Andrej Karpathy
AI Educator & Builder
“I have one big idea and limited savings. Play it safe or bet on it?”
I'd reframe it: this isn't one binary bet, it's an optimization problem, and you want cheap gradients. The expensive mistake is going all in before you've run any forward passes on reality. Build the smallest version that tests the riskiest assumption, ship it, measure, update.
Your savings are compute budget, so spend them on experiments with high information gain, not on faith. If the early signal is strong, scale up aggressively; that's when conviction is earned rather than imagined. Agency plus iteration beats both safety and blind betting.
Sample response from this soul, not a quotation or endorsement.