Andrej Karpathy

Featured Work

first-party slide deck — 2023-05-23

Why Now

He is not a pure CV pick anymore, but he is still highly relevant as a translator between modern foundation-model systems and the earlier end-to-end vision stack he helped build at Tesla.

Key Ideas

• -Modern AI systems are increasingly better understood as software stacks and operating systems for tokens, tools, and data pipelines.
• -A lot of frontier progress comes from system design, data curation, and interface choices rather than isolated model novelties.
• -The practical future of multimodal intelligence may belong to agents that compose models, memory, retrieval, and tools cleanly.

Open Questions

• ?What lessons from large-language-model systems transfer back into vision and autonomy stacks?
• ?How much of progress now depends on infrastructure and product interfaces rather than on new architectures?
• ?If multimodal models become general-purpose computers, what remains distinctively computer vision?