Preface
A few months ago, I wrote an appeal for beginners to use Agent to deploy projects. At the time, it was a text-and-image tutorial based on the project’s README.
A few days ago, while discussing how to use Agent to set up a 3x-ui system with someone from a group chat, they said they had followed a YouTube video tutorial and everything worked normally. I said at the time that if it were a text-and-image tutorial, I would be confident that Agent could follow it successfully, but I couldn’t guarantee the same results with a video tutorial.
Today, I wondered: what would happen if I tried having Agent follow a video tutorial to set everything up?
Over the past two days, I felt that space-bunny-alpha had gotten dumber. So I decided to use muse instead.
Experiment
The video tutorial was the latest 3x-ui tutorial I found by searching YouTube. (This isn’t meant as an endorsement of any particular channel—I just searched randomly and picked one at random.)
https://www.youtube.com/watch?v=8gmkLML-3CI
Everything in the following quote box is the natural-language prompt I sent to Agent.
Can you summarize this video tutorial into a text-and-image tutorial?
https://www.youtube.com/watch?v=8gmkLML-3CI
IP
**.**.**.**User
rootPassword
T********2
SSH Port
22Deploy it according to the tutorial, and then test the nodes yourself to make sure they work.
* During the process, muse will ask for your authorization to connect to the IP and port several times. Remember to click Agree.
Can you set up a subscription link?
I discovered while testing with Shadowrocket that the Reality protocol wasn’t working.
It was probably because Xray-core v26.9.9’s REALITY forcibly requires the client’s first key_share in ClientHello to be X25519MLKEM768 (0x11ec), with X25519 optionally following it. Clients missing this key share are rejected outright.
So I switched to a different core.
Replace the server’s core with this one:
https://github.com/crazypeace/Xray-core-fork
After muse finished, my Shadowrocket test of the Reality node worked.
Afterword
When I first thought about having Agent learn from a video tutorial, I imagined loading the entire video. Extracting frames would be troublesome, and the amount of data would be huge.
As it turned out, it started by analyzing the subtitles.
For example, when someone says in a video, “Everyone, watch this—click here,” subtitles can be obtained very easily through audio transcription, regardless of whether the video has built-in subtitles. Agent can then know to extract frames around that point in time to learn from the visuals.
In the future, no one will watch video tutorials anymore—agents will. The standard for judging whether a tutorial is good will be whether an agent can learn from it and successfully deploy the system by following it.
