For a long time, I was in some doubt and thought the models had very little improvement over time, including the chain-of-thought stuff. They did get quicker and faster and made fewer mistakes, but this AI Agent is truly amazing.

Technically, everything is doable without AI Agents with enough time and prompting. However, with AI Agents it really makes things a lot easier.

So the AI Agent spuns up in a remote docker (I don't want it running on my machine), and you can just spun up 3-5 AI Agent pipelines at a time, have tea for a while, and look at the output.

How Things Actually Work

For a long time I couldn't figure out how yt-dlp works. I mean, I know the source code but I didn't want to read it, and I just thought it was some weird InnerTube thing. Then I throw an AI Agent at it, "read the yt-dlp source code, and distill a small script for me to download youtube videos, strip everything else away". And it gave me this code and you can understand it very quickly. I think you can also understand stuff like the core of repos like basic DBs, network repos like shadowsocks much quicker.

Optimize Your Data Structures

Throw a dummy website you built at an AI Agent, give it the Github link, and you say, "optimize the data structure to make it faster". And it does all the unglamorous jobs for you. Though usually it's useless anyway.

Generate a Dataset or Build a Gymnasium Environment

You can say, "Build a gymnasium environment where people push things around", "Build a gymnasium environment similar to CLEVRER with balls colliding", and ask it to generate some videos and images, things like that. And it is very easy.

Figure Out the Python Env of a Github Repo Online

If you come across this really ugly conda thing on a Github repo, and you find it's very hard to get it running, you can past the Github repo link to an Agent and it will help you figure out all the dependencies.

Building Gymnasium Browser Pipelines

You can go on online games like slither.io or zombsroyale.io and ask the AI Agent to build a gymnasium pipeline for the game in headless firefox with Selenium, using the Javascript memory and variables as states instead of Screenshots stuff. Then you can try some heuristics or RL on it, and while it is playing you can let it capture the video on a headless browser.