Sundar Pichai: belief

Transcript context

…I wonder if you can comment on maybe earlier versions of Gemini were a little bit careful on the kind of things it’d be willing to answer. I just want to comment on I was really surprised, and pleasantly surprised, and enjoy the fact that Gemini 2.5 Pro is a lot less careful, in a good sense. Don’t ask me why, but I’ve been doing a lot of research on Genghis Khan and the Aztecs, so there’s a lot of violence there in that history. It’s a very violent history. I’ve also been doing a lot of research on World War I and World War II. Earlier versions of Gemini were very… Basically this sense, are you sure you want to learn about this. And now, it’s actually very factual, objective, talks about very difficult parts of human history, and does so with nuance and depth. It’s been really nice. But there’s a line there that I guess Google has to walk. And it’s also an engineering challenge how to do that at scale across all the weird queries that people ask. Can you just speak to that challenge? How do you allow Gemini to say… Again, forgive, pardon my French… crazy shit, but not too crazy? I think one of the good insights here has been as the models are getting more capable, the models are really good at this stuff. And so I think in some ways, maybe a year ago, the models weren’t fully there, so they would also do stupid things more often. So you’re trying to handle those edge cases, but then you make a mistake in how you handle those edge cases and it compounds. But I think with 2.5, what we particularly found is once the models cross a certain level of intelligence and sophistication, they are able to reason through these nuanced issues pretty well. And I think users really want that. You want as much access to the raw model as possible. I think it’s a great area to think about. Over time, we should allow more and more closer access to it. Obviously, let people custom prompts if they wanted to and experiment with it, et cetera. I think that’s an important direction. The first principles we want to think about it is, from a scientific standpoint, making sure the models… And I’m saying scientific in the sense of how you would approach math or physics or something like that. From first principles, having the models reason about the world, be nuanced, et cetera, from the ground up is the right way to build these things, not some subset of humans hard-coding things on top of it. I think it’s the direction we’ve been taking and I think you’ll see us continue to push in that direction. I took extensive notes and I gave them to Gemini and said, “Can you ask a novel question that’s not in these notes?”, and it wrote… Gemini continues to really surprise me, really surprise me. It’s been really beautiful. It’s an incredible model. The question it generated was, “You…”, meaning Sundar, “… told the world Gemini is churning out 480 trillion tokens a month. What’s the most life-changing, five-word sentence hiding in that haystack?”. That’s a Gemini question. I don’t think you can answer that, but it woke me up to all of these tokens are providing little aha moments for people across the globe. So, that’s like learning. Those tokens are people are curious, they ask a question and they find something out, and it truly could be life-changing.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.