The claim that 96% of Claude and GPT-5's weights are “literally useless”
The post cites a 2019 pruning experiment that it says showed 96% of a neural network's weights could be deleted without performance loss, and extends that argument to Claude and GPT-5.
TLDR
A user sharing a video they describe as an 18-minute scaling explainer argues that most of Claude and GPT-5's weights—the numerical parameters inside a neural network—are unnecessary. Citing what they describe as 2019 MIT research, they claim a much smaller “winning lottery ticket” network does the real work. The post uses that argument to claim AI labs waste 90%-plus of their Nvidia budgets.
Combined views
733.3K
2 Sources, first seen ago