This summary conveys, based on an interview with Pushmeet Kohli, head of the Science & Strategic Initiatives team at Google DeepMind, how AI is solving some of science's most challenging problems. Alongside representative innovations such as AlphaFold, AlphaEvolve, and AlphaEarth, it covers the team's research framework, diverse impacts, teamwork, and the vision for future science tools like the AI Co-scientist that anyone can use. Let's explore, with quotes, how AI is driving real paradigm shifts in the scientific world today!
1. DeepMind's Recent Alpha Innovations and the Evolution of AI
The conversation opens with the host's impression that "the pace of DeepMind's scientific innovation is incredibly fast." Within the past three months, three major Alpha-series models were released, and Pushmeet Kohli directly explains each one's impact and specific achievements.
"We're not looking for simple incremental improvements. We are truly seeking goals that will have a transformative impact on humanity."
A prominent example is AlphaEvolve, which attracted significant attention by solving a wide range of optimization problems, from data center optimization to improving Gemini training speed. AlphaGenome focuses on decoding the human genome, and AlphaEarth enables understanding of various phenomena at planetary scale using remote sensing data.
AlphaEarth is described with the following analogy:
"AlphaEarth, much like how Google Earth integrates various information so that anyone worldwide can easily understand the Earth, combines remote sensing satellite data into a single semantic representation to understand changes at a global scale, propagating attribute information so that, for example, 'if a particular species inhabits this area, we can infer about other areas with similar environments.'"
In this way, DeepMind is not limiting itself to one or two fields, but testing AI's potential across a broad range of problem domains.
2. DeepMind's 'Selection Formula' for Choosing Breakthrough Research
Behind DeepMind's ability to achieve world-class results across various domains lies a systematic research topic selection framework.
"The problems we choose to tackle must meet three conditions:
- They must have a transformative impact on humanity,
- Everyone must think they can't be solved within 5-10 years,
- But we believe we can solve them in half that time or less."
They don't pursue simple progress -- they challenge problems that are currently considered physically or intellectually impossible. For example, before AlphaFold was released, finding a single protein structure "took years and millions of dollars," but AI replaced this process "in seconds, for pennies," flipping the paradigm.
3. Diverse Impact: Science, Commerce, Society
The innovations achieved by the DeepMind science team are categorized into three dimensions of impact.
3.1. Scientific Impact
"AlphaFold became one of the most cited papers ever, and in 2024, Demis and John received the Nobel Prize." "In just about four years, it had an enormous impact on the scientific community."
The breakthrough in protein structure prediction provided critical advances for drug development, human health, and understanding life itself.
3.2. Commercial Impact
AlphaEvolve saved 0.7% of Google's total compute resources, generating hundreds of billions of won in real value.
"AlphaEvolve solved data center optimization problems that no computer scientist had been able to solve. It also significantly increased Gemini training speed."
In mathematical problem-solving, it "found best-in-class solutions for 75% of problems and even better solutions for 20%."
3.3. Social Impact
SynthID embeds 'imperceptible watermarks' in generative AI content (text, images, video), allowing users to distinguish "whether something is real or AI-generated."
"This kind of watermarking was pioneered by Google and is now applied to all content produced by generative AI."
4. AGI (Artificial General Intelligence) and Teamwork / Technology Convergence
DeepMind operates on the premise of "more general, more powerful AI" and focuses on what to use these resources for.
"Our role is to apply all that progress to solving 'the next impossible thing' and turning it into benefits for humanity."
As they focus on new challenges enabled by AI advances, there is active technology and data sharing, tech transfer between DeepMind and the Gemini development team. Each project is characterized by "close collaboration on the latest architectures, evaluation frameworks, and training data composition."
For example, the IMO (International Mathematical Olympiad) project shows the technology transition process from domain-specific models like AlphaProof and AlphaGeometry to the generalized Gemini-based DeepThink model.
"Until last year, we used domain-specific models like AlphaProof and AlphaGeometry, but this year the DeepThink model achieved an IMO gold medal using only 'natural language descriptions in English.' It's now an almost universal model accessible to everyone."
5. From Math/Reasoning AI to General-Purpose 'DeepThink': Technology Generalization and Data Utilization
At the previous IMO, they "challenged for gold by just one point from silver," mainly using customized models like AlphaProof/AlphaGeometry.
AlphaProof converted math problems into Lean, a formal language, to derive formal proofs, thereby guaranteeing the reliability of answers.
"AlphaProof formally specifies all problems, and if it finds a proof, it can guarantee that the answer is truly mathematically verified."
The large volume of verified mathematical data from these models was used for Gemini and DeepThink training, enhancing the generality and reasoning of their solutions.
"Through millions of proofs generated by AlphaProof, Gemini was also able to learn this kind of thinking and problem-solving."
Currently, DeepThink "can solve problems presented in plain English," no longer requiring manual translation or special domain languages, having evolved to a stage accessible to anyone.
6. Accessibility and the 'AI Co-scientist': Democratizing Scientific Innovation
DeepMind believes that for AI to have real impact on all of humanity, "AI tools must be easy for anyone in the world to use."
"With AlphaFold, we predicted the structure of nearly every known protein on Earth, uploaded them to the AlphaFold Database, and anyone worldwide can access them with a single click." "Researchers in Brazil, Africa, and elsewhere are inputting their own proteins and getting results, realizing global AI democratization."
In the same spirit, a new paradigm called AI Co-scientist is being envisioned. This system is designed so that a single model plays multiple roles (hypothesis proposer, critic, editor, etc.) simultaneously to automate/simulate "the entire process of scientific research."
"The AI Co-scientist... has Gemini take on multiple scientist roles, generating ideas on its own, critiquing them, prioritizing them, and conducting research just like a real science team." "Anyone, anywhere -- even in places without PhDs, or where bold scientific ideas were previously out of reach -- can now make such attempts."
Real-world scientists have been surprised that "the Co-scientist produced the same alternatives as ideas we had just submitted in our paper," showing how quickly the gap with actual scientific practice is being closed.
"These experiences keep repeating, and researchers are amazed to see AI proposing ideas that are new, and without even referencing their own papers."
7. Future Outlook: 'An API for Science' and an Era Where Anyone Can Be a Scientist
DeepMind's vision is like "an API for science," enabling even people without specialized expertise to create innovative research and results with AI's help.
"Just as the definition of who creates software has changed in coding, an era is coming in science where entirely different kinds of talent will produce important achievements." "The problem ultimately comes down to 'problem specification.' When we build much more natural interfaces for AI and developers to communicate, AGI and users can truly achieve innovation."
The host half-jokingly, half-seriously suggests, "If a 'Science API' comes out in the second half of 2026, let's connect it to AI Studio right away," and the conversation wraps up.
Closing
DeepMind's scientific innovation strategy is not about simple problem-solving but focuses on transformative change for humanity through AI. As demonstrated by cases like AlphaFold, they solve challenges previously thought impossible with unprecedented methods and broadly share those achievements with the scientific community and the world, realizing the 'truly meaningful impact' of AI. Going forward, tools like the AI Co-scientist will enable more people to become agents of innovation, bringing the era of science and technology democratization ever closer.