Miju Labs

← All buyers · Google DeepMind

Dumitru Erhan

Research lead, video models (Veo / Omni), Google DeepMind · Google DeepMind

medium confidenceTier A · Talk to firstFrontier labsvideo

Leads Google DeepMind's video model work in the Veo and Omni lineage. On an August 2026 panel he stated the position that matters most to a data vendor: high-quality data is the constraint, not volume, and he specifically prefers professionally shot content over random web video. He also called video evaluation the hardest open problem in his work.

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

  1. —Research scientist, Google Brain / Google Research — Dates not verified
  2. presentResearch lead, video models (Veo, Omni), Google DeepMind — Announced Veo 2, Dec 2024

On the record

Data quality over volume · 2026-08

High-quality data is the constraint, not volume; professionally shot content is preferable to random YouTube video. source

Video evaluation · 2026-08

Evaluating video is the hardest open problem in this work. source

When metrics tie · 2026-08

When two models score similarly the team watches clips side by side in a room and votes on preference. source

Reward hacking · 2026-08

Spurious behaviours such as adding wedding rings to every hand were caught by trusted testers, not by metrics. source

Recent posts

Dumitru Erhan
Senior Research Director @ Google DeepMind, co-lead of Gemini Omni
Be a Veo PM, work with me to ship Veo stuff to everyone :)
20
Dumitru Erhan
Senior Research Director @ Google DeepMind, co-lead of Gemini Omni
Join the Veo team! See details below:
212 reposts
Dumitru Erhan
Senior Research Director @ Google DeepMind, co-lead of Gemini Omni
Really proud of our team shipping this earlier this week! Veo 3 is spectacular and brings your imagination to life in a whole new way. We can't wait for everyone to try it out ASAP and so TODAY we're expanding access for all Pro users in 70+ countries. Go to gemini.google or flow.google to try it out! More details here: lnkd.in/eS4njHKK
436 comments

Interviews & talks

Why it matters here

Very high for the video half of Julian's thesis. He has said on the record both what data he wants (professional) and where his evaluation hurts (human video judgement at scale). Most vendors sell video annotation as bounding boxes and captions; almost nobody sells trained cinematographic judgement. That is a defensible niche and he has described the need for it without using the word.

How to reach

Show the panel, not the pitch. Speculation on channel: he is active on X as @doomie and on LinkedIn. The artefact that would land: take fifty public Veo and competitor clips, have a panel of working editors and cinematographers grade them on motion, continuity, lighting and artefacting with written rationale, report inter-rater agreement, and show three failure modes their automated metrics would not catch. Then say what a standing panel costs.

What we could not establish
  • Education and pre-Google career not verified in this pass.
  • Exact title (research lead vs director) not evidenced.
  • Whether Google DeepMind licenses professional footage externally and from whom.
  • The AI Engineer panel was read via a secondary summary page, not the primary transcript.

Why they are on the map

Veo co-lead (Google AI Release Notes podcast); discussed video eval and cautioned human preferences are unreliable optimisation target source — evidence dated 2025-10

Needs the data for a model; writes the spec.

How Google DeepMind buys data

Imagen 3 report (2024-08) collected 366,569 ratings from 3,225 raters in 71 countries via an unnamed 'external platform' under GDM data-enrichment policy (living wage). Veo 3.1 page cites human side-by-sides on MovieGenBench (1,003 prompts), VBench I2V and internal sets. Nano Banana leads (2025) describe large-scale human evals plus trusted testers and artist reviews; Gemini Data team job post (2026) covers data curation/generation/evaluation.

Aug 2026: Kavukcuoglu runs GDM incl. Gemini as SVP; Hassabis chair. Gemini has Data, Post-Training, Safety Post-Training and Model Behavior teams (job posts); Gemini app sits inside GDM under Josh Woodward. Gen-media (Imagen/Nano Banana/Veo/Flow) has product lead Brichtova and modality research leads (Wang, Erhan).

Sources

  1. Leads of Nano Banana, Imagen, Veo, Omni/Thinking (AI Engineer panel)
  2. Dumitru Erhan on X, Veo 2 announcement
  3. Dumitru Erhan, LinkedIn
  4. Veo 3.1 model page