Auto-grading decade-old Hacker News discussions with hindsight (karpathy.bearblog.dev)

Article summary

The author used a large language model (LLM) to analyze and grade decade-old Hacker News discussions, providing hindsight on the accuracy of predictions and opinions expressed at the time. The project involved downloading front pages of Hacker News from December 2015, parsing article and comment threads, and submitting them to the LLM for analysis. The results are presented on a website, allowing users to browse and explore the graded discussions. The author reflects on the potential implications of LLMs being able to scrutinize past online activity in great detail.

Main themes

• Hacker News discussion analysis
• Large language models
• Predictions and hindsight
• Online activity scrutiny
• Web archiving and preservation

What commenters say

• The ability of LLMs to analyze and grade past online discussions raises concerns about the potential for future scrutiny of current online activity.
• Some commenters believe that the value of maintaining a well-archived web is increasing as technology improves, allowing for more efficient processing and analysis of old content.
• Others argue that maintaining the web is not hard, but rather a matter of lack of will and forethought on the part of maintainers, with monetization also playing a role.
• There is a discussion about the potential for a decentralized and distributed web, with some commenters suggesting that IPFS could be a solution, while others express skepticism about its feasibility and effectiveness.
• Some commenters note that people's past opinions and predictions can be both accurate and inaccurate, and that it's important to consider the context and limitations of their knowledge and expertise.
• The concept of the 'Halo effect' is mentioned, where people tend to assume that someone who is right in one area will be right in another unrelated area, and the importance of recognizing and tempering this cognitive bias is emphasized.