Then have fun trying to debug an LLM and realising that not only are there billions of unlabeled variables (weights and activations) but that in fact no one kno

Then have fun trying to debug an LLM and realising that not only are there billions of unlabeled variables (weights and activations) but that in fact no one knows the specific purposes of those variables, not even the people who wrote the code and trained the model.