Auto-grading decade-old Hacker News discussions with hindsight
2 days ago
- #Hacker News
- #LLM
- #Future Prediction
- The author explores using LLMs to analyze decade-old Hacker News discussions for prescience.
- A project was created to download and analyze HN front pages from December 2015 using GPT 5.1 Thinking.
- The analysis includes summarizing articles, evaluating comment predictions, and grading commenters.
- Key findings highlight prescient and incorrect predictions from past discussions.
- The project aims to train future prediction skills and emphasizes the scrutiny future LLMs may bring.
- Examples of analyzed topics include Swift going open source, Figma's launch, and Theranos struggles.
- The project cost $58 and took about an hour to process 930 LLM queries.
- Results are hosted on karpathy.ai/hncapsule, with code available on GitHub.