An AI got stuck on a CAPTCHA game. Then it opened YouTube on its own. OpenAI's coding agent Codex struggled on level 15 of the "I'm Not a Robot" game, then watched a YouTube tutorial and solved the problem. It went viral on X with 3,700 likes.
The AI Stuck on Level 15, and the Method the Agent Chose
In the video, Codex was working through the CAPTCHA-style puzzle game one level at a time and got stuck on level 15. What it chose to do next is telling. It did not switch tools. It went looking for a solution through a YouTube search. It played the tutorial, then followed it to solve the level.
What Is That Game? Neal.fun's "I'm Not a Robot"
The game is "I'm Not a Robot" by developer Neal Agarwal (neal.fun). It is a puzzle that parodies CAPTCHAs only humans can solve, and level 15 is a "parking" problem where you fit a Waymo car into a parking space. The agent got stuck in this section, which requires understanding coordinates and the screen, and then had a breakthrough after watching a video walkthrough.
Playing a Video Is Not the Same as Understanding
It would be natural for a person, but it is unusual for an AI agent. It means that when stuck, instead of stopping or writing different code, it chose on its own the "search, then watch a tutorial" path that people use. It went beyond simply using tools and even imitated how people find a way to solve problems.
Why It's Getting Attention: Seen as a "Computer Use" Test
The game's creator, Neal Agarwal, replied that he would work on a sequel, and the post's author, Sharif Shameem, fired back that "this game would make a good benchmark for measuring computer use." The reason is that the whole chain of operating a browser, understanding a video, and acting on it comes together in one scenario.
The Limits Are Clear Too
Of course, this does not mean "AGI." The agent relied on human-made content to get past where it was stuck, and CAPTCHA games were originally a domain for telling humans from bots. Still, the agent following the basic human path of "when stuck, search, watch, and solve" offers a glimpse of the next stage for autonomous agents.
Related: X post (original) · neal.fun
