December 10, 2025
Your takes, now graded
Auto-grading decade-old Hacker News discussions with hindsight
Karpathy’s bot regrades 2015 takes, HN freaks out and bets on 2035
TLDR: Karpathy’s AI project publishes grades and summaries of Hacker News debates from December 2015, complete with a Hall of Fame. The crowd splits between joking “AI doom” and worrying about biased judgment, while others share tools and make bold 2035 predictions—turning nostalgia into a lively courtroom for old takes.
Hacker News just got a retro report card, and the class is buzzing. Andrej Karpathy dropped a nostalgic grenade with HN Time Capsule, where an AI skims the front page of December 2015 and grades everyone’s hot takes—complete with a Hall of Fame shoutout for legends like tptacek and moxie. Cue the drama: one commenter instantly yelled “Roko’s Basilisk,” turning “LLMs are watching” into a sci‑fi panic about future robo-judgment. Another went full dystopia, warning that we’re being evaluated by an “aligned” AI worldview, asking if we’re “Alpha or Delta‑Minus.” The vibes? Equal parts nostalgia, existential dread, and popcorn.
But it’s not all nosebleed-level philosophy. The thread quickly veered into future bets—one user declared there’ll be no app stores and most software will be AI‑generated by 2035. Meanwhile, a community member brought receipts with a slick replay visualizer to relive the old threads in real time, and another praised HN’s good web hygiene: stable links let our past be perfectly re-read. Between the jokes, the warnings, and the bold predictions, this feels less like a museum and more like judgment day with a vibe coder. The GitHub for the project is here: karpathy/hn-time-capsule.
Key Points
- •A pipeline was built to analyze and grade Hacker News front pages from December 2015 using GPT 5.1 Thinking.
- •Data is fetched via the Algolia API, packaged into markdown prompts, submitted through the OpenAI API, and rendered as static HTML.
- •The code is available on GitHub (karpathy/hn-time-capsule) and results are hosted at karpathy.ai/hncapsule, including intermediate data.
- •The project aims to study prescience in discussions and illustrates broader ideas about training future prediction and pervasive LLM-enabled scrutiny.
- •Example graded threads include Swift open-sourcing, Figma’s launch, OpenAI’s announcement, George Hotz’s Comma effort, SpaceX’s Orbcomm-2 webcast, and Theranos issues.