Everything Is a Leaky Abstraction

Elijah Potter contends that machine learning abstractions inevitably leak: unlike well-sealed systems such as lossless compression, TCP/IP, TLS, or PostgreSQL, ML requires intimate knowledge of its internals to succeed. Drawing on Andrej Karpathy’s classroom practice of having students implement forward and backward passes by hand, the argument emphasizes that understanding backpropagation and how weights and biases update is essential for debugging and improving models. Every architectural choice changes the shape of the cost function and therefore the learning dynamics, so seemingly high-level components - optimizers, layer choices, normalization, or even using a transformer - cannot be treated as black boxes without risking failure.

The practical takeaway is to slow down and build a deep, mechanistic understanding of models and training regimens before expecting reliable results. Success depends on knowing how design choices interact with optimization, not on throwing compute or popular architectures at problems. Potter urges methodical experimentation and careful study of training behavior, arguing that being right slowly beats being wrong fast when systems exhibit Rube Goldberg-level complexity and subtle failure modes.