Google revealed an interview with Google DeepMind SVP and Chief AI Architect Koray Kavukcuoglu. Kavukcuoglu shared that Google is more and more viewing Gemini as an AI agent, not a chatbot. That is vital as a result of CEO Sundar Pichai has expressed that agentic AI is the way forward for search.
Gemini Is Transitioning From A Mannequin Towards Agentic AI
Koray Kavukcuoglu shared that Google DeepMind more and more sees Gemini as an agent, rather more than only a chatbot or language mannequin. What impressed this variation is coding, that was the gateway to software program engineering, software use, and agentic workflows.
Kavukcuoglu mentioned that the purpose is just not actually about making a mannequin that solutions questions higher. The main focus now could be to create one thing that may take actions on behalf of and alongside a human.
The interviewer, Logan Kilpatrick requested Kavukcuoglu if he wished to speak about Google DeepMind’s “architectural innovation” and Kavukcuoglu answered that that wasn’t one thing he was prepared to speak about proper now, indicating there’s one thing happening behind the scenes.
However what he did talk about was how Google’s focus has transitioned towards agentic AI.
He shared:
“Going from the preliminary 3.0 launch, my reflection is we discovered rather a lot by way of understanding what it means to do coding, and never simply coding, proper?
Like what it means to do software program engineering, what it means to work with instruments or work with the features that folks use on daily basis.
Mainly flip this entire factor into an agent, from a mannequin to an agent.
And that transition, I believe we had been form of strolling round it, and that was the factor that we wished to immediately sort out. And naturally, software program engineering is probably the most crucial area and atmosphere that you really want your methods to achieve success in as a result of it’s on the root of many, many alternative issues that you are able to do.
Throughout that point, we discovered rather a lot by way of find out how to practice a mannequin, find out how to practice an agent that may truly code with you. And after that, I believe the steps have began turning into quicker.
And like in any analysis challenge, there are lots of parallel tracks that is occurring on the identical time too. So what we try to do proper now could be mix all these learnings, in fact, like add on prime of them, perceive what it means to do agentic actions and agentic workflows higher, along with new architectural enhancements, along with new concepts that we have now been engaged on for some time.
A lot of this stuff that we have now finished in 3.6, 3.7 return like a 12 months or extra, after which these begin paying off and also you converge them into the mannequin.
And it’s nice to see the standard influence that we acquired from that. We had been very excited within the run as much as 3.7. After we had been doing 3.6, in fact, we might see what 3.7 may very well be or what the following step may very well be.
After which it got here collectively and internally we began having fun with that mannequin Lots. In order that’s the parallel observe impact that you’re going, that you’re seeing.”
AI Is Altering However The Steps Stay The Similar
One thing else he talked about appeared contradictory in that he mentioned revolutionary change is going on however what’s inflicting it’s not revolutionary itself. He mentioned that the fundamental steps for creating AI hasn’t modified almost as a lot as the issues that AI is fixing have. What it does is revolutionary however the way in which it will get there hasn’t undergone a revolutionary change.
Kavukcuoglu mentioned that they’re nonetheless utilizing:
- Deep studying
- Pre-training
- Reinforcement studying
- Optimization methods that depend on comparable underlying rules as from the previous
What modified, he mentioned, is the atmosphere during which AI is working during which requires fashions to deduce intent, cope with ambiguity and collaborate with people.
Google Has Improved Agentic Workflows
Later within the dialogue Kilpatrick mentioned that Gemini 3 reached the frontier, however that the frontier subsequently moved towards agentic coding and agent-like capabilities.
Kavukcuoglu mentioned nonetheless had issues to study agentic actions and workflows, saying that the work surrounding Gemini 3.5 taught them about how individuals truly work with brokers, and that he now feels extra assured that Google has a greater understanding of these sorts of interactions.
He defined:
“I believe there are two issues that we want to bear in mind.
One is, by definition, in a really aggressive atmosphere, the frontier will all the time shift. There will probably be ebbs and flows of various issues. And the cadence and the frequency of which lab is producing their most succesful mannequin goes to vary. That’s #1.
However #2 is a good level. And we talked about that within the context of three.5. I believe we discovered rather a lot about agentic actions and agentic workflows. And having the ability to carry that to life in a mannequin, that’s the method that we went via.
The place I’m feeling proper now, I’m feeling very, very, very snug and good proper now the place we’re and {our capability} of understanding what customers want when they’re working with an agent that’s partnering with them on any sort of agentic job and workflow.
However we went via that strategy of constructing that.”
Google DeepMind: The Most Necessary Enchancment In AI
The interviewer, Logan, then requested Kavukcuoglu to share what’s an important enchancment in AI.
He requested:
“…in case you might form of wave your magic wand and form of get the fashions to do something and never want to spend so much of time and power, do you may have something on the highest of that record that you’d form of desire a functionality, form of higher habits on one thing?”
Kavukcuoglu answered:
“I believe if I had a magic wand, I might simply make them extra clever. I believe the fashions get extra clever they do all the pieces higher, they do all the pieces extra intuitively, and I believe that might be wonderful.”
Takeaway
Google DeepMind, and Google as a complete, is transitioning towards greater than offering solutions, it’s serving to customers accomplish duties. All of Google’s merchandise, from Gmail, to Google Sheets, to Maps, all assist individuals accomplish duties. Kavukcuoglu defined that AI fashions are conceptually turning into extra than simply chatbots, they’re additionally turning into extra agentic, and that displays how all the pieces at Google is altering, together with Search.
Watch the interview right here:
Featured Picture/Screenshot of interview
