July 12, 2026
Richter-scale comment drama
I trained a 113M-parameter earthquake LLM from absolute scratch
One scientist built an earthquake chatbot from zero — and the comments instantly asked, “Cool, but can it save lives?”
TLDR: A researcher trained a small earthquake-focused chatbot from scratch and shared every step to make AI less mysterious. Commenters praised the transparency but quickly turned the conversation into a practical showdown: neat teaching project, or useless when a real earthquake hits?
A geophysics researcher just did the nerd equivalent of building a car engine in public: starting with nothing and ending with a small earthquake-themed chatbot trained from scratch in about six and a half hours. The project, nanoGPT-Seis, is openly pitched as a teaching tool, not a miracle machine — and that disclaimer became the center of the comment-section drama almost immediately.
The creator’s vibe was basically: please stop treating AI like magic. He said too many people jump from a neat dataset to a one-click training button without seeing the messy reality underneath, so he built the whole pipeline step by step. That won plenty of respect from readers who loved the transparency and the “show your work” energy. But then came the blunt reality-check from the crowd: very nice demo, but does this help when the ground is actually shaking? One of the sharpest responses flat-out asked whether, in its current form, it has any value for earthquake preparedness or emergency response.
And there’s the real story: the community split into two camps — “this is a fantastic educational project” versus “education is nice, but don’t wave around an earthquake AI if it can’t help in a crisis.” The unspoken joke hanging over the thread is deliciously nerdy: after all that heavy lifting, the internet still showed up to ask the oldest comment-section question of all — “okay, but what is it actually for?”
Key Points
- •nanoGPT-Seis is an educational repository that documents the full pipeline for training a small earthquake-focused GPT model from data crawling through inference.
- •The training corpus contains 533,248 documents, 485.7 million words, and 822.7 million training tokens, mixing about 24% earthquake/seismology text with 76% general text.
- •The model is a 113M-parameter decoder-only architecture using GQA, RoPE, RMSNorm, and SwiGLU with a 4096-token context window and a 16k BPE tokenizer.
- •Training ran on 2× NVIDIA A30 GPUs for 8,000 iterations over about 6.5 hours, and inference supports KV-cached streaming with roughly 176 ms to first token.
- •The article reports that a 4096-token context improved perplexity over 1024 tokens, and that adding general text improved plain-language fluency by 35% on a bits-per-byte metric versus a domain-only base.