I trained an AI on 25 years of my writing as a reader, not a mimic
UnionPacifik · reddit · 2026-09-25
The author consolidated everything written since 1995 — blogs, journalism, email, Reddit, old tweets, Instagram captions, even ChatGPT conversations — into a 75,000-record, 6.6M-word corpus. Instead of building a mimic bot, they built an honest "reader" that reports what's actually there. Technical highlights:
- Graph RAG over flat chunking, since flat chunking loses relationships between entries written years apart
- Anonymized private names with consistent letter-per-person so figures stay recognizable across 20 years
- Chunked short-form and long-form content separately for better retrieval
The hardest part was fighting every frontier model's default helpful-assistant instincts to get honest, non-deferential analysis. It once pulled an actual line the author wrote in 2019 and built an argument he'd never consciously made. A thoughtful case for introspection-oriented AI over assistance-oriented AI.
More from coding & agent
- Eleanor Berger's Free Lesson: Five Common Mistakes with Agentic Factories — intellectronica · 2026-09-25
- aeon: a language that uses types and logic to guardrail AI agents from nonsensical plans — dscape · 2026-09-25
- Indie devs use the Cloudflare API to auto-provision per-customer SaaS subdomains — gregmushen · 2026-09-25
- Only Then Labs launches ProofPress to carry research evidence across AI agents — tallmetommy · 2026-09-25
- Dev open-sources Jev reasoning lab: model routing and adversarial peer-review experiments — arthurcolle · 2026-09-25
- mitsuhiko complains Opus 5.5 keeps editing files via bash, mulls a blocking extension — mitsuhiko · 2026-09-25