Generative Vision Interview Questions #6 - The Time-Arrow Inversion

Why treating Flow Matching as just a continuous diffusion model exposes you as an architecture tourist, and how flipping the timeline changes the entire training objective from passive denoising to a

You’re in a Senior AI Engineer interview at Anthropic. The interviewer sets a trap:

“How does the formulation of time and noise in Flow Matching fundamentally differ from standard DDPMs?”

90% of candidates walk right into it.

Most candidates say Flow Matching just swaps out the score network for a vector field. They mumble something about continuous ODEs …