Princeton's Logion suggests restorations for damaged ancient Greek texts; however, it treats every letter swap as equally likely. I tested an improvement swap based on how easily scribes would actually confuse letters. This improvement performed significantly better in a short test on Photius.

A benchmark for whether a language model can propose creative good ideas. Each idea is scored by a panel of model judges on feasibility, impact, and originality. The benchmark is based on consulting prep cases because they were readily available and included pre-scored, uncreative/basic ideas.

A small A/B harness for comparing language models across providers, built to verify this paper, which still seems to work.

A 3D map of US public companies built from ~17 years of public SEC 10-K filings (~12k companies), built to visualize interesting clustering patterns and aberrations in businesses. Highly recommend rolling the dice next to the search bar to find outlier/interesting companies and the insights subpage for a more traditional set of visualizations.

Live USDC transfers across nine blockchains, turned into sound. Inspired by listen to wikipedia, a new and more natural way to understand this data.

A bulk-ingest tool for Claude Code that runs document search agents in parallel with prompt caching; testing showed 89% cheaper than running as subagents within a Claude Code session.