Jeremy Howard: fast.ai Deep Learning Courses and Research | Lex Fridman Podcast #35

Howard argues the most valuable AI work comes from domain experts applying deep learning to real problems (e.g., medicine, fisheries, media bias), not from abstract benchmarks; you should pair technical skills with a field you deeply understand.

Through transfer learning and careful experimentation, fast.ai repeatedly achieves top performance on modest hardware and smaller datasets, showing most organizations can do serious deep learning without Google-scale resources.

Because Python is slow and forces critical loops into low-level CUDA C, many ideas (e.g., sparse CNNs, novel RNN kernels) are practically un-researchable; Howard is optimistic about Swift + MLIR-style tensor DSLs that let researchers write fast, hackable kernels in a high-level language.

He notes that fields like transfer learning and active learning, which radically reduce data and labeling needs, were largely ignored until someone showed dramatic results, while papers that improve well-studied benchmarks by small margins are rewarded.

Howard stresses that learners should quickly fine-tune pretrained models on personally relevant datasets, inspect inputs/outputs and errors, and iterate; this builds intuition far more effectively than reading theory alone.