Karpathy retrospective Hacker News analysis

Andrej Karpathy ran a retrospective analysis on Hacker News using GPT-5.1 Thinking to score December 2015 HN content. - Dataset: 930 HN posts from 2015-12. - Method: GPT-5.1 Thinking reviewed and ranked comments for prescience about future outcomes. - Cost/time: project took ~3–4 hours to code and run, costing about $60. - Results: identified top prescient commenters and statements (examples listed in original analysis). - Broader note: exercise framed as training predictive judgment and highlighted that future LLMs will cheaply audit public archives.
Key insights
Karpathy conducted a cost-efficient retrospective HN analysis with GPT-5.1 Thinking
Karpathy used GPT-5.1 Thinking to analyze 930 Hacker News posts from December 2015, scoring comments for hindsight accuracy and identifying the most prescient contributors. The workflow required a few hours of coding and execution and cost roughly $60, demonstrating affordability for retrospective large-scale text analysis.
Retrospective scoring aimed to calibrate prediction skill and highlight archival value
The analysis served to surface historically prescient commentary and to illustrate how modern LLMs can audit public discourse to evaluate past predictions. Karpathy noted that such retrospective exercises could improve forecasting skill and that cheaper future LLMs will enable extensive archival reviews.