HN Time Capsule
https://github.com/karpathy/hn-time-capsule- Category
- AI Tools
- Rank
- No. 1429Tools index
Previous survey · No. 1436 ·
- Pricing
- Open Source
- Type
- TOOL
- Builder
- karpathy
- GitHub
- 673 stars
- Date
About
Andrej Karpathy's project analyzing decade-old Hacker News discussions in hindsight using LLMs to score predictions and ideas.
What it does
HN Time Capsule reconstructs a historical Hacker News front page, collects linked article text and full comment trees, then runs a staged command-line workflow. It prepares model prompts, stores responses, extracts commenter grades, aggregates results, and produces a browsable HTML summary.
Why it's ranked here
Worth ten minutes as a compact example of an end-to-end archival analysis pipeline. Its separate fetch, prompt, analysis, parsing, and rendering stages make the data flow unusually easy to inspect. The candid lack of intended support keeps it in experiment territory, not dependable infrastructure.
What's good
Each stage can run independently, which supports cheap test runs and avoids repeating every operation. Inputs, prompts, model responses, parsed grades, errors, and final reports remain as separate artifacts. A date selector and article limit make focused experiments practical before paying for a full model run.
Tradeoffs
The author explicitly provides the code as is and does not intend to support it. Analysis requires a paid OpenAI API call. Article extraction accepts HTML only, skips PDFs and several social or video hosts, rejects very short results, and truncates long text. Model judgments and letter grades remain generated assessments.
How to use it well
Use it for a reproducible research experiment, a small archival project, or as a readable pipeline example. Start with an article limit and a cheaper model, inspect saved prompts and responses, then render the report. It does not cover current Hacker News monitoring or production-grade supported analysis.
Technical notes+
pipeline.py is the single command-line entry point and contains the fetch, prompt, analyze, parse, render, and clean stages. Its visible fetching code parses Hacker News HTML, retrieves comment trees from the Algolia API, retries front-page requests with exponential waits, extracts text with Python's HTMLParser, caps article text at 15,000 characters, and records fetch failures separately. pyproject.toml requires Python 3.10 or newer and declares openai, python-dotenv, and requests; .python-version selects 3.10, while uv.lock pins the resolved dependency graph. README.md documents uv sync, an OPENAI_API_KEY in .env, per-stage CLI commands, and the date-oriented artifact layout.
Observed
- License
- MIT
- Primary language
- Python
- Runtime
- Python 3.10 or newer
- Install surface
- Dependencies install through uv sync using pyproject.toml and uv.lock
- Interface
- Command-line Python pipeline with independently runnable stages
- Declared dependencies
- OpenAI client, python-dotenv, and requests
Read from README.md, pyproject.toml, pipeline.py, uv.lock, .python-version.
What it can do
Analyze decade-old Hacker News discussions
Historical Hacker News posts and comments → Analyzed discussion data with hindsight context
Score predictions from old discussions
Past predictions and claims from Hacker News → Accuracy scores and evaluation metrics
Evaluate ideas with hindsight knowledge
Historical ideas and concepts discussed on Hacker News → Assessment of idea quality and accuracy using current knowledge
Extract predictions from historical posts
Decade-old Hacker News discussions → Identified predictions and forecasts made by users
Compare past expectations with actual outcomes
Historical predictions and current reality data → Comparison analysis showing prediction accuracy
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.