A Solo Coder Trained an AI for $164 — and Found Hidden Bugs

Most AI models are built with Python and a toolkit called PyTorch, the way most furniture is built with a standard screwdriver. One developer decided to build his entire model with Rust instead — a pickier, faster programming language — working alone, with no team, no Python anywhere in the process. The total bill: $164 in rented computing time, spread over about 55 hours on a single high-end chip. The model he produced is tiny by industry standards (roughly 0.4 billion settings, versus hundreds of billions in ChatGPT-class systems) and tuned for Bengali, a language with far less AI support than English.

The real value of the experiment isn't the model — it's the bugs he found along the way. Two leading Rust AI toolkits, Candle and Burn, each had failures that produced no error message. In one case, part of the model simply stopped learning while everything looked normal. In another, the training step ran at about 3% of the chip's expected speed. In a third, the program crashed mid-training at large scale. Loss curves — the standard dashboard people watch to confirm training is working — showed nothing wrong in any of these cases. He only caught them by running a small test that checks every part of the model is actually receiving a learning signal.

He also found a text-splitting trap. Before an AI reads text, software chops it into chunks called tokens. Bengali script, full of stacked symbols, got chopped into tiny fragments — about 1.4 characters per chunk, versus 3.9 for English. That meant his Bengali data was being weighted far heavier than intended, silently tipping the balance of what the model learned. Fixing it brought the ratio to about 4.1.

His honest conclusion: after this run, he moved training back to PyTorch and kept Rust only for running models on devices. In his hands, Rust isn't yet a competitive place to train an AI — though it may be a good place to serve one.