Shared Residual Blocks
Karpathy explains that residual blocks can share connectivity, so a vision model starts to resemble a recurrent neural network. Lex Fridman's interview keeps the claim technical, with no autonomy slogan attached. The written record supports this trajectory: his 2015 image description paper and DenseCap work on dense captioning show a sustained interest in visual recognition and localization. The point is architectural, not promotional.






