Korva is a scroll of the internet’s best ideas.
People train on 20 billion tokens
Many labs seem very close to solving continual learning and sample-efficient learning. A model trained on 10 trillion tokens and then released into the world to keep learning efficiently would need far less training compute than one trained on 300 trillion.
Even then, training would not drop to zero. Training as a share of semiconductor demand and compute would settle at a small figure, so the change could look like a temporary discontinuity in demand.
Safe Superintelligence is planning to release its model in August, and a whole generation of new labs is focused on this problem. Nvidia is heavily involved with these startups, and it is hard to see the change being negative for AI infrastructure demand.
Why do models need so much training data?
A model learns by making many small adjustments to its weights, and each adjustment captures only a sliver of what is in the data. DeepMind's 2022 Chinchilla paper found that for a given compute budget, the best results came from training on about 20 tokens for every parameter in the model. Larger models therefore call for much larger datasets. Humans pull far more from each example, which is the gap that sample-efficient learning aims at.
Why do models forget when they keep learning?
In 1989, Michael McCloskey and Neal Cohen showed that neural networks trained on a new task could abruptly lose what they had learned on an earlier one. The effect is called catastrophic forgetting. New training moves the same weights that held the old knowledge, overwriting it. Continual learning research tries to let a model add knowledge without erasing what it already has.