Google DeepMind Says Gemini Is Evolving From Chatbot To AI Agent

Google published an question and reply pinch Google DeepMind SVP and Chief AI Architect Koray Kavukcuoglu. Kavukcuoglu shared that Google is progressively viewing Gemini arsenic an AI agent, not a chatbot. This is important because CEO Sundar Pichai has expressed that agentic AI is the early of search.

Gemini Is Transitioning From A Model Toward Agentic AI

Koray Kavukcuoglu shared that Google DeepMind progressively sees Gemini arsenic an agent, overmuch much than conscionable a chatbot aliases connection model. What inspired this alteration is coding, that was the gateway to package engineering, instrumentality use, and agentic workflows.

Kavukcuoglu said that the extremity is not really astir creating a exemplary that answers questions better. The attraction now is to create thing that tin return actions connected behalf of and alongside a human.

The interviewer, Logan Kilpatrick asked Kavukcuoglu if he wanted to talk astir Google DeepMind’s “architectural innovation” and Kavukcuoglu answered that that wasn’t thing he was consenting to talk astir correct now, indicating there’s thing going connected down the scenes.

But what he did talk was really Google’s attraction has transitioned toward agentic AI.

He shared:

“Going from the first 3.0 launch, my reflection is we learned a batch successful position of knowing what it intends to do coding, and not conscionable coding, right?

Like what it intends to do package engineering, what it intends to activity pinch devices aliases activity pinch the functions that group usage each day.

Basically move this full point into an agent, from a exemplary to an agent.

And that transition, I deliberation we were benignant of stepping astir it, and that was the point that we wanted to straight tackle. And of course, package engineering is the astir captious domain and situation that you want your systems to beryllium successful successful because it’s astatine the guidelines of many, galore different things that you tin do.

During that time, we learned a batch successful position of really to train a model, really to train an supplier that tin really codification pinch you. And aft that, I deliberation the steps person started becoming faster.

And for illustration successful immoderate investigation project, location are galore parallel tracks that is going connected astatine the aforesaid clip too. So what we are trying to do correct now is harvester each these learnings, of course, for illustration adhd connected apical of them, understand what it intends to do agentic actions and agentic workflows better, together pinch caller architectural improvements, together pinch caller ideas that we person been moving connected for a while.

Many of these things that we person done successful 3.6, 3.7 spell backmost for illustration a twelvemonth aliases more, and past those commencement paying disconnected and you converge them into the model.

And it’s awesome to spot the value effect that we sewage from that. We were very excited successful the tally up to 3.7. When we were doing 3.6, of course, we could spot what 3.7 could beryllium aliases what the adjacent measurement could be.

And past it came together and internally we started enjoying that exemplary A lot. So that’s the parallel way effect that you are going, that you are seeing.”

AI Is Changing But The Steps Remain The Same

Something other he talked astir seemed contradictory successful that he said revolutionary alteration is happening but what’s causing it is not revolutionary itself. He said that the basal steps for creating AI hasn’t changed astir arsenic overmuch arsenic the problems that AI is solving have. What it does is revolutionary but the measurement it gets location hasn’t undergone a revolutionary change.

Kavukcuoglu said that they are still using:

• Deep learning
• Pre-training
• Reinforcement learning
• Optimization techniques that trust connected akin underlying principles arsenic from the past

What changed, he said, is the situation successful which AI is operating successful which requires models to infer intent, woody pinch ambiguity and collaborate pinch humans.

Google Has Improved Agentic Workflows

Later successful the chat Kilpatrick said that Gemini 3 reached the frontier, but that the frontier subsequently moved toward agentic coding and agent-like capabilities.

Kavukcuoglu said still had things to study astir agentic actions and workflows, saying that the activity surrounding Gemini 3.5 taught them astir really group really activity pinch agents, and that he now feels much assured that Google has a amended knowing of those kinds of interactions.

He explained:

“I deliberation location are 2 things that we request to support successful mind.

One is, by definition, successful a very competitory environment, the frontier will ever shift. There will beryllium ebbs and flows of different things. And the cadence and the wave of which laboratory is producing their astir tin exemplary is going to change. That’s #1.

But #2 is simply a adjacent point. And we talked astir that successful the discourse of 3.5. I deliberation we learned a batch astir agentic actions and agentic workflows. And being capable to bring that to life successful a model, that’s the process that we went through.

Where I’m emotion correct now, I’m emotion very, very, very comfortable and bully correct now wherever we are and our capacity of knowing what users request erstwhile they are moving pinch an supplier that is partnering pinch them connected immoderate benignant of agentic task and workflow.

But we went done that process of building that.”

Google DeepMind: The Most Important Improvement In AI

The interviewer, Logan, past asked Kavukcuoglu to stock what is the astir important betterment successful AI.

He asked:

“…if you could benignant of activity your magic wand and benignant of get the models to do thing and not request to walk a batch of clip and energy, do you person thing connected the apical of that database that you would benignant of want a capability, benignant of amended behaviour connected something?”

Kavukcuoglu answered:

“I deliberation if I had a magic wand, I would conscionable make them much intelligent. I deliberation the models get much intelligent they do everything better, they do everything much intuitively, and I deliberation that would beryllium excellent.”

Takeaway

Google DeepMind, and Google arsenic a whole, is transitioning toward much than providing answers, it’s helping users execute tasks. All of Google’s products, from Gmail, to Google Sheets, to Maps, each thief group execute tasks. Kavukcuoglu explained that AI models are conceptually becoming much than conscionable chatbots, they are besides becoming much agentic, and that reflects really everything astatine Google is changing, including Search.

Watch the question and reply here:

Featured Image/Screenshot of interview