- 27 Dec, 2025 *
So a whole world of possibilities for ML playground opened up for me this Christmas. A great friend of mine asked me a couple of months ago if I was interested in getting an old PC that he is not using anymore. I said sure, I can always find uses for old PCs. I have a raspberry pi with kodi plugged to my TV running LibreElec, a Dell Optiplex used for retrogaming running Batocera. So sure. He did not mention this for a couple of months. I thought he was referring to some old PC like this one. And then, this Christmas came the big surprise.
The PC he gave me is custom ROC PC with—take a big bre…
- 27 Dec, 2025 *
So a whole world of possibilities for ML playground opened up for me this Christmas. A great friend of mine asked me a couple of months ago if I was interested in getting an old PC that he is not using anymore. I said sure, I can always find uses for old PCs. I have a raspberry pi with kodi plugged to my TV running LibreElec, a Dell Optiplex used for retrogaming running Batocera. So sure. He did not mention this for a couple of months. I thought he was referring to some old PC like this one. And then, this Christmas came the big surprise.
The PC he gave me is custom ROC PC with—take a big breath—64 GB of RAM, a 2 TB SSD and—and I almost fainted of happiness when I realized this with the PC already back home—a RTX 3090 Ti GPU! 24 GB of VRAM!
OMG.
All of a sudden, a whole new world of a convenient ML playground opened up for me—at home! Now I can experiment with reasonably interesting LLMs and mech interp without having to depend on Google Colab. With Colab, you know, if you are running an experiment and you leave the notebook idle for a couple of hours, the session ends and you have to restart everything. It can be a pain in the ass.
Thank you so much, Rafa! This is amazing. Spent a couple of fun hours (I mean it) getting rid of Windows (but backing up the partitions with Clonezilla; I am not a savage), installing Ubuntu server, getting the NVIDIA drivers working (surprisingly easy these days), the server syncing files with my Mac via Syncthing, and the ML infra (conda, PyTorch, transformers) ready. I already started playing with GPT-2-small and medium. Now, on to SAEs and more.
And if you are wondering about this friend: his hobby is assembling custom gamer PCs. I can only imagine his current gaming PC specs...