According to @karpathy, he used the GPT-5.1 Thinking API to auto-grade all 930 December 2015 Hacker News frontpage article-discussion pairs to identify the most and least prescient comments, taking about 3 hours to write the code and roughly 1 hour and $60 to run, source: twitter.com/karpathy/status/1998803709468487877 and karpathy.bearblog.dev/auto-grade-hn. According to @karpathy, the project repository is available at github.com/karpathy/hn-time-capsule and the full results are browsable at karpathy.ai/hncapsule, source: twitter.com/karpathy/status/1998803709468487877. According to @karpathy, he emphasized in-hindsight analysis as a practical way to train forward prediction models and noted that future LLMs will perform such work cheaper, faster, and better, source: twitter.com/karpathy/status/1998803709468487877. According to @karpathy, the top 10 most prescient HN accounts for that month were pcwalton, tptacek, paulmd, cstross, greglindahl, moxie, hannob, 0xcde4c3db, Manishearth, and johncolanduoni, source: twitter.com/karpathy/status/1998803709468487877. According to @karpathy, these run-time and cost figures provide a concrete real-world datapoint for large-scale LLM evaluation workflows using GPT-5.1 Thinking, anchored at approximately $60 for a 930-thread pass in about one hour, which traders tracking AI infrastructure efficiency can use as a benchmark, source: twitter.com/karpathy/status/1998803709468487877 and karpathy.bearblog.dev/auto-grade-hn.