BIO

GDM’s Pushmeet Kohli on solving science's biggest challenges with AI

Featured

I recently sat down with Pushmeet Kohli, Head of Science and Strategic Initiatives at Google DeepMind, for a wide-ranging conversation about how AI is being shaped to tackle the toughest problems in science, industry, and society. The discussion—hosted by Google for Developers and led by me—covered recent "Alpha" launches like AlphaEvolve, AlphaFold, AlphaGenome, and AlphaEarth; the team's criteria for picking high-impact problems; how specialized scientific breakthroughs translate into more general AI capabilities; the rise of AI co‑scientists; and what an "API for science" might look like.

This write-up is my take on that conversation: a news-style, first-person account that mixes direct explanations Pushmeet shared with deeper context about why these projects matter, how the work gets transferred into broader AI systems like Gemini, and what the future might hold when scientific discovery becomes more broadly accessible.

Table of Contents

🔬 Recent Alpha launches and why they matter

Over the past few months, DeepMind's science team released a string of alpha projects, each tackling a different domain but tied by the same ambition: use AI to solve problems that were previously too hard or too expensive to solve at scale.

  • AlphaEvolve — a code and optimization agent built on Gemini technologies. AlphaEvolve was applied to real operational problems like optimizing job scheduling and compute distribution across Google’s data centers. The reported impact was substantial: about a 0.7% saving on the overall compute fleet and improvements that accelerated Gemini training itself. That fraction might sound small in isolation, but when applied across the scale of Google's infrastructure it represents large monetary and energy savings.
  • AlphaGenome — a model designed to analyze and make sense of the human genome. The video painted AlphaGenome as a step toward deciphering the biological code at richer resolution, enabling researchers to ask and answer questions about mutations, genomic function, and disease-linked variation more quickly and at higher scale.
  • AlphaEarth — described as an evolution of the Google Earth concept. AlphaEarth ingests satellite and remote sensing imagery and builds a semantic geospatial representation of the planet. That representation lets researchers and conservationists propagate known information (e.g., habitats for a species) across the globe and reason about where else similar conditions exist. It’s far more than mapping — it’s about understanding and predicting planetary-scale patterns.

Across these examples you can see the breadth of the team’s ambitions: optimizing infrastructure, decoding life’s blueprint, and building planetary-scale models. The common thread is leveraging AI not to chase incremental gains but to push the boundary of what’s possible.

🧭 How we choose which scientific problems to attack

One of the clearest takeaways from Pushmeet's remarks was the explicit framework his team uses for selecting projects. This is not a scattershot list of interesting problems; it is a disciplined filter designed to prioritize transformative, feasible, and time-sensitive challenges.

In short, the filter is threefold:

  1. Transformative impact: The problem must have the potential to meaningfully change a scientific field, a commercial sector, or the public good. Incremental improvements aren’t enough — the objective must be game-changing.
  2. Feasibility: The target should be achievable with a credible path forward. They avoid projects that are essentially speculative science fiction; the problem should be something the community expects will be solved eventually, just perhaps not on the near-term horizon.
  3. Acceleration hypothesis: There has to be a belief the team can deliver the result substantially faster than the community consensus — for example, what the community expects to take five to ten years, the team believes they can do in two to five years. If the community already believes success is imminent, the problem isn't the right fit for DeepMind's science unit.

That last point is vital. The team focuses on problems that are hard, amenable to AI-driven approaches, and where DeepMind's combination of research expertise, engineering, compute, and data can pull ahead of the broader field. It's how AlphaFold followed that logic: the protein folding problem was widely understood to be fundamental and hard, but DeepMind pushed it forward far faster than many expected.

🏆 Three types of impact: scientific, commercial, social

Pushmeet framed the outcomes of the science program into three categories: scientific impact, commercial impact, and social impact. These categories help explain why projects look so different from one another but still belong in the same portfolio.

Scientific impact: AlphaFold as the benchmark

AlphaFold is the canonical example of scientific impact. The protein folding problem — predicting a protein's 3D structure from its amino acid sequence — had been an open challenge for decades. Before AlphaFold, determining one protein structure could cost millions of dollars and take a long time in the lab; AlphaFold reduced that to seconds and pennies computationally.

The ripple effects have been enormous. AlphaFold's predictions accelerated research across biology, drug discovery, and basic science. It was also recognized at the highest levels, with significant awards honoring the work that made it possible.

Commercial impact: AlphaEvolve optimizing infrastructure

AlphaEvolve demonstrates commercial value: an AI that optimizes complex compute and scheduling problems, and does so better than hand‑tuned algorithms developed over years. That 0.7% fleet-wide reduction in compute use doesn’t just save money — it reduces energy consumption and enables faster iteration for other AI models. Pushmeet stressed that some of the gains AlphaEvolve produced were beyond what human designers had achieved, and it also contributed to solving open mathematical problems (75% of a test set matched state-of-the-art human solutions, and 20% went beyond them).

Social impact: SynthID and the need for provenance

Generative AI is powerful, but with power comes responsibility. SynthID is DeepMind's watermarking technology: an imperceptible signal embedded in generated content so that users (and systems) can detect whether an image, video, or other media was synthetically produced. Pushmeet framed SynthID as a social-impact tool — preserving trust in information ecosystems by letting people know what is synthetic. It also illustrates the team’s broader view: new capabilities don’t just require technical breakthroughs; they require guardrails, transparency, and tools to manage societal implications.

🔍 The interesting tension: low-hanging fruit or deep exploration?

A question I had — and I suspect many readers have — is whether the science team is plucking "low-hanging fruit" or rigorously digging into genuinely hard problems. Pushmeet's answer was explicit: the unit is not targeting easy wins. Instead they look for problems that are broadly agreed to be challenging and where DeepMind can accelerate progress in a way that the community isn’t expecting in the immediate term.

There are many research directions where the team could make useful gains, but the priority is always impact and hardness. Where an incremental improvement won’t be transformational, they move on. Where a problem requires combining the best research, the best engineering, and massive compute to make an outsized difference — that’s where they focus.

🧠 Wielding AGI: if general intelligence arrives, what will we use it for?

Pushmeet and I spent time thinking about what happens as models become more capable and general. His framing was pragmatic and task-driven: progress in AI will produce more powerful, more general models — that seems inevitable — so the real question is how to wield that intelligence for human benefit.

He argued that our job (as a research/engineering community) is to leverage the progress in model capability to solve the next set of "impossible" problems. The implication is twofold:

  • We should continue to invest in general-purpose models (like Gemini) because they unlock many possibilities.
  • We also need to design targeted systems and pipelines that harness those models for domain-specific breakthroughs — whether that’s in materials science, biology, earth observation, or something else.

In other words: generality is necessary but not sufficient. Building interfaces, evaluation protocols, and domain-specific data sets remains crucial to converting general intelligence into real-world impact.

🤝 Tech transfer and collaboration with Gemini

DeepMind doesn’t operate in isolation. The science team and Gemini (the broader Google AI ecosystem) collaborate closely across architectures, data, and evaluation strategies. Pushmeet described the collaboration as deeply integrated:

  • Base architecture: Science teams work with Gemini teams to experiment with architectures and mechanisms that would make a general model better at scientific tasks.
  • Evaluation: Shared benchmarks help both sides understand where progress is occurring and where it still falls short — whether that's in math, reasoning, or domain-specific knowledge.
  • Data curation: The science unit helps identify and assemble the right data to train models that are stronger in biology, chemistry, materials science, code, cybersecurity, and more.
  • Project collaboration: Some initiatives are developed together from the ground up; other specialized systems eventually produce data or approaches that are integrated back into general models.

One concrete example is the IMO (International Mathematical Olympiad) project. Earlier systems like AlphaProof and AlphaGeometry were more domain-specific and specialized. Over time, the lessons and training data from those systems were transferred into Gemini-based "Deep Think" models, which combined general language understanding with improved reasoning — and the results were impressive.

🧩 From AlphaProof & AlphaGeometry to Deep Think (IMO gold)

I asked Pushmeet to walk me through the IMO story because it’s such a clear illustration of how domain expertise, formal methods, and general models can interact.

Historically, the team worked on two models:

  • AlphaGeometry — focused on geometry problem-solving.
  • AlphaProof — built around the idea of formal proofs and reasoning. AlphaProof converted math problems into a formal language (Lean) and searched the space of mathematical proofs with an LLM guiding search. The key property: when it produced a proof, that proof could be formally verified, meaning correctness was guaranteed.

The verification property is huge. Unlike typical LLM outputs, a formal proof in Lean is a rigorously checkable artifact. That shifts some of the trust concerns to objective verification: if the formal system proves something, the result is mathematically valid.

AlphaProof and AlphaGeometry together had impressive results in a controlled environment. But they were specialized. The big leap after that was to take what these systems taught us about solving hard mathematical problems and tee them into a more widely available model: Deep Think built on top of Gemini architectures.

Deep Think's key differences were:

  • It accepts natural-language problem statements instead of requiring encoding into a proof assistant language; that means more people can use it with ordinary English prompts.
  • It leverages general-purpose reasoning capacities from Gemini while incorporating training signals derived from formal proofs generated by systems like AlphaProof.
  • It is more accessible — not a tucked-away research-only agent but something that can be made broadly available.

The result: a step from silver to gold on the IMO benchmark. Deep Think's performance suggested that domain-specific reasoning can be distilled back into general models to improve their capabilities on rigorous reasoning tasks.

📜 Formal proofs and verifiability: why proof systems matter

The use of formal systems like Lean is more than an academic curiosity. Pushmeet emphasized why formalizable proofs are powerful for training and evaluation:

  • Verifiable correctness: A formal proof, once checked in a proof assistant, offers guarantees that an LLM may not otherwise provide.
  • High-quality training data: If a system like AlphaProof can produce a verified solution to a problem, that pair (problem, verified solution) becomes a gold-standard training example for more general models.
  • Scalability: With a verified solver you can generate hundreds of thousands or even millions of correct problem-solution pairs, creating a powerful augment for a general model's training data.

One natural question I raised is whether the presence of formal proof data primarily helps math tasks or transfers more broadly. Pushmeet admitted that's an open empirical question. It might improve reasoning and instruction-following in other domains, but the team runs ablation studies and careful evaluations to understand cross-domain transfer effects.

🔁 Synthetic data generation: turning verification into training

One pattern that emerges repeatedly is this loop: build a high-quality, verifiable solver for a narrow domain → use it to produce verified outputs at scale → train a general model on those verified outputs so it gains the capabilities without having to be handcrafted for that narrow domain.

AlphaProof is an ideal source for this kind of synthetic data generation. When it finds proofs, those proofs are correct by construction, so they can seed training sets for Gemini variants. The general model then learns the reasoning patterns required to solve similar problems but in natural language, making the capability accessible to users outside the original proof-assistant setup.

This mechanism—automatic, verified data generation—has two key advantages:

  1. It sidesteps the need for manual labeling at scale.
  2. It converts a research-only breakthrough (formal provers) into a broadly useful dataset for general models.

🧮 Does math skill generalize to other domains?

This is a question that came up naturally: if a model becomes very good at math and formal reasoning, does that improvement help in non-math tasks like code generation, customer support, or biological sequence reasoning?

Short answer: nobody knows for sure yet. Longer answer: it's an empirical research question and the team approaches it that way — by running ablation studies, inserting or removing targeted training signals, and measuring impacts across benchmarks.

There are reasons to be hopeful that math-oriented training can benefit other areas:

  • Instruction following: Math tasks often require strict multi-step reasoning and careful instruction following—skills that transfer to code generation, procedural tasks, and logical analysis.
  • Chain-of-thought structure: Good math solvers develop structured chains of thought, and that structure can support clear reasoning in other domains.
  • Rigorous specification: Math enforces precise specification and verification, which is critical in domains like security, system design, and certain aspects of biology.

But the effect size and scope of transfer remain an empirical question. It’s one of the reasons DeepMind integrates specialized systems and general models rather than choosing one approach exclusively.

🌍 Democratizing science: making tools available to everyone

One of the themes that excited me most in the conversation was democratization. Breakthroughs matter only if they’re used, and the team is intentional about serving the broader scientific community.

AlphaFold is the poster child for operationalizing a technical breakthrough. DeepMind didn’t just publish papers; the team predicted structures for nearly every known protein and made them available through the AlphaFold Protein Structure Database. That database allowed researchers everywhere—regardless of resources—to access high-quality structural predictions and accelerate their own science.

Other projects followed a similar mindset:

  • AlphaGenome: built user interfaces for researchers to probe the effects of genomic variation.
  • AlphaEarth: was framed as a semantic layer over satellite data to empower conservationists and climate scientists.
  • AlphaEvolve: yielded operational gains that could be integrated into infrastructure pipelines to reduce costs and energy usage.

Pushmeet stressed that from day one, their plan is not just to demonstrate capability but to provide accessible tools—APIs, databases, and UIs—so other scientists can make practical use of the models.

🧪 AI Co-Scientist: simulating the scientific process with multi-agent systems

One of the most exciting projects we discussed is AI Co-Scientist, a multi-agent setup in which a single system plays multiple roles in the scientific workflow: hypothesis generator, critic, editor, reviewer, and ranker.

Think of it as a synthetic lab partner that can generate novel ideas, critique them, iterate on them, and prioritize the most promising ones. Pushmeet described a powerful anecdote: a researcher at Imperial College provided the system with a problem and expected routine suggestions; the system returned ideas that matched work the researcher’s team had been developing privately and even suggested other novel directions. The researcher initially suspected the system had access to their unpublished work, but it hadn’t — the convergence was an example of independent ideation.

That capability changes the dynamic of scientific research. It doesn’t replace expertise, but it augments it. Laboratories and researchers can explore a broader hypothesis space, accelerate iteration, and test directions that might never have been considered otherwise.

Pushmeet emphasized that AI Co-Scientist isn't magic; it's an engineered process that leverages multi-agent dynamics and the generative power of models like Gemini to emulate aspects of scientific collaboration. The early results—especially in areas like antimicrobial discovery—are promising.

🔐 Responsibility and verification: managing the risks of powerful tools

With great capability comes great responsibility. The team invests heavily in safeguards: formal verification where possible (e.g., proofs), watermarking for provenance (SynthID), and careful release strategies for public accessibility. The goal is to create systems that are useful, trustworthy, and auditable.

SynthID is one concrete instance of that philosophy: it’s not enough to generate content responsibly; we must also build signals that help consumers and systems distinguish synthetic content from natural content. That’s critical for maintaining trust in information ecosystems at the scale we’re heading toward.

🔧 The specification problem: designing interfaces for scientific work

If an API for science is going to transform who can do research, there is a technical and human-centered challenge that precedes it: the specification problem. How do you let people specify the problem they want solved? How do you capture intent so a model can produce useful, verifiable results?

Pushmeet compared this to the historical arc of software development: coding used to require high levels of expertise, but with higher-level languages, frameworks, and tooling, more people can build software. For science to become more democratically accessible, we need similar leaps in how people specify experiments, hypotheses, and evaluation criteria.

The interface layer matters: rich UIs, domain-specific languages, guided prompts, and validated defaults will all play roles in enabling people with domain knowledge (but not necessarily years of training in formal methods or AI) to harness powerful systems responsibly.

🧭 An "API for science": what it could look like

We wrapped our conversation talking about the possibility of an "API for science" — a programmable interface that lets researchers query AI services to help generate hypotheses, design experiments, analyze data, and reason about tradeoffs. Pushmeet was optimistic: the technical foundations exist, but we must solve the specification and interface questions to make the API broadly useful.

Key components of such an API might include:

  • Problem templating: predefined schemas for common scientific tasks (e.g., protein design, sequence analysis, environmental inference) so users can provide structured inputs quickly.
  • Verified output channels: integration with verifiers or lab automation so that outputs can be checked and replicated.
  • Multi-agent orchestration: prebuilt agent "roles" (hypothesizer, critic, experiment designer) to simulate the collaborative scientific process.
  • Provenance & watermarking: built-in signals for synthetic content and model-derived hypotheses to preserve trust.
  • Human-in-the-loop controls: ways to inject expert knowledge, constraints, and manual review at critical points.

That suite of capabilities would make it possible for people outside elite labs to engage in high-impact research, accelerating innovation across the board.

📦 Deployment and access: making science tools widely available

Deployment strategy has been central to the team's approach. Pushmeet made clear that releasing models and services widely is part of the plan, not an afterthought. We've seen a few playbooks emerge:

  • Databases and public access: AlphaFold’s protein structure database is a textbook example: predict a global set of entities and publish them for universal access.
  • Web UIs & APIs: for models requiring interaction or problem-specific queries (e.g., AlphaGenome), the team builds dedicated interfaces so researchers can interrogate models without deep engineering pipelines.
  • Integration into product ecosystems: for capabilities like Deep Think or math reasoning, the models can be surfaced through broader products (e.g., Gemini integrations) so developers and end-users can access them in familiar environments.

The intent is to avoid the trap of creating breakthroughs that only a handful of institutions can use. Democratization is baked into the mission.

🔬 What success looks like: Nobel-level breakthroughs and everyday empowerment

Pushmeet framed the highest-level aspiration in striking terms: enable Nobel-level breakthroughs more frequently and from a broader pool of people. AlphaFold was one instance where AI materially changed the trajectory of an entire field. The next step is building AI systems that make high-impact discovery possible for far more researchers, not just those at elite institutions.

That democratization could reshape the practice of science. Imagine a biologist in a resource-constrained lab designing protein variants with help from a co‑scientist, or a climate researcher leveraging AlphaEarth’s semantic layer to discover previously unnoticed habitat shifts. The combination of higher-level tools and powerful AI models makes that future plausible.

🔮 Looking ahead: the next five years

So what should we expect in the near term? Based on Pushmeet’s remarks and the direction of the projects discussed, a few trends seem likely:

  • More mixed systems: Expect systems that combine verified, domain-specific solvers with general models. The hybrid approach (formal proof + LLM) will be a blueprint for other domains.
  • Data-driven capability transfer: High-quality synthetic data generated by specialized systems will be a staple for amplifying general models in critical domains.
  • Broader access: More tools will be deployed via APIs, UIs, and shared databases so a wider set of users can leverage advanced AI for research.
  • Societal safeguards: Provenance, watermarking, and verification tools will be integrated into release strategies as standard practice, not optional add-ons.
  • Human + AI collaboration: Co-scientist workflows will scale up, enabling humans and AI to iteratively explore large hypothesis spaces together.

💬 A few surprising anecdotes from the conversation

I want to close this section with a couple of real-world stories that stuck with me:

  • When researchers at Imperial College gave a problem to the co‑scientist system they were astonished that the first idea the AI suggested matched work they were themselves preparing to submit. They initially believed the system had seen their unpublished paper. That wasn’t the case — instead, it shows the AI can converge on independently valuable ideas.
  • AlphaEvolve didn’t just shave a tiny percentage off an experimental metric; that 0.7% saving in fleet compute translated into very large real-world savings and also made Gemini training materially faster — a direct cross-product benefit.
  • AlphaFold's deployment model — predict globally and publish a database — is a masterclass in how fundamental research can be turned into an equitable public resource.

🧾 Final thoughts — why I’m optimistic but cautious

My conversation with Pushmeet reinforced a view I already had: AI is rapidly maturing into a tool that can amplify human intelligence in ways we can already see. The DeepMind science team's explicit focus on transformative problems, rigorous verification, and public accessibility is a healthy roadmap for realizing that potential responsibly.

That said, the size and speed of these advances underscore the necessity of safeguards. Verification, transparency, provenance, and thoughtful deployment strategies will determine whether these tools uplift science or introduce new failure modes. I appreciated how Pushmeet kept returning to those themes — the technical ambition is matched by a clear-eyed view of responsibility.

Going forward, I’m especially excited about two complementary directions. First, continued development of hybrid approaches: marrying formal, verifiable methods with the generality of LLMs seems like the fastest path to trustworthy breakthroughs. Second, broadening access through APIs, databases, and carefully designed interfaces so more people can participate in scientific discovery.

That’s the vision Pushmeet and I discussed: not simply making AI smarter, but making it useful and safe for everyone who wants to work on humanity’s biggest problems.

❓ FAQ

What is the DeepMind science team's selection criteria for projects?

I described it earlier but to repeat succinctly: they pick projects that are (1) transformative in impact (scientific, commercial, or social), (2) feasible with a credible technical pathway, and (3) where they can accelerate progress significantly relative to the community consensus (e.g., cutting a 5–10 year horizon substantially shorter). They avoid small, incremental improvements and avoid problems that are effectively impossible today.

How do specialized systems like AlphaProof help general models like Gemini?

Specialized systems can generate high-quality, verifiable outputs (like formal proofs). Those outputs become trusted training data for general models. By training Gemini variants on that verified data, general models inherit abilities without requiring everyone to run specialized provers. In essence, the specialized system scales its expertise by teaching the general model.

Does being better at math make a model better at other tasks?

It’s an active research question. There are theoretical reasons to expect some transfer—structured reasoning, instruction following, and chain-of-thought can be helpful across many tasks. But the size and breadth of transfer effects must be proven empirically. DeepMind runs targeted ablations and evaluations to measure the exact impact.

What is AI Co-Scientist and how does it work?

AI Co-Scientist is a multi-agent configuration where an LLM plays multiple roles in the scientific process: hypothesis generator, critic, editor, reviewer, and ranker. The system iterates on ideas, critiques them, and surfaces the most promising directions. Early results show it can generate hypotheses that human researchers find valuable and sometimes novel.

How does DeepMind make models and results available to researchers?

They try to make research accessible through APIs, public databases (the AlphaFold database is the canonical example), web UIs, and product integrations. The philosophy is to operationalize breakthroughs so a researcher in any part of the world can access and apply them.

What is SynthID and why is it important?

SynthID is a watermarking system for generative content. It embeds an imperceptible signal into images, videos, and other media so that generated content can be detected, even after common transformations. This is important for provenance, trust, and combating misinformation as generative models scale.

Will there be an API for science?

Pushmeet and I both think it’s likely and desirable. The technical components are largely within reach: multi-agent orchestration, verified solvers, and high-quality synthetic data generation. But we need better specification languages, UI paradigms, and human-in-the-loop controls to make such an API broadly useful and safe. If built correctly, an API for science could significantly lower the barrier to high-impact research.

What should researchers and developers be doing now?

If you’re working on scientific or industrial problems, start thinking about how to structure problems so they can be consumed by models (clear specs, datasets, evaluation criteria). Engage with the interfaces that are being released, provide feedback, and if possible, participate in collaborations that test AI co‑scientist workflows. The next few years will be about iterating on both capability and workflow design.

Any final advice for people curious about this space?

Stay curious but critical. Explore these tools, experiment, and bring domain expertise to the table. The most exciting work will come from teams that combine deep domain knowledge with an understanding of how to pose problems to AI systems and verify the outputs. And don’t forget: building the interfaces to make AI useful to more people is as important as building the models themselves.

📢 Closing note

I want to thank Pushmeet Kohli for a candid, thoughtful conversation. The science and strategic initiatives team is pursuing a deliberate strategy: pick problems that matter, design solutions that are verifiable and accessible, and use the power of general AI models to amplify impact. The results we've already seen—from AlphaFold to AlphaEvolve to Co-Scientist prototypes—show that approach can produce both Nobel-level breakthroughs and practical improvements at global scale.

If you’re excited by the possibility of an API that helps more people do science, you’re not alone. The next few years will be decisive in shaping how AI becomes a partner for discovery—and whether that partnership is equitable, trustworthy, and genuinely transformative.

Share this post
  • Read More
  • Featured image for the blog, "Jak TPG zrychluje investiční výzkum s ChatGPT"
    Jak TPG zrychluje investiční výzkum s ChatGPT
  • Featured image for the blog, "Sephora mění beauty byznys pomocí AI: jak vzniká personalizovaný nákupní zážitek nové generace"
    Sephora mění beauty byznys pomocí AI: jak vzniká personalizovaný nákupní zážitek nové generace
  • Featured image for the blog, "Jak vypadá AI native budoucnost pro jednu z největších automobilových tržišť na světě"
    Jak vypadá AI native budoucnost pro jednu z největších automobilových tržišť na světě
  • Featured image for the blog, "Builders Unscripted: Derya Unutmaz ukazuje, jak AI mění biologii, imunologii i budoucnost medicíny"
    Builders Unscripted: Derya Unutmaz ukazuje, jak AI mění biologii, imunologii i budoucnost medicíny
    • AI World Vision

    AI and Technology News