On the sidelines of the keynote, HT spoke with Angela Sun, Director of Product, Gemini apps, to get a better understanding of how the AI assistant vision was fine-tuned over time, key technical challenges and ethical considerations, as well as the user experience that is undergoing a significant transformation. Sun gives credit to the Ironwood silicon, the seventh-generation Tensor processing unit, or TPU, optimised for large-scale AI inference. “It is what fuels all of these AI advancements. Our focus on optimisation and efficiency from a hardware and silicon standpoint every day. It really is the backbone of AI,” she says. Edited excerpts.

Q. The vision that we have seen for a Universal AI Assistant, is nothing short of astounding. Could you give us a sense of how this vision was fine tuned over time. And in that regard, were the two Projects (Astra and Mariner) being primed for this moment in time?

Angela Sun: Our vision is really this true manifestation of the AI assistant. Everyone heard for the first time today, an evolution of that and a further definition of what that means for us. Our vision really is to make Gemini the most personal, proactive and powerful assistant out there. That is, helping users with their everyday life. As everyone would think about our roadmap and the things such as Project Astra and Project Mariner, it is about how to work on with users and also on users’ behalf to really bring those three P’s together to really deliver value.

Q. Can you elaborate on the key technical challenges and ethical considerations involved in making Gemini truly universal, and how will Google address these evolving challenges and concerns?

AS: I think one of the top things for that is user feedback and continuing to iterate on that feedback. And here I’ll point to our AI principles which is really to be both bold as well as responsible. And so what does bold mean? It means that we can to innovate. We continue to push the boundaries of this technology and as everyone heard from a lot the announcements at I/O 2025, how we really frame it. Here’s Project Astra, here’s Project Mariner, and here are these technologies that are just in these nascent research prototype stages and we have this group of the trusted tester program where we really try to test and understand the both the strengths of the technology as well as the limitations. But then you can evolution. And I think this is really part of Sundar’s keynote.

How does that evolution work? How does Astra turn into a more generally available product like Gemini Live? How does Mariner turn into a more generally available product? And so going through that evolution and having that, and I will say, Google’s very transparent about what it has in research. Then seeing that life cycle, which sometimes can take months, sometimes a little bit longer, helps with that transparency. Not just publicly, but also, as I mentioned, continuing testing that we do with users and testers, is really important for us.

Q. What are the underlying architectural or training innovations that enabled this level of advanced reasoning with Deep Think, particularly with mathematics, code, and multimodality?

AS: I would say that coding is definitely a big one that we focus on, but it also just goes back to a lot of the evaluations that we do on these models, and so those continue to grow and evolve as well. I believe math, coding, multimodal are some of the headliners that you saw today. But evaluation sets are growing and evolving continuously as we move forward with this technology as well.

For deeper thinking, or as such with the Gemini 2.5 models, is really just a stronger, more powerful LLM. Architecturally it is consistent with the large language model architecture, but it is able to have that capacity to show you its thought process throughout 2.0. And the flash models are a smaller size. The size of the model and the efficiency of the model definitely matters. For our more efficient and optimised models, we do say, they are more for simple everyday tasks and queries and if you want a stronger model, you do have that availability to use one of the deeper thinking models which are more capacity intensive but will show you that thought process.

Q. With Gemini Live camera and screen sharing, Workspace and Chrome layers, and Android XR, the user experience is undergoing a significant transformation. How challenging is it to ensure these new capabilities are intuitive?

AS: That is a very important every day question we ask ourselves. And especially as user behaviour changes, adopting these new technologies, which may not feel intuitive at the beginning because it’s never existed before, we try to make it as seamless as possible. I think one of the beauties of AI powered assistants is natural language. This really was not as prevalent, as it’s been in the last few years. And so how do you make things as simple as a prompt? How do you make things really be intuitive to how people naturally speak, whether or not that’s from a language standpoint or from a stylistically, how people speak. I have two little kids, my five year old speaks to Gemini very differently than how I speak, even though it’s the same English language. And so making sure and really anchoring on that natural language, the natural interaction I do feel is an advantage in today’s technology.