the novice-expert gap on routine tasks and widens it on judgment-heavy ones, how model deference compounds the difference turn by turn, and what the position paper cannot yet show.
Citation bundles a ledger (registration, priority) and a currency (reputation) in one act, coupled only because reading was the sole transport layer of science. I argued LLM reader-side consumption is splitting them, and made three falsifiable predictions. Updated with the measurements: reuse and citation did come apart, by roughly half over a decade, but my mechanism lost. No break at model release, no field-exposure gradient, and the 2015 cohort declines as steeply as the 2023 one.
Companion note to the Tuto audit report and the paper behind it: fabricated references are a rounding error (2 confirmed, 0.001%); claim-support rates did not replicate across three identical-pipeline runs, so the paper takes non-replication, not any single rate, as the principal finding.
Interactive companion to "When Should the Agent Speak?": twenty years of research learned what it costs to interrupt a person, and had no agent capable of earning that cost back. The agents arrived. The price did not come with them.
Interactive companion to the working paper "Registers, Not Plans": an independent replication of Anthropic's global-workspace claim, and why only context registers, not content plans, survive a strict readout test.
Interactive companion to "Fidelity Before Structure" (arXiv:2601.00821): explore the benchmark results and see why extraction loses to verbatim chunks at write time.
Interactive companion to "Consensus Density Predicts Output Dispersion in Aligned LLMs" (TMLR 2026): sample clouds, judge-predicted dispersion, alignment as amplifier, and why instruction form is not consensus.
Cognitive Workspace (2025) argued for actively curated memory; my own 2026 ablation showed curation is lossy deletion. What failed, what survived, and the meta-lesson about measuring claims.