Pushmeet Kohli, vice president of science and strategic initiatives at Google DeepMind, said in a speech at the Hindustan Times Leadership Summit 2025 that DeepMind's mission continues to be to responsibly build artificial intelligence (AI) for the benefit of humanity.
He said that after 15 years of operation, DeepMind believes that AI can push the boundaries of human knowledge. At a time when many frontier AI companies are experimenting with similar research, approaching these as “scientific problems” is key to DeepMind's persistence at the intersection of science and AI. “Science is in the DNA of this organization,” Kohli said.
“We do this through science, and we are making progress in many scientific fields. We have been fortunate to demonstrate the potential of AI in problems such as protein structure prediction with AlphaFold. We are focused on these problems where AI can have a transformative impact,” Kohli said. He argued that these improvements cannot be made incrementally, but will “change the way society does things.”
AlphaFold is an AI program developed by DeepMind that uses predictive techniques to understand the structure of proteins. Examples include designing crops that are more resilient in a warming climate and understanding the key proteins behind heat stroke.
“As a case study, before AlphaFold was released by Google DeepMind, it could take nearly five years to solve the structure of a single protein. And these proteins are the building blocks of life. From drug discovery to designing new enzymes to combating pollution, everything is essentially made of proteins. But we didn't know the structure of these proteins. Uncovering this knowledge would have required enormous effort and effort,” Kohli said.
big focus and narrow focus
Conventional wisdom holds that large-scale language models (LLMs) are improving and should be applied to more problems. This differs from DeepMind's relatively narrow approach of tailoring its AI to what is needed within the domain.
Kohli claimed that DeepMind's main focus is on pushing boundaries and developing the most powerful AI models.
“The question we ask is: What are the strongest and most capable models? Essentially, the way we measure intelligence is to know how quickly a model can accomplish a task. We need to develop models that are more capable than ever at solving more difficult problems in a more general way,” he said.
Kohli said less data and less monitoring will be key to success. “In some ways, LLM is a natural evolution of our long-term focus. We are working on building fundamental breakthroughs that increase the efficiency of AI models, all the way to special models like AlphaFold,” he said.
In 2016, Google's AI program AlphaGo competed against Lee Sedol, a legendary Korean player of the ancient board game Go. Considering that Go is a much more complex game than Chess for a computer to master, AlphaGo won the match, which was a major milestone for AI at the time.
Kohli pointed to Google's Gemini family of models, which he said have shown some ability in a variety of tasks beyond just answering questions in everyday chatbot usage scenarios. This week, Google rolled out its latest and most capable model, Gemini 3 Deep Think, to Ultra subscribers. Last month, the Gemini 3 and Gemini 3 Pro models were released. Gemini 3 Pro in the Gemini app, Gemini 3 in AI mode in search for complex inference, and the latest Gemini models as part of search.
“If you look at the spectrum, yes, we are working on specialized models, but we are also improving the general model,” Kohli said. “At the end of the day, it’s all about the problems and how you solve the most impactful problems.”
Scientists who trust AI
Replying to a question on whether scientists have reached a stage where they can trust the output of AI and use it for research, Kohli said AI still makes some mistakes. He added that knowing when AI fails is a key element. “If you're a biologist, you'll tell me that it's very accurate, but it's still wrong sometimes. You don't want to spend years of your life thinking it's right, only to later find out it's wrong.”
He said that if AlphaFold made a mistake, he would “raise his hand and say, “Maybe I wasn't sure about this, so don't take too much credit.''
To tackle the problem of hallucinations in modern LLMs, Kohli said he is building a tool that can identify such instances and alert users. “Can we use AI to solve energy problems? Can we discover new types of materials?”
future vision
Kohli said there will be more emphasis on structural biology as fundamental advances have been made. He added that the key to this is the democratization of technology, allowing more users to benefit from it. Kohli said accelerating science and agent systems that bring AI capabilities to more tasks will be key themes for next year.
“The impact on healthcare and drug discovery is going to really accelerate, especially in countries like India. Kohli believes there is a lot of scope for AI in healthcare in India,” he said.
Kohli said 180,000 researchers and students are using AlphaFold in India. He said he was surprised by the number of people in this country researching protein structures and disease treatments. “It also shows the vast research ecosystem in this country.”
