Case context: A startup team is tackling a new speech recognition problem. The lead engineer insists on spending three months researching the theoretical optimum neural network architecture before writing any code, arguing that this will save time in the long run.

Question: As a machine learning consultant, what is the flaw in the lead engineer's plan, and what alternative approach should you recommend based on Andrew Ng's advice?

Sample answer: The flaw in the engineer's plan is the assumption that one can theoretically determine the best ML approach in advance without empirical testing. Andrew Ng notes that for a new problem, it is very difficult to know in advance what will work best, and even experienced researchers try dozens of ideas. The recommended alternative is to quickly build an initial system to establish a baseline and then iterate through ideas based on empirical results.

Key points:

• It is practically impossible to know the best ML approach in advance for a new problem.
• Even experienced practitioners try many dozens of ideas.
• The team should build an initial system quickly and iterate rather than relying on theoretical upfront design.

Rubric: The response must diagnose the flaw (assuming the best approach can be known in advance) and recommend the alternative (rapid initial implementation followed by iteration) citing the unpredictability of ML approaches.