May 24, 2025 · 32m · big-technology
Google DeepMind CTO: Advancing AI Frontier, New Reasoning Methods, Video Generation’s Potential
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the Big Technology Podcast at Google I/O, Google DeepMind CTO Koray Kavukcuoglu discusses the research roadmap toward AGI, detailing advances in multimodal foundation models, test-time reasoning with DeepThink, and generative video systems.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Alex holds 27.1% of the talking time here. How this is scored →
Alex as informed peer 4.9 Guest teaching 4.6 Guest disagreement 2.0 Alex pushing back 3.8
speaking balance: gold is Alex, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 9:37 Kavukcuoglu dismisses LeCun's critique as a strawman hypothesis
Kavukcuoglu firmly rejects the premise that labs are merely scaling LLMs to reach AGI, dismissing LeCun's argument as an untested hypothesis not reflective of actual frontier research.
Hardest push from Alex ▶ 9:13 Kantrowitz challenges DeepMind's path with LeCun's AGI critiqueKantrowitz confronts the guest directly by citing former lab director Yann LeCun's emphatic claim that scaling LLMs cannot reach human-level intelligence.
Biggest teaching moment ▶ 13:05 Kavukcuoglu corrects product framing and explains parallel reasoningKavukcuoglu clarifies that DeepThink is a reasoning mode rather than a new standalone product and educates the host on parallel hypothesis generation versus sequential chain of thought.
Alex holds their own ▶ 5:41 Kantrowitz cites industry-wide model stalls to question scaling limitsKantrowitz demonstrates strong domain command by listing unreleased frontier models across OpenAI, Meta, and Anthropic to challenge the guest on scaling diminishing returns.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Alex as informed peer | Guest teaching | Guest disagreement | Alex pushing back | Why |
|---|---|---|---|---|---|---|
| Google DeepMind's Unified AGI Mission and Research Pillars | 6 | 4 | 2 | 5 | Kantrowitz presses on whether scaling laws are hitting diminishing returns, citing industry-wide delays across OpenAI, Meta, and Anthropic. Kavukcuoglu provides a balanced breakdown of how compute scale interacts with algorithmic architecture, data quality, and inference techniques. | |
| Model Progress Across the Industry and the Gemini Roadmap | 4 | 5 | 2 | 3 | Kavukcuoglu reframes the scaling discussion around the tripartite nature of scaling laws (data, compute, parameters) rather than single-dimension scaling. Kantrowitz listens as the guest outlines DeepMind's progress from Gemini 1.5 to 2.5. | |
| Debating AGI Pathways and Yann LeCun's Critique | 6 | 6 | 4 | 6 | Kantrowitz challenges Kavukcuoglu with Yann LeCun's assertion that scaling LLMs cannot achieve human-level intelligence. Kavukcuoglu counters by calling it an unproven hypothesis and noting that no serious frontier lab is solely scaling LLMs without broader architectural innovations. | |
| DeepThink and Parallel Reasoning in Gemini 2.5 Pro | 5 | 6 | 2 | 4 | Kantrowitz asks about the DeepThink announcement and inference-time compute. Kavukcuoglu clarifies that DeepThink is a reasoning mode rather than a standalone product, educating the host on parallel hypothesis exploration versus single chain-of-thought processing. | |
| Native Multimodality and Continuous Frontier Velocity | 5 | 4 | 3 | 5 | Kantrowitz asks if frontier model improvement velocity is slowing down compared to earlier leaps like GPT-3 to GPT-4. Kavukcuoglu disputes any perception of slowdown, arguing that native multimodality and unified reasoning architectures are accelerating overall capability. | |
| Quantifying Model Improvement and User Feedback | 4 | 5 | 2 | 4 | Kantrowitz poses a hypothetical question about the real-world value of 10% or 50% model improvements. Kavukcuoglu reframes the question around non-linear evaluation metrics and feedback loops between research and product deployments. | |
| Generative Video Breakthroughs: Veo 3, Physics, and Flow | 5 | 4 | 1 | 2 | Kantrowitz highlights Google's Veo 3 and Flow video tools, noting synchronized audio and storyboard capabilities. Kavukcuoglu explains the technical milestones from physical object dynamics in Veo 2 to orthogonal audio-visual modeling in Veo 3. | |
| Open Weights vs. Closed Frontier Model Governance | 6 | 5 | 2 | 4 | Kantrowitz asks how Google navigates the tension between open-source contributions and proprietary frontier models like Gemini, referencing DeepSeek and transformer history. Kavukcuoglu articulates Google's dual strategy of releasing Gemma open weights while restricting Gemini weights for safety and governance. | |
| The Potential of 'Vibe Coding' for Accessibility | 3 | 2 | 0 | 1 | Kantrowitz ends on a light note asking about vibe coding. Kavukcuoglu enthusiastically endorses the concept, explaining how it democratizes software creation for non-technical users. |
Statements from this episode (8)
DeepMind CTO: Gemini and Veo are our primary pathways to AGI
“So like with Gemini models, with VO models, all those kinds of generative AI models and exciting things that is happening there. That's our main line of AGI research that's going on.”
DeepMind CTO: AI algorithms and architecture matter just as much as scale
“When we are thinking about our architectures,
Like the architectural elements, the algorithms that we put in there that come, that make up the model, right?
They are as important as the scale.”
DeepMind CTO denies AI slowdown, says progress across the field remains stellar
“Rather than thinking about scaling only in one dimension, there's actually many different ways to think about it, and investing in those, and we can see the returns that I think across the field, really, not just not just here at Google, but across the field, …”
Kavukcuoglu: No major AI lab relies solely on scaling LLMs for AGI
“But also, I don't think that there's any research lab that is trying to only do scaling up the LLM.”
Gemini 2.5 Pro uses DeepThink to generate parallel hypotheses, not single chains
“It is a mode that we are enabling our 2.5 pro model so that, like, it can spend a lot more time during inference time to think, to build hypotheses, and the important thing is to build parallel hypotheses rather than a single chain of thought. It can build par…”
DeepMind CTO: A 10% gain in mathematical reasoning vastly expands AI applicability
“So when we say by improving 10%, if we can improve 10% by its understanding in math, right, understanding of really highly complex reasoning problems, I think that is a huge improvement because then that actually expands the general knowledge that, that would …”
Kavukcuoglu: Veo 2 is DeepMind's first video model to truly understand physics
“With VO two, I think for the first time we could comfortably say that for many, many cases, right, the model has understood the dynamics of the world well.”
DeepMind CTO: 'Vibe coding' allows non-programmers to easily build software applications
“I find it really exciting, right? Like, I mean, because, like, what it does is all of a sudden it enables a lot of people who are not necessarily, who do not necessarily have that coding background to build applications.”