Frustrated with YouTube's algorithm recommendations, Karpathy built HackerTube using Claude in two hours.

Author: TechFlow

TechFlow Summary: When top AI researchers grew tired of algorithmic feeding from commercial platforms, the personal software revolution quietly began. On November 2, OpenAI co-founder Andrej Karpathy revealed that, fed up with YouTube's monotonous recommendations and clickbait thumbnails, he had Claude independently build the video site HackerTube within two hours.

The core of this experiment was not merely coding, but pitting nearly two decades of genuine wisdom from the tech community against black-box recommendation algorithms. Leveraging Claude's efficient cleaning and semantic recognition of approximately 43 million historical comments on Hacker News, Karpathy achieved end-to-end delivery from mining vast datasets to production deployment.

In a human-AI collaboration costing 500 million Tokens, Karpathy only handled cloud infrastructure configuration and API keys, while coding, security hardening, and performance tuning were entirely autonomously executed by the Agent. This highly efficient practice not only heralds new possibilities for anti-algorithm recommendations but also sets a clear benchmark for AI-native full-stack development.

A Two-Hour Turnkey Project: Farewell to Algorithm Manipulation and Clickbait Thumbnails

According to a post on Karpathy's social media platform, after growing weary of YouTube's rigid, uniform recommendation mechanism and exaggerated thumbnails filled with visual stimulation, he decided to completely bypass the platform's official recommendation feed and build his own video navigation tool purely based on geeky tech aesthetics.

The resulting HackerTube embodies a minimalist style. The site features 6,454 high-quality videos covering 150 vertical topics, all drawn from nearly two decades of accumulated discussions in the Hacker News community. Users do not need to log in, data is stored locally, and the hassle of browsing traces being tracked is completely eliminated.

HackerTube completely discards the exaggerated thumbnails creators intentionally design for click-through rates, opting instead to uniformly capture a static frame from the center of each video via script. This minimalist design, which counters commercial tricks, returns video content to its objective essence and eliminates the visual pollution caused by clickbait titles.

Gold Rush from 43 Million Comments: Data Refining via Dual-Model Relay

Selecting high-quality tech videos from the vast ruins of the internet requires an astonishingly massive underlying data engineering effort. Throughout the process, Claude first performs a dragnet-style cleanup on approximately 43 million historical discussion records since Hacker News's inception, quickly narrowing down the potential video pool through regex matching.

Subsequently, the system invokes the Sonnet model for precise semantic disambiguation. The model must comprehend programmers' genuine evaluations within context, effectively separating authentic recommendations from ungrounded spam promotions. After multiple rounds of cross-referencing, this valuable set of content indexes was ultimately extracted from approximately 40,000 valid recommendations.

This capability of empowering data mining with large models demonstrates a powerful industrial leverage effect. Unstructured data cleansing that previously required weeks of manual review is compressed into mere minutes under AI's multi-layer filtering pipeline, extracting high-quality technical knowledge at a remarkably low time cost.

Claude Code Commands the Entire Process: Engineers Step Back to Infrastructure Commanders

In the coding and deployment phase, Karpathy experienced truly keyboard-free development. The entire HackerTube project was driven entirely by the terminal development tool Claude Code, where the more logically capable Opus 5.5 was responsible for writing data pipelines, full-stack website code, and deployment scripts.

Karpathy noted that over the two-hour duration, he did not write a single line of business code nor type complex commands in the terminal; he only manually applied for API keys and purchased the domain name. From cloud server initialization to security hardening and high-concurrency performance tuning, everything was autonomously advanced by the model.

This division of labor reflects a clear collaborative paradigm: the flagship model Opus 5.5 serves as the system architect and DevOps engineer, while the cost-effective Sonnet acts as the mainline processor handling massive classification tasks. Developers are completely freed from tedious syntactic details, transitioning into product managers and requirement specifiers for the overall business.

The Economics of 500 Million Tokens: The Dawn of the Personal Exclusive Software Era

The computational cost ledger behind the project is equally thought-provoking. Karpathy disclosed that the project consumed a cumulative 504 million input Tokens and 7.6 million output Tokens. Thanks to the context prompt caching mechanism, the hit rate for the input portion reached approximately 90%, significantly reducing actual computing costs.

Of this computational consumption, over 90% of the cost was allocated to Sonnet for corpus classification, while core architecture construction and website deployment itself only consumed about 1 million Opus 5.5 Tokens. This indicates that high-value architecture generation has become extremely inexpensive, with most compute power now being directed toward massive data reasoning.

HackerTube is currently live, and the cleaned dataset has also been open-sourced to the public.