Developer extracts 500GB of personal data to train local models, open-sources toolkit
BLUECOW009 · x · 2026-08-21
- Practice: A developer extracted 8 years of digital footprint (33k AI sessions, 12k chats, 800 notes, 20k tweets, 400 codebases, browser history, etc.), totaling 500GB, to train local models.
- Toolkit: The author released the "AI Coding Assistant Training Data Extraction Toolkit" on GitHub, which automatically extracts complete conversation history, code context, tool usage, and diffs from tools like Claude Code, Cursor, and Codex.
- Findings: The article details the "weird stuff" that happened after training on this massive personal dataset and provides a link to the full process.
More from coding & agent
- Reevaluating Agent Stack Preferences: Stability vs. Cost — generativist · 2026-08-21
- Git Visualizer: Turn Commit History into a 3D Mario Kart Track — MickeySteamboat · 2026-08-21
- Merge, PostHog, Redis Host 'Self-Driving Product' Tech Talks — shensi · 2026-08-21
- Codex Organizes 24,649 Unread Emails: From Cancer Cure to Inbox Zero — andrew_n_carr · 2026-08-21
- Zig-Based Coding Agent 'fx' Boots in 10 Microseconds — udmrzn · 2026-08-21
- Anthropic ships cookbook pairing Claude Managed Agents with CopilotKit's AG-UI protocol — EricBuess · 2026-08-21