I just trained a physics-based earthquake forecasting model on a $1000 GPU
7 コメント
Right. Because compute power and/or a physics based model is the limiting factor for accurately predicting when a seismic event happens. Training on historic data is hardly the problem that need's solving.
It's the leading indicators that are actually measurable that are missing. You know the ones that allow for evacuations and other protective measures.
It's the leading indicators that are actually measurable that are missing. You know the ones that allow for evacuations and other protective measures.
This is wild — optimizing I/O and memory flow instead of brute-forcing with clusters is exactly the kind of rethink AI infrastructure needs. You basically inverted the whole scaling narrative. Curious if the zero disk reads trick could generalize to other physics-heavy domains (fluid sims, EM propagation, etc.) or if it depends on the dataset’s uniformity. Either way, killer proof that smarter beats bigger.
This is interesting. Can you share the model/github?
Given the swath of sensors that Japan has, and the long history of a lot of them. I do wonder what the result of training off their datasets would be.
So when/where is the next big one coming?
How do I get access to this?
How are you doing your train/test split?
Pre-loads all 15GB of training data into RAM at startup Zero disk reads during training (that's the bottleneck everyone hits) Uses only 0.2GB of VRAM somehow Trains 40 epochs in under 3 hours Best validation Brier score: 0.0175
For context, traditional seismic models get Brier scores around 0.05-0.15. Lower is better.