TJDAX / AI lab
AI lab
We research AI and invest in building it. Reinforcement learning references are a large part of that work. Over time we will develop and run our own language model.
How we work
- Research and investment. The lab is where we study AI and put time and resources into developing it, not only into using models built by someone else.
- Many reinforcement learning references. We keep a wide set of tasks so methods can be compared. A few names below are in training or in planning. They are references, not the whole lab.
- Our own language model, later. The long-term aim is to develop and operate a language model of our own.
- Numbers only after we measure them. A result appears on this page when we have it. Failed runs stay in the record. An AI result is labeled as AI.
Reinforcement learning references
The reference set covers sequential decisions, control, and planning. Most of it is not a product and not a game write-up. We use the tasks to see whether a method learns, where it fails, and what is worth carrying into later systems, including the language model.
Apple
Apple is a Tetris AI model, and the first one tjdax trained. Development closed on 10 October 2026 after generation 10,211. The recorded training time on that checkpoint is 1 hour 13 minutes 33 seconds. The final ladder score is Elo 970.4 against a heuristic anchored at 1000, with 8 wins, 525 losses, and 8,149 draws. The checkpoint files and their SHA-256 hashes are on the result page. Training was not continued: public open-source weights for this game already exist, and this run does not replace them.
Gomoku
One reinforcement learning reference among many. An agent learns the task from its own attempts and the score it receives. Training is running. The setup and the log will be published here when the first evaluation is done.
Project 1:06
A driving task kept as a reinforcement learning reference, not as a game page. It is in planning. We will publish the method and the measurements when there is something to measure. An agent result will not be submitted to a human leaderboard.
Our own language model
Reinforcement learning is the work in front of us now. The longer project is a language model we develop and run ourselves. That model is not built yet. This page will record the research as it becomes something we can show.
Follow along
Notes from the lab will be posted here as the references and the language-model work move forward.
For questions: tjdaxmail@gmail.com