Originally Posted by
Gobbledygook
Hey brother. This is some cool shit. I remember speaking with Kris at length about his work on the decap AI bot. Are you retaining your data in a .csv or .xlxs format by chance with a specific delimiter?
Thank you bro.
The training data is not Excel. It is JSONL: one data entry per line.
The bot saves stats, evaluations, worker logs, checkpoint info, etc. as text logs, and the actual model is saved separately as neural-network checkpoint weights.
Originally Posted by
Moonshake
In and of itself, I don't have a problem that this entire project is (in all likelihood) "vibe coded". I don't even have a problem with the simple fact that the original post seems to be entirely written by a LLM. But, when you made the decision to include 11 (?!) AI-generated banners, on top of the fact the post is already unnecessarily long -- and now you're doubling down by throwing the questions of actual users into the prompt and pasting the output here in your responses -- all this insults the intelligence of the reader.
And if that's what you're going to do, how about you publicize your repository and have your agent create a skill out of the project and chat history? It'll save time for everyone involved!
You are half right, which is dangerous because it makes the wrong half sound more confident.
Yes, the main thread was cleaned up with ChatGPT/Codex help. That is not a secret. The whole project has been handled with Codex from the start, so when Fran asks me for the algorithm in detail, I’m obviously going back to the thing that has been staring at the codebase for days and asking it to help me answer properly.
That is not insulting the intelligence of the reader. Giving a vague “yeah bro it learns with AI stuff” answer would be insulting.
Also, let’s not pretend this is some free one-click magic trick. I pay for the AI tokens, I spent almost four days babysitting this thing, testing runs, fixing crashes, checking dashboards, reading logs, comparing checkpoints, watching the bot invent stupid strategies, and then forcing it to stop being stupid. All unpaid, obviously, because apparently my new hobby is arguing with Lua, Python, Toribash and now forum formatting.
So yes, when someone asks a technical question, I will use the tool that has the project context to answer it properly.
Some replies are mine, some are cleaned up, some are generated from the actual code/config. Welcome to 2026, we have tools now.
About the banners: they are there because the thread is a giant pile of experiments, failures, fixes, training campaigns, dashboard stuff, and bot drama. The images are supposed to break the wall of text. You may not like them, fair enough, but pretending the purpose was to hypnotize the reader with 11 sacred rectangles is a bit theatrical.
About open sourcing the repo: I could dump it tomorrow.
But then we are not just sharing a cute “AI learns ABD” experiment. We are also sharing a working Toribash automation bridge, training pipeline, checkpoint loader, dashboard, client launcher, and play/watch tooling.
Even if my scope is offline/local, handing that out without thinking is how you speedrun “why are there suddenly bots everywhere?”
So no, I’m not open sourcing it right now just because someone discovered that the AI project uses AI in the writing process too. Shocking development, I know.
I’m happy to explain the architecture, the algorithm, the data format, the results, the failures, and what changed between iterations.
But I’m not going to publish the whole repo just to win a forum purity contest.