Published 1,000-transcript collection · relevance and completeness still require review
Explore 1,000 recovered public AI conversations referenced on GitHub. Follow the collection process and compare monthly counts alongside major model releases.
Explore the timelineA share URL enters this collection only after extraction of an ordered user-to-assistant exchange. Relevance and transcript completeness still require review.
Find share URLs in public GitHub issue and PR bodies and individually verified comments.
Both source creation and update dates must be in 2025. These are source dates, not chat creation dates.
Run resumable browser extraction on the cluster. Repair changed selectors and retain failure reasons.
Deduplicate URLs and normalized transcript hashes. Mask six token-shaped examples before publication.
Why no Claude transcripts? All 220 eligible Claude links attempted on the cluster hit access challenges. Claude was excluded from the final recovery job; zero recoveries does not mean zero usage. The collector stopped at 1,000 transcripts, leaving eligible candidates unattempted. Read the run record ↗
Open the interactive explorer to compare platforms, zoom into the year, and inspect model releases alongside monthly recovered-conversation counts.
A GitHub artifact’s timestamp does not establish when its chat was generated, or precisely when its link was added. The explorer uses the creation date of the issue, pull request, or comment containing each recovered chat’s share link. Multiple sources are assigned to the earliest qualifying source creation date, so each recovered chat contributes only once.
Model versions are unverified. Milestones come from linked provider announcements and describe availability, not attribution, adoption or causal effects on the collected chats.
Public sharing, discovery order and retrieval success shape these counts. Platform proportions do not estimate usage or market share. Development relevance, transcript completeness and model versions remain unverified.
Counts are generated from the published dataset at build time.