Dotplan

Updated September 24, 2026

September 24, 2026

TCG PPO ablations have lead down another path. I’m looking into an ablation where we give the model premade decks. At the start of the training 80% of games are played with pre-made decks, currated from our tournament decks, while the other 20% of games have the model draft their own deck. We then anneal this 80% pre-made deck ratio down throughout the training run.

The reason for this is because with previous attempts, when the model purely drafts its own decks (even with a prefix of up to 4 forced random cards), the model quickly converges on deck building strategies that lead to quick cheap wins. While the model is learning card lines, it cannot wield deeper strategies with the cards it converges away from, so it quickly removes exposure from those cards and the chance to learn them better.

So far the ablations are working well. I am now doing some refinements on the backprop for the LSTM and running short 15m step training runs to confirm my findings. I have currently removed all other reward shaping outside of the terminal reward. Depending on these results, I may re-add some reward shaping to maybe speed up training. Then depending on which path I take there, I will hopefully be able to start another production run.

Other

circuit class, content study

September 20, 2026

All TCG agent ablations aren’t going as originally thought. Need to probably break everything down into core components and really cull out a lot of previous ablations that aren’t serving the goal.

Today’s Reading

  • PBRS

    • Been using this in the TCG agent training and I’ve realized I need to familiarize myself with this more than just using it based on rough knowledge I read online.

September 16, 2026

TCG PPO ablations should finish up today in about 16 hours. After finalizing which recipe leads to strategic thinking, I will finally start a full production level run.

Continue courses on DB circuit analysis, still doing solves using Kirchhoff.

Research and plan to get better at storytelling

September 15, 2026

TCG PPO model ablations continue, ablations are quite slow having only a 3090…I should have bought that 6000 Pro before the price bumps. Whoops.

Ablations are looking promising now. Optimizing for winrates only were a mistake, as the model was converging on aggresive behavior that in the long term had a low cieling of improvement. Now I’m mostly focusing on pulling out elemental specific strategic plays, then running training for longer periods to optimize on those.

Continue courses on DC circuit analysis, more lessons on using Kirchhoff to solve currents, voltages, and resistances.

Today’s Readings

  • How I Learned to Read Way, Way More

    • Attention is precious and there’s distractions everywhere.

    • The more you read, the more you easily narrow in on the text that you’re hungry for.

    • The snowball starts with what you’re familiar with, then it rolls and grows into inspirations from those previous authors leading to deeper texts.

    • Read with curiousity as the lead, not with expectations nor even to finish a text. The right text will pull you in.

    • Social media pulls you from the present. It’s a trap that scatters your mind, consumes your curiousity with cheap content, and which pulls you into a doom loop of the next refresh in search of enlightenment.

  • The optimal choice of hypothesis is the weakest, not the shortest

  • Closed Source AI = Neofeudalis

    • Open source ai is anti-fuedal. Absolute power always corrupt, it’s a mistake to fall for the message that the chosen few will keep us safe. Giving up freedom for the illusion of safety.

Reading Eventually

September 6, 2026

Working on more ablations for the TCG model today. I want to see stronger strategies emerge in deck building and combat before kicking off another production run.

Then gonna spend some time taking courses on Kirchhoff’s Voltage Law and solving circuits with Kirchhoff’s laws.

Aug, 8, 2026

Ran a ton of ablations that lead to better TCG model performance. Added a deck building phase for the model, deck prefixing, more reward shaping to help discovery, league training improvements.

There’s been improvements across the board, the last nit is that the model still builds similar decks for same element sibling portals.

Overall, my biggest blocker is im compute poor, so it’ll be tough hitting the scale I want. I’ll be looking into more training speed improvements as well as reading the echo 1 & 2 papers before implementing a distributed training run.

Depending on the budget I can get for compute will decide if this is a full train or a learning experience.

July 5, 2026

Fable ended up taking the JAX conversion for the TCG sim to the finish line, results were not great. I think due to all the branching that is required in the sim, there’s a lot of wasted compute. Also the memory requirments are HUGE for compilation, too big even for my 128GB RAM and 3090 system. If I really wanted to keep pushing this, I’d have to rent a larger system to compile and probably even further explore new ways to run sims in parallel. Overall, I’m gonna put this on the shelf and take it as a learning experience. It’s cool to see coding agents can create software in 1:1 parity given the original software to validate against.

Next steps, start on distributed compute for training and ablation on rewards until I see strong gate/leader specific strateigies emerge in the deck building and combat.

June 25, 2026

Factory is amazing but it’s not quite what I want, I’m going to have to build my own version. Being able to tap into sessions across devices is sick, but not being able to use local AI models on specific BYOM machines is kind of lame, it forces you into paying for their subscription for inference if you’re on your phone for example. It’s going to be a pain to replicate a lot of their functionalities I want, but I think it’ll be worth it.

  • build out architecture for OpenFactory
  • get hacksaw and clamps + film progress

June 24, 2026

Embrace Cringe

Moving kicked my butt the past few weeks, absolute marathon. It’s finally mostly done, the place is still a mess but it’s an organized mess at least. It’s at a place where I can just continue to make progress without the whole thing taking over my life.

One thing with all this moving chaos, my brain is kind of fragmented. I need to regain control of that, reduce my stimulation to regain my focus.

Things to do today

I’m gonna buy a hacksaw and some clamps, as I need to cut some aluminum rods for a robotics assembly. I should also film my progress while working on this, I keep saying I want to start making content. Embrace the cringe. Then, before I stat working on the assembly, I need to finish organizing my office. It’s probably the worst offender, out of all the rooms in my apartment.

Also, need to research on doing visualizations for this TCG model and also build out a rough model on distributed compute so I can wrap up training after this JAX sim wraps up. Hoping to wrap up this whole project ASAP.

Research Factory or OMP for more agent automation across organizations and projects.

C Sim -> JAX Sim Update

GPT 5.5 took over to try and bring it to the finish line. 5.5 is still running as we speak, I wouldn’t be surpised if I’ve burned $5k in compute, thank God for the sub. It’s still making progress, but it’s gotten to a point where I think I should have just learned JAX myself and did it myself. But tbh, with all this moving I probably wouldn’t have had time anyways, so hopefully it gets the 1:1 parity I’m looking for and in the meantime I can focus on other things.

June 12, 2026

Finishing up the first build outs of the portfolio website. For now it’ll be simple, just a place I can write daily notes and thoughts down (dotplan) and also do writeups of my projects in /posts.

Current focus is in finishing up the Azuki TCG PPO model while I finish moving into my new apartment. Right now I’m testing out Anthropic’s Fable with some goals towards that end.

The two goals running right now:

  1. Fable running on a lambda instance to do a conversion from the original C engine used for simulation to a JAX based engine. This is based on a paper I found that explored utilizing models to re-create sims in more performant languages. The models are able to use the pre-existing, less performant engine as a validation while putting together the new engine in the new language. The goal has been running for 24 hours now, but from seeing existing metrics, I should expect about a 5x performance improvement in the full training flow. Outside of the training flow, the JAX sim seems to run 100x faster than the C sim. This makes me think that after this, I’ll probably have to see if there’s any improvements I can make in the actual training pipeline itself. But I’ll take it step by step, as while it’ll be nice, if I don’t have to do that to get to the end state of a model that can both build a deck and play the game intelligently, I’ll stick with that so I can move onto my robotics projects.

  2. Another fable goal, but this one is simply doing ablations based on several papers to verify that both deck building and game playing is possible and also structuring the rewards in a way that leads to the outcome I want. Original models did show some strategy, but deck building did show right now that models seem to really only prioritize aggresion instead of utilizing the different playstyles that come out of playing different types of decks (gates, leaders, and overall deck composition).

Relevant Papers

Well…

They killed fable.