The Future Map of AI Research: Innovations and Prospects Revealed in the DeepMind CTO Interview
In this article, based on the interview with Korai Kavukolu, CTO of Google DeepMind, on the Big Technology Podcast released in May 2025, we explain the company's vision for AI research, its initiatives in cutting-edge technology, and its future prospects. We have organized the content clearly, incorporating specific examples and quotes, ranging from technical theory and product application examples to the relationship with open source.
1. DeepMind's Vision and Research Areas
1-1. Shared Vision for AGI
At the beginning, Korai stated clearly, "Our goal is to build AGI (Artificial General Intelligence)." While DeepMind holds AGI as its ultimate goal, it is also actively working on the impact that current AI technology has on society.
"We are passionate about both research toward AGI and the social impact of current technologies."
1-2. Expansion of the Research Spectrum
At DeepMind, three major research areas are currently underway in parallel.
Generative AI models (Gemini, VO, etc.)
Applications to scientific fields (AlphaFold, etc., application of AI to mathematics and chemistry research)
Fundamental computer science research
As a result, while exploring the path toward AGI, they are also creating examples where current AI is opening up "new science."
2. Balance Between Scaling and New Technology
2-1. Significance and Limitations of Scaling
Korai states, "Scale is one of the important factors, but it is not the only solution." While it is true that increasing computational resources improves model performance,
"In research, data, algorithms, and architectural ingenuity are just as important as scale," he emphasized, highlighting a stance that does not rely solely on scaling.
2-2. Necessity of New Technology Development
Korai focuses on techniques such as "various inference-time techniques" and "parallel chains of thought." They are researching new attempts such as "DeepThink," which verifies multiple hypotheses simultaneously rather than a single chain to derive the optimal answer.
3. Inference Capabilities and the Potential of "DeepThink"
3-1. Differences from Traditional Chain-of-Thought Models
While conventional models follow a single chain in detail, DeepThink is characterized by "parallel hypothesis construction -> comparison -> refinement."
"By building multiple chains in parallel and cross-verifying each hypothesis, we achieve a deeper understanding."
3-2. Initial Research Results and Future Challenges
Currently, we are in the research phase, having initiated collaborative experiments with safety research teams and trusted testers. We plan to optimize our training methods while identifying actual use cases and application areas.
4. Evolution of the Multimodal Model 'Gemini'
4-1. The Path of Version Upgrades
According to Mr. Korai, the Gemini model has steadily improved in performance from 1.0 to 2.5 Pro. In particular, reasoning capabilities, coding skills, and multimodal support have been strengthened, and he summarizes,
"The fact that a single model can handle everything from image understanding to code generation is a major catalyst."
4-2. Product Application and User Feedback
When actually integrated into Google products, we work with product teams to determine which areas of performance improvement will provide the most user value. We quantify user feedback and feed it back into research and development.
5. Progress in Video Generation Technology and 'V3'
5-1. The Leap from V1 to V2 to V3
In the V2 generation, it became possible to reproduce physical behavior and dynamics between objects with high precision, and in V3, 'simultaneous generation of video and audio' has become a major topic. Demos that naturally reproduce the cross-section of a cut tomato and the sound of it falling have attracted significant attention.
"Generation that captures the mutual complementarity of sight and sound demonstrates a new level of capability."
5-2. Story Flow Generation with 'Flow'
With the Flow feature, users can sequentially prompt scenes, enabling the continuous generation of short videos as a story. A wide range of applications is expected, including content creation, education, and entertainment.
6. Collaboration Between Open Source and Proprietary
6-1. Publication of Research and Responsible Development
Google has a history of publishing foundational research, such as Transformers, through papers and open source. On the other hand, the cutting-edge Gemini model is kept closed, and we maintain a stance of,
"fulfilling our responsibility to bring positive impact to society while managing the risks of technology misuse."
6-2. Providing Models for Each Community
We provide a diverse range of models for developer communities with different needs, such as GMA (Gemini Open Weights) and Llama-like open-weight models. These can be used selectively depending on the application and scale.
7. Future Outlook: The Path to AGI
7-1. A Research Portfolio Pursuing Multiple Hypotheses
Mr. Korai emphasized the importance of "continuing to invest in multiple approaches rather than just one to achieve AGI." By advancing research in parallel across areas such as scale, data, algorithms, and reasoning methods, they are searching for the shortest route.
7-2. Social Impact and Ethics
In the final section, he reiterated that "the goal is not AGI itself, but maximizing the benefits that AGI brings," demonstrating a development stance that balances positive impact with safety.
The above is a summary of the key points from the latest interview with the Google DeepMind CTO. The company's efforts to achieve AGI through diverse research approaches and product applications will likely continue to lead the industry.
Related Articles