Generative Vision Interview Questions, issue 11, Jun 19, 2026
The Concatenation Trap
Senior AI Engineer interview at OpenAI, and the interviewer asks:
“You’re building a diffusion transformer. How do you inject the timestep and class condition into each block, and why?”
Don’t say: “I’ll concatenate the condition to the input tokens”
Why feeding timesteps as extra tokens silently burns quadratic compute, and how adaptive modulation retunes your entire diffusion block for free.
The full answer, with the mechanism and the arithmetic, is for paid subscribers on Substack.