Koray Kavukcuoglu on frontier models, coding agents, and building AGI

00:00:00: There's always a lot of discussion about is there a test for AGI?

00:00:04: There's no test for AGI.

00:00:05: I don't think anyone has a test that they can say, okay,

00:00:08: if you get this much on this test, that means you are at AGI.

00:00:12: And I don't think that is—

00:00:13: - If you do have that benchmark, send it to us, please,

00:00:14: because we want to— we'll use it, we'll run our models on it.

00:00:24: Hey, everyone, welcome back to Release Notes.

00:00:26: My name is Logan Kilpatrick.

00:00:27: I'm on the Google DeepMind team.

00:00:28: Today we're joined by Koray Kavukcuoglu.

00:00:30: Koray, I'm super excited for this conversation.

00:00:32: We sat down probably three or four months ago,

00:00:35: and so I'm excited to catch up,

00:00:37: reflecting on a lot of the progress that's been made over the last

00:00:41: few months, looking into the future, Gemini 4 on the horizon.

00:00:45: The third iteration of the 3.5,

00:00:49: post-3.5 series, lots of progress in literally only three weeks.

00:00:55: We launched 3.5 at I/O, 3.6, and then three weeks later landed 3.7.

00:00:59: Reception has been super positive.

00:01:01: Congrats to you and the team.

00:01:02: Yeah, it's been positive. Do you want to— I don't know if there's

00:01:04: anything you want to say about what made that model.

00:01:07: We teased a little bit of

00:01:10: architectural innovation or something like that,

00:01:13: which I'm not sure we'll spill the beans on, but—

00:01:16: - Not really.

00:01:16: - Yeah.

00:01:17: - Going from the initial 3.0 launch,

00:01:21: my reflection is we learned a lot in terms of

00:01:24: understanding what it means to do coding and not just coding, right?

00:01:28: Like what it means to do software engineering, what it means to

00:01:31: work with tools or work with the functions that people use every day.

00:01:35: Basically, turn this whole thing into an agent,

00:01:37: from a model to an agent.

00:01:39: And that transition, I think we were walking around it,

00:01:44: and that was the thing that we wanted to directly tackle.

00:01:47: And of course, software engineering is the most critical domain

00:01:50: and environment that you want your systems to be successful at,

00:01:54: because it's at the root of many different things that you can do.

00:01:58: During that time, we learned a lot in terms of how to

00:02:01: train a model, how to train an agent

00:02:03: that can actually code with you.

00:02:06: And after that,

00:02:06: I think the steps have started becoming faster.

00:02:09: And in any research project,

00:02:11: there are many parallel tracks that is going on at the same time too.

00:02:15: So, what we are trying to do right now is combine all these learnings,

00:02:19: of course, add on top of them, understand, what it means to do

00:02:24: agentic actions and agentic workflows better together with new architectural

00:02:29: improvements, together with new ideas

00:02:31: that we have been working on for a while.

00:02:34: Many of these things that we have done in 3.6, 3.7 go back like a year or more,

00:02:39: and then those start paying off and you converge them into the model.

00:02:44: And it's great to see the quality impact that we got from that.

00:02:48: We were very excited, in the run-up to 3.7,

00:02:51: when we were doing 3.6, of course, we could see

00:02:54: we could see what 3.7 could be or what the next step could be.

00:02:58: And then it came together.

00:02:59: And internally we started enjoying that model a lot.

00:03:03: So that's the parallel track effect that you are going, that you are seeing.

00:03:07: - There's a lot of parallel tracks.

00:03:08: I don't know if we've done this in the past, but a few weeks ago

00:03:10: we publicly said that Gemini 4, we're working on it.

00:03:14: It's the most ambitious pre-training run

00:03:16: that we've done so far, which is really exciting.

00:03:19: People seem excited by it.

00:03:20: I think folks are excited internally.

00:03:21: Lots of research going into this.

00:03:24: Yeah, I'm curious, this has historically been one of our

00:03:28: strengths and the team,

00:03:30: seems incredible and they've always done a great job.

00:03:32: And so I don't know if there's anything

00:03:34: top of mind or anything more you want to say about that.

00:03:36: Yes, it's the most ambitious run, correct.

00:03:39: And yes, so far, touch wood, it's going well.

00:03:44: I am very excited.

00:03:45: The team is very excited.

00:03:47: I look at these as step by step, right?

00:03:50: So, far, everything is looking good.

00:03:52: We are excited.

00:03:53: But it's all about realizing that potential

00:03:56: until we get to a moment that we are actually putting the

00:04:00: model in front of users, first internally, then externally.

00:04:04: I'm always in a very cautiously optimistic state.

00:04:08: - I think maybe to tee you up for a very specific question

00:04:11: there's a meme about whether or not we care about being at the frontier.

00:04:16: I see you in meetings all the time and know what your point of view is

00:04:20: but what's your point of view as far as whether or not,

00:04:22: you know, you think

00:04:23: it's important for us as Google DeepMind and as a team to trying to be at the frontier?

00:04:28: - I was surprised.

00:04:29: I've seen these comments too, and I can understand obviously

00:04:33: our quality of current models are a little bit below the frontier.

00:04:38: - Yeah. - And that's fair enough.

00:04:40: To put it very bluntly, there's nothing

00:04:43: other than being at the frontier that is important for us.

00:04:46: Yeah, that's it.

00:04:48: All our goals, all our focus is always about that.

00:04:51: And when we are making decisions, when we are making prioritizations

00:04:55: in terms of ideas, in terms of what to do and what goals we should follow,

00:05:00: it's all about being at the frontier,

00:05:01: and I'm 100% certain that, we will be at the frontier.

00:05:07: We have an amazing team.

00:05:09: I have so much trust in the team.

00:05:11: There is a great amount of creativity and dedication in the team,

00:05:16: and being part of Google, we have amazing resources available to us.

00:05:23: We have the whole full stack that we can and we are

00:05:27: able to optimize to be able to achieve our goals.

00:05:32: If you think about Google as a company,

00:05:36: it is a company that has always

00:05:38: been at the frontier of investing in the next big technology.

00:05:42: It has never shied from investing in a long-term important project, right?

00:05:50: AI investment in Google has been not the last five years or ten years, right?

00:05:55: Fifteen years ago when there was no AI, Google was investing in AI chips.

00:05:59: - Yeah.

00:06:00: - Not chips only, the AI chips.

00:06:03: We were not even talking about AI at the time.

00:06:06: So, that kind of long-term vision and investment

00:06:09: has always been part of Google,

00:06:10: like the quantum computing, Waymo.

00:06:12: We talk about these things as if it's normal, but these are very,

00:06:18: technologically critical and heavily scientific and long-term ambitious

00:06:23: endeavors that it's in the DNA of Google.

00:06:25: So, that's why I feel I have the whole trust

00:06:29: for us to be able to achieve that frontier.

00:06:32: That's all our goal. That's why we were set up.

00:06:34: That's what the whole team feels.

00:06:36: Yeah, there's nothing less than that.

00:06:38: - I love that. And I think in this role now, that you have leading all of Google DeepMind,

00:06:42: the frontier AI folks, we actually have a frontier AI team,

00:06:46: and the products and other stuff are now part of this story.

00:06:48: And I'm curious— is there anything from this framing of like

00:06:52: the frontier is what we're focused on that you're

00:06:56: excited now that we're bringing these teams together?

00:06:59: I know there'll be a lot of business as usual for a lot of folks,

00:07:01: but anything particularly that you're excited about?

00:07:05: - Gemini, of course, is very much focused on

00:07:09: building AGI and delivering the goals.

00:07:11: At the end of the day, it relies on two things.

00:07:13: One, good execution, right?

00:07:15: It's very detailed.

00:07:16: It's very complex.

00:07:17: To build AGI, something that you

00:07:19: can put in front of people that they can work with.

00:07:22: But also at the same time, it relies on a very large

00:07:25: funnel to get ideas from everywhere.

00:07:28: So by bringing together Google DeepMind this way, what I'm excited about is, as you said,

00:07:34: there's the frontier AI, there's the products to be able to really

00:07:38: orient and steer at least the idea,

00:07:41: the knowledge and the focus towards this.

00:07:44: You don't want to continuously streamline

00:07:46: everything on one because the idea is at the end exploration.

00:07:50: Success

00:07:52: in building AGI still depends on innovation and exploration

00:07:56: and channeling that in the right way through Gemini.

00:08:00: I'm excited about being able to do that more.

00:08:02: The path to building AGI then,

00:08:04: really depends on our interaction with the users.

00:08:10: And when we say users, yes,

00:08:12: some people are using these models in their daily lives

00:08:14: to go through their emails, but some scientists are using

00:08:18: these models to be better scientists to do actual research to invent things

00:08:22: that has never been invented and getting help from the models.

00:08:26: So, the spectrum of usage of these models

00:08:30: covers everything that is happening in the world.

00:08:33: So that usage

00:08:37: is our guide, right?

00:08:38: That tells us what are the problems that are worth solving,

00:08:42: that people want us to solve, and we need to solve those problems.

00:08:45: And that leads us towards AGI.

00:08:47: That's how we have something general that people like interacting with,

00:08:51: that people rely on, that people trust.

00:08:53: - Yeah, I think last time we were talking you had mentioned

00:08:56: this co-building AGI with our customers framing,

00:09:00: which I think has just continued to like it makes so much sense.

00:09:02: It resonates. I think it makes it clear why

00:09:06: the deployment matters so much,

00:09:08: why this feedback flywheel matters so much.

00:09:10: - There's always a lot of discussion about is there a test for AGI?

00:09:15: There's no test for AGI.

00:09:16: I don't anyone has a test that they can say, okay,

00:09:19: if you get this much on this test, that means you are at AGI.

00:09:23: - And I don't think that is— if you do have that benchmark,

00:09:25: send it to us please, because we'll use it, we'll run our models on it.

00:09:29: - All right.

00:09:29: When you look at the history of AI, history of machine learning,

00:09:33: it has always been like that.

00:09:34: It's the progress.

00:09:36: It's the journey that makes the difference, right?

00:09:39: Go back like fifteen years ago, like twenty years ago

00:09:42: or even more, there was always some notion, some concept of

00:09:47: if we achieve these kinds of tasks,

00:09:49: that means that is the true intelligence.

00:09:52: And at every time period, each and every one of those

00:09:56: were the legitimate targets to follow up on at that point.

00:10:00: And that's what leads to this kind of progress that is happening.

00:10:03: And that's the progress in itself, right?

00:10:05: Picking the next goals is also a big part of that progress.

00:10:09: And that's what we are doing, right?

00:10:11: And you can see that the intelligence, from the very early days of AI

00:10:15: being very narrow to now it's becoming very general.

00:10:18: The goals that we are picking are very general goals.

00:10:21: That's why we are building general intelligence.

00:10:24: And I think it's going to continue like that.

00:10:26: You go back twenty years ago, you look at what we are doing right now.

00:10:30: It like, okay, maybe AI, it's already there.

00:10:34: Like, what people were thinking twenty years ago in terms of what

00:10:37: the goals was a lot of things are achieved.

00:10:39: - Yeah.

00:10:40: - But I think the more important thing is, at the end of the day,

00:10:43: building an intelligent entity that you can trust, that you can work with.

00:10:48: And if we can get there,

00:10:50: and I don't think that there is a sudden threshold of that,

00:10:54: it's more of a building up that trust and getting better and better at that.

00:10:57: - Yeah, I'm with you. You mentioned twenty years ago,

00:11:00: I want to go twenty years back.

00:11:02: I don't know, maybe not twenty years back exactly,

00:11:05: but you were Google DeepMind's first deep learning researcher, which is very interesting.

00:11:09: And I want to double-click on that.

00:11:11: You were at Yann's lab with a bunch of other folks.

00:11:14: The deep learning community was quite small at that time.

00:11:17: Previous iterations,

00:11:18: DeepMind CTO leading a bunch of technical efforts.

00:11:21: The story that you were telling about how during the

00:11:24: Google DeepMind due diligence process, Jeff was asking you questions about

00:11:27: random places in the Google DeepMind codebase because you literally reviewed all of the code.

00:11:32: You had to, you had to be the one to explain random things,

00:11:35: which seems like a crazy problem

00:11:37: now in today's world where AI helps us do that.

00:11:39: You've been at the helm and pushing on the frontier

00:11:43: now for a long time specifically inside of Google DeepMind,

00:11:45: before Google DeepMind. And so much of the world has changed.

00:11:49: I'm curious if you want to reflect back on any of those moments,

00:11:53: any of the papers or technical breakthroughs, obviously,

00:11:55: Google DeepMind has done so much that stand out to you as you think back.

00:11:59: - First, of all, thank you very much.

00:12:00: Yes, I'm old.

00:12:04: I think the journey in Google DeepMind has been very exciting.

00:12:08: There was a time at which me and Karol Gregor,

00:12:10: we were at Yann's lab, we joined Google DeepMind at the same time.

00:12:12: I started the deep learning team and it grew from there.

00:12:15: There was no big plan about how all this was going to happen.

00:12:19: - And just to clarify, other folks were just doing,

00:12:21: they were not doing deep learning, they were doing like what?

00:12:24: Like just like RL stuff?

00:12:26: - I think when you're in a startup,

00:12:27: when you are doing research, everyone does everything, right?

00:12:30: There were like three, four deep learning labs at the time.

00:12:33: Bringing that deep learning background

00:12:36: together with Karol into DeepMind, it was an amazing environment.

00:12:40: I remember, I think in research there were

00:12:42: ten, twelve people in research and everyone is working on some really interesting idea

00:12:48: from their own background, like a really interesting group of people.

00:12:53: At some point there was a discussion of,

00:12:55: okay, let's pick an ambitious task that we can corral

00:12:58: the whole research team around and try to show progress on that task.

00:13:03: And that's how Atari started,

00:13:05: the Atari games and trying to train an agent that

00:13:09: can learn by itself

00:13:11: playing those games and being successful in those games.

00:13:14: It was an internal project because Atari at the time, there were RL researchers

00:13:18: who were thinking that this can be a good environment to measure success.

00:13:22: That was the first example of: we are known for these focus

00:13:26: projects in DeepMind and we carry that over to now Google DeepMind as well, that we bring

00:13:30: a diverse group of researchers and engineers together

00:13:34: to solve a hard problem.

00:13:35: That was the first example of where we did that.

00:13:38: A large group of researchers that came together every week

00:13:41: we would measure progress and that's what led to DQN.

00:13:44: That's the first example of how deep learning and RL could come together

00:13:48: and create a successful agent that can learn to achieve a task by itself.

00:13:54: I think it convinced me and it convinced many people

00:13:58: this is actually possible.

00:13:59: That one, games are the right environment that you

00:14:03: can actually study intelligence and it can scale.

00:14:07: And two, at the time, deep learning was already

00:14:10: very popular, deep learning and RL together

00:14:14: is the basis of the recipe for building intelligent agents.

00:14:19: It opened that door.

00:14:22: - Was it that obvious?

00:14:24: It's more obvious in hindsight,

00:14:25: but like it really felt like the path was lit then,

00:14:27: or at least there was a sliver of light at the end of the tunnel perhaps?

00:14:30: - I'm very cautious about these things.

00:14:33: I'm not sure about the path that was lit but we all felt like it was

00:14:36: an important moment because

00:14:38: it gave us confidence that we should continue.

00:14:42: When you look at the history of DeepMind, history of Google DeepMind,

00:14:45: bigger and bigger goals have been achieved in this trajectory, right?

00:14:49: Like DQN was the first step and then

00:14:51: every step after that was bigger and bigger and bigger, right?

00:14:55: Like Go, Chess, being able to do those from scratch,

00:14:59: and then using that knowledge for AlphaFold— all sorts of things, StarCraft.

00:15:04: More and more we started building these agents.

00:15:07: That was very exciting.

00:15:08: - Yeah.

00:15:09: I want to talk about agent stuff, but I'm actually curious

00:15:12: if you contrast maybe some of the stuff that was happening

00:15:18: back then to some of the stuff that's happening now.

00:15:20: And one example of this is you mentioned this before about

00:15:24: a lot of these ideas were purely in the research domain

00:15:27: and it was, “Can you actually prove that this thing is possible?”

00:15:30: And it feels like there's been at least some transition to

00:15:32: we know a lot of it's possible.

00:15:34: We have some amount of the recipe.

00:15:36: It's really just a lot of really difficult research

00:15:39: engineering work to actually bring the thing to life.

00:15:42: And so, I'm curious if you want to talk about

00:15:46: actually bringing the stuff to life now and how

00:15:49: does it feel super different?

00:15:51: - It does feel different.

00:15:52: It's also the same in a strange way because we are still doing RL,

00:15:57: we are still doing deep learning, we are still pre-training

00:16:00: these models with very similar loss functions to what we used to do before.

00:16:05: We are still doing RL pretty much using similar RL algorithms.

00:16:09: The fundamentals of learning did not change.

00:16:12: The fundamentals of optimization did not change,

00:16:15: but the domains kept getting harder, right?

00:16:19: Just as from Atari to chess to go to StarCraft,

00:16:23: the domains kept getting harder.

00:16:25: And the way they get harder is in real life you have ambiguity,

00:16:28: you have depth. You have depth of ambiguity.

00:16:31: That's the biggest change that you need to be able to deal with that.

00:16:33: Language in itself is much richer, much complex, much more multi-domain,

00:16:40: multifaceted than any constrained action space that you can work with.

00:16:45: So, of course, like things became

00:16:47: a lot more complex, a lot more larger scale.

00:16:50: Like the basics of it is still the same.

00:16:52: - Yeah.

00:16:53: - But also at the same time, there's a lot more room

00:16:57: to extract the information and be successful in this domain.

00:17:00: But everything is a lot more open-ended,

00:17:03: a lot more open to interpretation,

00:17:05: and that requires a lot

00:17:07: more intelligence to be able to deal with that to

00:17:09: anticipate what someone is saying for the model

00:17:12: to be able to partner with that person or to respond to that person.

00:17:17: - Yeah, I want to talk about the future and stuff,

00:17:20: but maybe one last question on this.

00:17:22: Anything that's surprising to you as you look back

00:17:26: and I'm sure there's lots of stuff,

00:17:28: surprising new innovations that have come up,

00:17:30: but anything that folks were working on before that

00:17:32: has just turned out to be extremely successful that was not obvious it

00:17:37: was going to be super successful or don't any way you want to take that.

00:17:41: - I take a step back and if you look at it statistically,

00:17:44: it is surprising that something

00:17:48: has done this exponential takeoff.

00:17:50: - Yeah.

00:17:51: - Right? It's not like you have hundreds of these kinds of things where different

00:17:55: scientific or technical domains are in this exponential takeoff stage.

00:18:00: So, there's a little bit of an element of I feel lucky that we are living

00:18:08: in this age and all the knowledge

00:18:11: buildup has happened up until this point.

00:18:13: All the technological progress in terms of chips and internet and data

00:18:17: has happened up until this point that we are the ones living through this.

00:18:23: I quite enjoy that.

00:18:25: But also at the same time, like of course,

00:18:28: it's intense and it's competitive.

00:18:30: - Yeah. - I feel lucky.

00:18:31: I feel surprised that a lot of the things that, twenty years ago,

00:18:35: where we were exploring as students are now actually

00:18:39: some of the biggest practical implications in the world.

00:18:42: And I'm pretty sure a lot of people who were

00:18:44: students at the time are feeling the same way.

00:18:47: - Yeah. Somebody just describes you as somebody who loves the journey.

00:18:52: A, is that actually true?

00:18:54: And B, what that means in the context of what we're doing as Google DeepMind,

00:18:58: because I think it's actually an important framing

00:19:02: to capture your headspace and your belief about what we're doing.

00:19:07: - I guess you have to like the journey to be able to achieve something.

00:19:10: So yes, I like the journey.

00:19:11: As I said, I think progress, especially in technology,

00:19:15: in my experience sometimes happen through these stacked sigmoids

00:19:19: where something happens, something gets unlocked, and then you have

00:19:22: big progress, and then you put a lot of incremental stages on top which actually

00:19:27: builds a lot of your understanding,

00:19:29: which leads to another discovery, and then

00:19:32: a little bit more of those, and then more understanding,

00:19:35: and then more discovery.

00:19:37: I think about that process a lot about what leads to progress,

00:19:41: what leads to these jumps and how to achieve that

00:19:45: in the context of any of the projects that we do.

00:19:47: Because as much as I always like talking about, we are building AGI,

00:19:52: we used to research AGI, we used to write papers about AI,

00:19:56: we used to write papers about intelligence.

00:19:58: And we did that,

00:20:01: it was pure idea to pure science.

00:20:03: Now, we are living in the world that we are actually building AGI

00:20:07: because people are using it.

00:20:09: So you have to build it,

00:20:10: you have to do it with that responsibility.

00:20:13: - Yeah. - Right? With that seriousness.

00:20:15: But at the same time, you cannot let go of the fact that

00:20:19: it is still research and innovation that fuels all this.

00:20:21: That funnel is the critical thing, is the biggest differentiator.

00:20:25: So I always talk a lot,

00:20:27: think a lot about what makes it successful.

00:20:31: What are the elements that will make it successful for us

00:20:35: in the next step and the next step after that and the next step after that?

00:20:39: And that's why I think that journey is very important,

00:20:43: but the goal and the mission,

00:20:44: that north star is extremely important, right?

00:20:46: Like, journey only matters,

00:20:49: you know that you are going somewhere good and valuable.

00:20:51: And I feel very strongly and the whole team here feels very strongly

00:20:56: that building AGI and bringing that positive impact to the world,

00:21:00: to our users, to the world, I think is the best thing that we can do.

00:21:05: - Yeah, And you made this comment, I don't know if this was yesterday,

00:21:07: but there's something about doing it specifically at Google,

00:21:11: I think is like a really interesting way.

00:21:14: The mission is obviously important,

00:21:16: but in context of Google and all these products we have

00:21:19: and the broader work that's happening is such a unique way of doing this.

00:21:24: - I think it's the right place.

00:21:26: We talked about Google having the DNA of having,

00:21:29: doing these technological investments, but the other part is Google

00:21:34: has very large number of multi-billion people products.

00:21:39: This is a company that really identifies itself by

00:21:42: providing good services and good value for its users.

00:21:46: Yeah. And understands that, and users like using

00:21:50: our products, and that is really valuable.

00:21:53: And as I said,

00:21:53: I see that as a critical element of building AGI.

00:21:57: The more we can put Gemini in a place where we understand and learn

00:22:01: and get feedback from that interaction, the better we will be.

00:22:06: And also, being in an organization

00:22:10: where bringing value to the user is the utmost important thing.

00:22:15: That's why I think it makes it the right place.

00:22:19: - Yeah. In the context of us being very user-centric,

00:22:22: lots of users asking, where's Gemini 3.5 Pro?

00:22:26: So I'm curious if you can give us a quick spiel on where the model is.

00:22:31: As much as we're willing to say.

00:22:34: - Look, as I said,

00:22:35: there are parallel tracks. In our model family,

00:22:38: we always have the Pro, the Flash, and the Flash-Lite.

00:22:41: But at the same time,

00:22:42: we are seeing really fast progress with our Flash models.

00:22:45: We are working on 3.5 Pro at the same time because it's one of those parallel tracks,

00:22:49: because research happens in multiple scales at multiple times.

00:22:54: Right? But also in terms of bringing the best thing forward for the users,

00:22:58: we are seeing really good progress with Flash, we can iterate really fast.

00:23:02: And I can see that like

00:23:05: we talked about from 3.5 to 3.6 to 3.7,

00:23:08: we have a Flash model that is getting better and better

00:23:10: and more and more competitive, getting closer to that frontier.

00:23:14: - Yeah. - So, that is very exciting

00:23:16: for all the researchers internally as well.

00:23:18: So, we are just using that as our progress path right now.

00:23:22: And 3.5 Pro, it's still like, we are still working on it.

00:23:26: But we are working on Gemini 4 as well at the same time.

00:23:29: And

00:23:31: there's a lot of exciting progress happening there too.

00:23:33: - I love it. I'm excited.

00:23:34: Another one that was top of mind,

00:23:36: Gemini 3 felt like a huge moment for us.

00:23:39: We were at the frontier.

00:23:40: And then I think interestingly, it was almost like the frontier

00:23:44: shifted and where there was a lot of new emergent capabilities

00:23:48: and all that stuff happened to be in

00:23:50: agentic coding and all these other things.

00:23:51: And I think we were just starting to rev the engine

00:23:54: and we were at the frontier in that moment and then

00:23:57: the landscape shifted and how you think about that.

00:24:00: - I think there are two things that we need to keep in mind.

00:24:04: One is by definition in a very competitive environment,

00:24:08: the frontier will always shift.

00:24:10: There will be ebbs and flows of different things.

00:24:13: And the cadence and the frequency of

00:24:17: which lab is producing their most capable model is going to change.

00:24:21: That's number one.

00:24:22: But number two is a fair point.

00:24:24: And we talked about that in the context of 3.5.

00:24:27: I think we learned a lot about agentic actions and agentic workflows

00:24:31: and being able to bring that to life, in a model.

00:24:35: That's the process that we went through where I'm feeling right now,

00:24:38: I'm feeling very comfortable and good right now where we are and our

00:24:43: capability of understanding what users need when they are working with an agent

00:24:49: that is partnering with them on any agentic task and workflow.

00:24:53: But we went through that process of building that.

00:24:56: - I love that. I'll ask you one more, which is a little bit

00:25:00: hypothetical but if you could wave your magic wand and get

00:25:04: the models to do anything and not need to spend a lot of time and energy,

00:25:07: do you have anything on like the top of that list that you would want

00:25:10: like a capability better behavior on something?

00:25:15: It can be small or big.

00:25:17: - I don't know. I if I had a magic wand I would just make them more intelligent.

00:25:23: As the models get more intelligent,

00:25:24: they do everything better, everything more intuitively.

00:25:28: And I think that will be excellent.

00:25:31: - Yeah, I love that.

00:25:32: I sat down with Sergey at I/O, I think 2025 or something like that.

00:25:37: I made this comment to him.

00:25:38: This was after you and I were sitting next to each

00:25:40: other listening to him and Demis do their talk.

00:25:43: And I made the comment to Sergey that I/O and just like some of these

00:25:47: moments of bringing people together, makes me feel the warmth of humanity.

00:25:50: And I was saying that in context of

00:25:52: some of the conversations that we were having.

00:25:55: And so, I feel like this, it's just exciting to see,

00:25:58: or I'm excited to see you take this leadership role

00:26:01: because I feel that warmth of humanity

00:26:04: from you at the same time that I feel

00:26:06: the progress of like your focus on the frontier.

00:26:09: And so on a personal level, I'm excited and I feel like

00:26:13: we've got the things going in the right direction.

00:26:15: I feel like you're well positioned to help us push. So...

00:26:19: - Thank you very much.

00:26:20: - Yeah.

00:26:22: No fake, no free praise.

00:26:24: I feel like you earn it. It's a lot of fun.

00:26:26: Koray, this was an awesome conversation.

00:26:28: Thank you for sitting down and chatting about all this stuff.

00:26:31: I'm excited. Hopefully we'll have

00:26:31: more to talk about soon.

00:26:34: And thank you everyone for tuning in to this episode of Release Notes.

00:26:38: We'll see you in the next one.