Mountain view, California: The Google Gemini Artificial Intelligence (AI) model has taken an important step forward. In this year Google I/O 2025 Keenot, beautiful Pichai, who is CEO for Google and alphabet, and Damis Hasabis who is CEO in Google Deepmind, revealed the company’s vision for the chatbot, developed as a versatile universal AI auxiliary. Or as some other companies are this trend, AI agents or agent AI. Two of Google’s current projects contributed significantly to Gemini’s planned change-project Estra, to give AI status references, such as video understanding, screen sharing and memory, as well as project marinner, which examines the future of human-agent interactions, starts with browsers.
On the occasion of Kenot, HT spoke with Angela Sun, Director of Product, Mithun apps, so that how to correct AI auxiliary vision over time, major technical challenges and moral ideas as well as user experiences that are undergoing a significant change. Sun, Ironwood Silicon, the seventh -generation tensor processing unit, or TPU credits, adapted for a large -scale AI estimate. She says, “This promotes all these AI progresses. Every day a hardware and silicon’s perspective focuses on adaptation and efficiency. It is actually the backbone of AI,” she says. Edited excerpt.
Q. The view with which we have seen for a universal AI assistant is no less than amazing. Can you understand us how this vision was fine over time. And in that regard, were two projects (Estra and Meriner) being prime for this time?
Angela Sun: Our vision is actually the right expression of AI assistant. Everyone heard for the first time today, one of his development and what it means to us, another definition of it. Our vision is actually to make Gemini the most personal, active and powerful assistant. That is, help users with their everyday life. As everyone thinks of things like our roadmap and project Estra and Project Meriner, it is about how to work with users and to give users really to really bring those three p together.
Q. Can you explain in detail on the major technical challenges and moral ideas involved in making Gemini really universal, and how will Google address these developed challenges and concerns?
As: I think one of the top things for this is a user response and continues to recur on that response. And here I will point to my AI principles which are really courageous and both responsible. And what does bold mean? This means that we can do it new. We carry forward the boundaries of this technique and as everyone has heard a lot of announcements in I/O 2025, how do we actually frame it. Here is Project Estra, here is the project mergeer, and here are these techniques that are only in these newborn research prototype stages and we have this group of reliable tester programs where we actually try to test and understand the boundaries of technology along with both powers of technology. But then you can develop. And I think it is really part of the keynote speaker of Sundar.
How does that development work? How does Estra Gemini Live like how to turn into a more commonly available product? How does Mariner more commonly turn into a available product? And so going through that development and so, and I would say, is very transparent about Google what it is in research. Then by looking at that life cycle, which can take months sometimes, sometimes it is slightly longer, helps in that transparency. Not only publicly, but, as I mentioned, it is really important to us to continue the test which we do with users and examiners.
Q. What are the built -in architectural or training innovation that enable this level of advanced logic with deep ideas, especially with mathematics, code and multimoda?
As: I would say that coding is certainly a large that we focus on, but it just goes back to a lot of evaluation we do on these models, and so they keep growing and develop. I believe that mathematics, coding, multimodal are some headliners that you saw today. But evaluation sets are growing and constantly developing because we move forward with this technique.
For deeper thinking, or as Gemini 2.5 model, is actually just a strong, more powerful LLM. Architecturally, this large language model corresponds to architecture, but it is capable of the ability that is capable of showing you its idea process in 2.0. And the flash models are of a small size. The size of the model and the efficiency of the model definitely matters. For our more efficient and customized models, we say, they are more for simple everyday tasks and questions and if you want a strong model, you have availability to use one of the deep -minded models that are more intense but will show you that idea process.
Q. Gemini live camera and screen sharing, workpace and chrome layers and Android XR, user experience is undergoing a significant change. How challenging is it to ensure that these new capabilities are comfortable?
As: This is a very important thing, we ask ourselves the question of every day. And especially the user’s behavior changes, adopting these new techniques, which cannot feel comfortable in the beginning because it has never existed before, we try to make it as comfortable as possible. I think AI is a natural language of the beauties of assistants. It was not really as prevalent as in the last few years. And so how do you make things simple as a sign? How do you really make things how people speaking naturally, whether it is from the language point of view or not, how do people speak. I have two young children, my five -year -old child speaks very differently to Mithun how I speak, even if it is the same English language. And so to ensure this and actually anchoring on that natural language, I think the natural conversation I think is an advantage in today’s technology.