Interview
Andrej Karpathy on why we still need humans in the loop
- GUIs are expected to remain "extremely important" for future auditing systems, leveraging computer vision GPUs to facilitate rapid information processing as text-based reading is deemed "effortful."
- AI development plans prioritize keeping AI "on the leash," reflecting a belief that current enthusiasm for AI agents is "way overexcited" and that automated code generation will not remove the human necessity for verifying correctness and security.
- Predictions indicate that even if AI instantaneously generates "1,000 lines of code," humans will continue to serve as "the bottleneck" by personally ensuring outputs are free of "bugs," perform the "correct thing," and avoid "security issues."
- Material expectations suggest that reviewing a "diff of 1,000 lines of code" for a repository will not be "useful" without requisite human oversight to guarantee safety and accuracy.