Andrej Karpathy's recent podcast interview is worth your time
Key ideas: agent orchestration over single-session prompting,
AutoResearch loops that remove human researcher from hyperparameter tuning, and a prediction that digital transformation leads while physical robotics lags by years
His take on open source (~6-8 months behind frontier) being a healthy power balance is worth sitting with
"Centralization has a very poor track record." Hard to argue
