One Guy Trains a Chatbot on Three Old Gaming Chips

A developer going by the username jjusko20 has been training a large language model (AI chatbot) from scratch on his own hardware—three Nvidia V100 graphics cards, each with 32GB of memory. These are older chips, originally designed for data centers but now affordable for hobbyists. His goal: create an AI that can chat naturally and perform tasks (agentic work). He's using a technique called distillation, where a smaller model learns by mimicking a larger one, and he's generating all the training data locally on the same cards.

After his first training round, the model learned to reason step-by-step (chain of thought) and could hold basic conversations. But it was underfit—meaning it hadn't seen enough varied examples. When asked ambiguous or off-topic questions, it produced garbled responses. He decided not to release this checkpoint because it wasn't useful. Instead, he's generating a new, broader dataset of about 5 million tokens focused on general conversation rather than just coding, and will retrain from the first checkpoint with a lower learning rate.

What's remarkable is the scale: training an 80-billion-parameter model (roughly the size of early GPT-3) on three consumer-accessible GPUs. This would have been unthinkable a few years ago. The developer also built an open-source tool called sftmill to make generating training data easier, which he's shared on GitHub. He plans to release a working version of the model once it's ready, and may livestream the training again.

This project shows how AI development is democratizing. You no longer need a million-dollar cluster to train a capable model; with some know-how and secondhand hardware, individuals can experiment. That could lead to more diverse, specialized AI models—and more competition for big tech.