When one of the world’s most prominent artificial intelligence researchers publicly admits that YouTube’s algorithmic feed is ruining his attention span, tech workers pay close attention.
Over the weekend, Andrej Karpathy—former Director of AI at Tesla and founding member of OpenAI—quietly published a minimalist weekend side project called HackerTube.
His goal was deceptively simple: create a clean, distraction-free YouTube frontend that completely strips out the platform’s machine learning recommendation engine.
Instead of autoplay loops, red alert notification badges, and algorithmically engineered outrage thumbnails, HackerTube only surfaces videos that real human engineers have explicitly recommended in comments and posts on technical discussion boards.
Within 24 hours of going live, Karpathy’s post racked up over 1,850 upvotes and 480 comments, touching off an impassioned discussion on algorithmic fatigue.
The Problem with Optimization for ‘Watch Time’
As an engineer who helped pioneer the modern deep learning models powering today’s internet, Karpathy’s critique carries unique technical weight.
Modern social video platforms do not optimize their feeds for human learning, focus, or intellectual fulfillment.
They optimize strictly for watch time and session retention—a mathematical objective that naturally rewards sensationalist hooks, anxiety-inducing clickbait, and endless 10-second dopamine hits.
“The algorithms are too effective at finding what captures short-term gaze rather than what nourishes long-term curiosity,” Karpathy observed in his blog post explaining the tool’s architecture. “I wanted a library of video lectures and technical demonstrations, not a digital slot machine.”
How HackerTube Actually Works
HackerTube functions as a lightweight, zero-bloat web client:
-
Human Vetting Only: It scrapes video URLs shared within high-signal discussion threads across technical communities.
-
No Algorithmic Sidebar: When a user opens a video, there is no infinite scroll column of related sensational videos pulling attention sideways.
-
Zero Shorts and Reels: The interface completely filters out short-form video clips designed for rapid doomscrolling.
The project touched such a sensitive nerve that within hours, community developers began cloning the concept for other media.
A spin-off called HackerBooks appeared by Sunday morning, filtering reading recommendations exclusively sourced from community bookmarks.
The Internet Reacts: The Great Algorithm Exodus
In the comments section, hundreds of software developers, educators, and knowledge workers shared their own exhaustion with algorithmic feeds:
-
“I spent two hours on YouTube yesterday trying to watch a 15-minute tutorial on linear algebra, and ended up watching a video on celebrity house tours,” lamented one senior engineer. “The feed is literally designed by thousands of PhDs to hijack my executive function.”
-
Another commenter praised the return to early web mechanics: “Human curation via forums was how the internet worked in 2005, and it was infinitely healthier. Bringing human taste back over engagement algorithms is the best trend of 2026.”
-
A third user noted the irony: “When the people who built modern AI systems start engineering tools to protect themselves from their own creations, that tells you everything you need to know about where consumer tech is headed.”
The Wider Battle for Digital Sanity
The immediate virality of HackerTube reflects an accelerating pushback against algorithmic manipulation among high-output professionals.
From minimalist “dumbphones” to custom RSS feeds and distraction-free web browsers, users are increasingly willing to build custom software just to escape the corporate attention economy.
For Karpathy, the project was just a personal weekend hack.
For the thousands of knowledge workers sharing it, it looks like a blueprint for how people might choose to consume media in the future.