The research question is simple: How close are we to being able to extend a human personality and patterns of thought in a meaningful way?
Why Paul Graham? Instead of using an average person with low signal-to-noise ratio sources like emails, this experiment works with a best-case scenario:
- Someone who has been thinking and documenting their thought process for an extended period
- Someone with a developed and distinct thinking and writing style
- Someone who writes many short-form pieces rather than few long books
- Someone moderately famous online - well represented in LLM training data
- Someone still alive - who could potentially assess the quality of their extension
- Someone with an existing community - for independent evaluation
The system:
- Allows users to generate essays by providing only a title
- Uses all of Paul Graham's essays as reference material
- Does RAG search of the reference material to provide the generator with the most relevant information in addition to the full corpus
- Employs a modern LLM (currently Gemini 2.5 flash preview) with large context window to generate new essays in his style (RAG's role is just to emphasize the most relevant parts)
- Collects ratings and comments on the generated essays
All generated essay data is freely available for research and analysis. (Download from https://wwpgw.georgestrakhov.com/about)
Technologies:
- scraper with python and Groq for smart MD conversion
- worker writer using cloudflare worker, queue, cloudflare autorag and openrouter
- website on nuxt3, deployed to nuxthub
Everything is open source. Have fun.
