Machine Learning System Design Interview #13 - The Flat Loss Trap

Why tuning a broken model is pointless - and how the single-batch overfit saves you in every deep learning interview.

You’re in a Senior ML Interview at a OpenAI. The interviewer sets a trap:

“You just implemented a complex Transformer from a new paper. The code runs without errors. The training loop executes. But the loss curve is completely flat. What is your first move?”

90% of candidates walk right into the trap.

Most candidates immediately jump to optimization.

They start listing hyperparameter fixes:

- “I’d lower the learning rate from 1e-3 to 1e-4.”

- “I’d swap AdamW for SGD to stabilize convergence.”

- “I’d double-check the data normalization stats.”