Background
Erhan is a long-standing deep-learning researcher who moved into generative video at Google and then Google DeepMind, shipping the Veo line (he announced Veo 2 publicly in December 2024) and subsequently the Omni video work. His public posture is that of a research lead who has been through several generations of the same measurement problem.
The August 2026 AI Engineer panel with Nicole Brichtova and Shane Gu is the richest public source on how his team actually judges video. The stack he describes has four layers: automated metrics where the task is objective (OCR for text rendering), large-scale human evaluation for subjective quality, live experiments for small differences, and trusted testers to catch workflow-breaking regressions. When two models score the same, the team sits in a room, watches clips side by side and votes — and treats that vote as the decision, not a ceremony. He also recounts a reward-hacking failure that only trusted testers caught: the model adding wedding rings to every hand.
What they run now
Video generation research: the Veo family and the Omni video line, and above all the evaluation of video, which he describes as unsolved. His team's evaluation stack mixes automated metrics, large-scale human evaluation, live experiments and a trusted-tester programme.
What they buy
The clearest video-side signal available. Stated in public: professionally shot footage over scraped web video, and a human-evaluation layer that currently consumes internal staff time (a room of people voting). The buyable products are (1) licensed or commissioned professional footage with production metadata, (2) a structured video preference panel staffed by cinematographers, editors and colourists that replaces the room-of-people vote with a repeatable graded protocol, and (3) trusted-tester-style diagnostic programmes that surface reward-hacking artefacts before release. Vendor relationships not public.
Career
- —Research scientist, Google Brain / Google Research — Dates not verified
- presentResearch lead, video models (Veo, Omni), Google DeepMind — Announced Veo 2, Dec 2024
On the record
High-quality data is the constraint, not volume; professionally shot content is preferable to random YouTube video. source
Evaluating video is the hardest open problem in this work. source
When two models score similarly the team watches clips side by side in a room and votes on preference. source
Spurious behaviours such as adding wedding rings to every hand were caught by trusted testers, not by metrics. source
