Host: And welcome back for those on the stream and those in person. We tend to basically take these longer sessions between all the sort of mainstage keynotes to reflect on things that are particularly important but don't have a significant launch moments. Today we're very lucky to have people working on Omni and Veo, Nano Banana, like the world's best generative models here with us. Dumitru, I first saw you when you were posting about your office. I think you're probably Google's number one office influencer at least in San Francisco. I think you like to bike as well. You like to take photos of...

Dumitru Erhan: Bike here.

Host: Yeah. But also you work on video models.

Dumitru Erhan: That's right.

Host: Shane, I met you I think at like a dinner.

Shane Gu: Yeah.

Host: And I remember you were trying to get me invested in like one of the companies. I forget which one.

Shane Gu: Forget about that. [laughter]

Host: But now you're working on Omni Thinking, and just a bunch of other...

Shane Gu: Gemini RL.

Host: Yeah. And Nicole also, the rest of the gen media models, Nano Banana and everything you just launched actually even this week.

Nicole Brichtova: Yeah. We launched some APIs.

Host: Yeah, yeah.

Nicole Brichtova: And I haven't tried to convince you to invest in anything, but maybe I should.

Host: I try not to be an investor, people just convince me anyway. For those of us who are not working at a Frontier Lab, this is the closest we'll ever get. So yeah, actually let's kind of recap since you're closest to it and we just did it. What was launched this week? What should people go try out?

Nicole Brichtova: Yeah. So yesterday we had two launch moments. One of them we launched Nano Banana 2 Light which has amazing quality and it's super fast. And the second thing is that we brought the Nano Banana 2 API to Google AI Studio, which means that developers can now build with it. Before that, you were only able to use it in Gemini or on Vertex, and now any developer can just go and build with it on AI Studio.

Host: Yeah, that's incredible. So when you guys launched Omni for the first time... That is my favorite use case. Everybody should do that. I got a cat which is probably like the most interesting use case.

Shane Gu: Furer is the number one guy you should follow for ideas on okay what can this thing do.

Host: Yes.

Shane Gu: He's amazing at that.

Host: I've tried to get him for the last two years to come to AIE. He hasn't made it yet. He's actually coming up with amazing demos.

Shane Gu: I know.

Host: I want to say his real name but I can't say his real name.

Dumitru Erhan: One of the cool things he did was a live camera demo with his cat, with real-time audio and vision.

Host: Right, he asked what the cat was thinking and the model responded in character.

Host: So obviously the hero capability of the model, or maybe there's two: one is the ability to kind of take in live video and do end-to-end streaming comprehension. Can you talk about the technical challenges behind building an omni-modal system versus traditional cascading pipelines?

Shane Gu: The key difference is the native end-to-end architecture. In traditional cascading systems, transcribing speech to text loses prosody, emotion, tone, and background audio. Then TTS is another disconnected step. An omni-modal model processes tokens across modalities directly, enabling real-time nuance and low-latency interaction.

Dumitru Erhan: In generative video like Veo, we also adopt native multimodal modeling. Video generation must preserve temporal consistency and visual fidelity, while understanding physical dynamics and enabling joint audio-visual generation.

Host: Shane, you've worked on RL for a long time. Now the industry is discussing reasoning models and RL. What role does reinforcement learning play in multimodal generation and Omni interaction?

Shane Gu: RL is evolving from games and robotics into a core driver for reasoning and multimodal behavior. At the intersection of vision and language, RL uses reward signals to guide multi-step reasoning. For generative media, feedback mechanisms aligned with aesthetic and physical plausibility are becoming crucial for controllability.

Host: Evaluation is known to be the hardest part in multimodal models. Text has deterministic benchmarks, but how do you evaluate video/image generative quality, aesthetics, and interactive UX?

Nicole Brichtova: We do a tremendous amount of human evaluations. Traditional automated metrics like FID are no longer sufficient to measure frontier model quality. Subtle lighting differences, composition, natural anatomy, and unintended artifacts require holistic evaluation from professional creators.

Dumitru Erhan: As these tools mature, the creative paradigm will fundamentally shift. Models won't just output isolated clips; they will act as collaborative partners throughout the workflow—from storyboarding and style adjustment to multi-track audio synthesis.

Host: Great discussion. Thank you everyone!