Why do we train bigger models instead of finding better neural pathways?
retained text
I saw a post of a guy with 90% of his brain missing and he was apparently living a normal life. There is another case of a student who finished university with half of his brain missing. There are also people with full brains, but they are even less functional than these two. This proves that bigger doesn't mean better. So why not aim to create better connections? I understand that a bigger model means less chance of catastrophic interference occurring, but then are all the neural pathways being formed efficiently? Wouldn't this result in a lot of redundant neurons? Wouldn't it make more sense to create the smallest neural network possible for a specific task and then learn to fuse multiple of these neurons together in order to create an optimized larger model? That way we'd always have a template for each specific task and would allow us to create a programming language that creates a model just by our syntax.
document fetched 2026-10-01 00:30 · canonical https://reddit.com/r/OpenAI/comments/1wucfxv/why_do_we_train_bigger_models_instead_of_finding