AI is the new gradient operator.
The loop isn't new — iterate, test, update. That's been engineering forever.
What's new: AI can finally compute the gradient when you plug it into pretty much *any* system.
Read the failure, understand why, write the fix. That step used to be reserved for human intelligence.
It works when the oracle is clear (compile Linux in
@AnthropicAI 's ccc , or optimize a metric in
@karpathy 's autoresearch).
The only place it breaks is when the loss is fuzzy.
The human's job now shifts from computing the gradient to defining the loss landscape.