At the end of Google I/O’s keynote, Demis Hassabis didn’t just tease another lab breakthrough or a near-term product. He made a statement that sounded almost like science fiction—delivered with the kind of deadpan confidence that only a seasoned executive can pull off. Google DeepMind, he said, hopes to “reimagine the drug discovery process” with the goal of “one day solving all disease.”
It’s the sort of line that invites skepticism on first read. “All disease” is an enormous claim, and drug discovery is famously slow, expensive, and full of biological surprises. But the more interesting question isn’t whether the ambition is literally achievable in the timeframe implied by the phrase. The more interesting question is what kind of reimagining Hassabis is pointing to—and why this particular moment at I/O is where it’s being said out loud.
Because this wasn’t framed as a single model doing one task. It was framed as a redesign of the pipeline: how we go from biology to hypotheses, from hypotheses to molecules, from molecules to candidates, and from candidates to therapies that actually work in humans. In other words, the target isn’t just better prediction. It’s better decision-making across the entire system.
To understand why “solve all disease” is showing up in a mainstream tech keynote, you have to look at what’s changed in the last few years. AI in healthcare has moved from “assist clinicians” to “assist discovery.” That shift matters because discovery is where the bottlenecks are. If you can compress the time between identifying a promising mechanism and producing a viable candidate, you can change the economics of medicine. And if you can do that across many diseases—rather than one narrow area—you can start to imagine a future where the same underlying approach scales.
DeepMind’s long-running work in protein structure prediction is often treated as the headline story, but the deeper implication is about representation: learning models that can internalize biological structure and then use that internal knowledge to propose new experiments. When Hassabis talks about reimagining drug discovery, he’s essentially pointing to the idea that the “knowledge layer” of biology could become more computationally navigable. Instead of treating biology as something you probe experimentally from scratch each time, you treat it as something you can explore in silico—then validate in the lab.
That’s not a small change. It’s a change in how you spend your time and money.
Drug discovery, at its core, is a search problem under uncertainty. You’re trying to find a molecule that will bind to a target in the right way, modulate the relevant pathway, avoid harmful off-target effects, survive the body’s chemistry, and do all of that with enough potency and selectivity to matter clinically. Each step introduces new failure modes. Even when you have a plausible target, the path from “it should work” to “it works reliably” is littered with surprises.
Traditional approaches handle this by iterating slowly: screen compounds, test them, refine the hypothesis, repeat. AI changes the iteration pattern. It can propose candidates faster, prioritize which experiments to run next, and potentially reduce the number of dead ends. But the real leap comes when AI doesn’t just generate candidates—it helps decide which candidates are worth testing, and which experimental results are most informative for updating the model.
That’s where “reimagine the process” becomes more than a slogan. It implies a closed loop: model → proposal → experiment → updated model. The goal is to make the loop tighter and smarter, so the system learns from each round rather than starting over.
This is also why the “all disease” phrasing lands differently than it might otherwise. If you’re talking about a single disease, you can focus on a specific target class, a specific dataset, and a specific clinical endpoint. But if you’re talking about many diseases, you need a framework that can generalize. You need methods that can transfer across targets, across modalities (sequence, structure, chemistry, phenotypes), and across the messy reality of incomplete data.
In that sense, “all disease” is less a promise of immediate cures and more a statement about ambition: build a platform-level capability that can be applied broadly. It’s a bet that the underlying scientific and computational machinery can scale, even if the outcomes will vary by disease.
There’s another reason Hassabis’ comment feels timely: the industry is converging on a shared view of what “AI for science” means. It’s not only about generating text or images. It’s about building systems that can reason over complex scientific representations—molecules, proteins, pathways, and interactions—and then connect those representations to measurable outcomes.
At Google I/O, the emphasis on AI’s role in science and health reflects a broader shift in how tech companies position themselves. The pitch is no longer “we’ll help you use existing knowledge.” It’s “we’ll help you create new knowledge.” That’s a much bigger claim, and it requires a different kind of infrastructure: data pipelines, model training strategies, evaluation methods, and partnerships with researchers who can run the experiments.
DeepMind’s history suggests it’s comfortable operating in that space. But the key point is that drug discovery isn’t just a technical challenge. It’s a coordination challenge. You need access to biological targets, compound libraries, assay results, and clinical knowledge. You need to align incentives across academia, biotech, regulators, and industry. You need to ensure that models are evaluated in ways that reflect real-world constraints, not just benchmark scores.
So when Hassabis says “reimagine,” it’s also a signal that DeepMind sees the bottleneck as systemic. The bottleneck isn’t only the model’s ability to predict binding or activity. The bottleneck is the entire workflow: how decisions are made, how experiments are prioritized, and how knowledge is accumulated.
A unique angle on this ambition is to consider what “solving disease” would actually require. Disease isn’t a single entity. It’s a set of conditions defined by symptoms, biomarkers, genetics, environment, and time. Even within a single diagnosis, there can be multiple underlying mechanisms. That means a universal cure isn’t just a matter of finding one magic molecule. It’s about mapping heterogeneity and tailoring interventions.
AI could help with that mapping. Models can integrate diverse data types—genomics, proteomics, imaging, electronic health records, and more—to identify patterns that humans might miss. But drug discovery is still constrained by biology’s complexity. A model can propose a target, but it can’t guarantee that modulating that target will produce the desired clinical effect without side effects. Biology is not a deterministic machine; it’s a network with feedback loops and compensatory pathways.
So the “all disease” goal likely implies a strategy that embraces modularity. Instead of one monolithic cure, it’s a portfolio of interventions: drugs, biologics, combinations, and perhaps personalized therapies. The “solving” could mean reducing disease burden dramatically across many conditions, not eliminating every illness forever.
That interpretation makes the ambition more plausible while still acknowledging the scale.
Another insight is that reimagining drug discovery may involve shifting what counts as success. Historically, drug discovery has been measured by whether a candidate reaches clinical trials and eventually wins regulatory approval. That’s a high bar, but it’s also a late measurement. AI systems could aim to improve earlier stages: better target selection, better candidate prioritization, better understanding of failure modes, and better prediction of which experiments will yield the most useful information.
If you can improve those earlier stages, you can increase the probability that the expensive later stages are worth it. That’s how you change the overall throughput of discovery. It’s also how you reduce the number of times you discover too late that a candidate doesn’t behave as expected.
This is where the “golden age of scientific discovery” framing—visible in the imagery around the keynote—becomes more than marketing. A golden age isn’t just about having powerful models. It’s about having a scientific ecosystem that can absorb those models into practice. That means reproducibility, robust evaluation, and careful integration with lab workflows.
It also means confronting the limitations of AI. Models can hallucinate in the sense that they can produce plausible-looking outputs that don’t correspond to real chemistry or real biology. They can overfit to datasets that don’t represent the full diversity of targets. They can fail silently when the input distribution shifts. And they can optimize for surrogate metrics that don’t translate to clinical outcomes.
Reimagining the process therefore requires more than generation. It requires verification. It requires uncertainty estimation. It requires designing experiments that test the model’s most consequential claims. It requires building trust through transparency and rigorous validation.
The most compelling version of Hassabis’ ambition is the one that treats AI as a scientific instrument rather than a magic oracle. An instrument doesn’t replace the scientist; it extends the scientist’s reach. It helps you see patterns, test hypotheses faster, and explore more of the search space than you could manually.
In drug discovery, that could mean exploring chemical space more efficiently, identifying binding modes more accurately, predicting off-target risks earlier, and simulating how molecules might behave in biological contexts. It could also mean using AI to design experiments that are maximally informative—so each lab run teaches the system something new.
If that sounds like a lot, it is. But it’s also exactly the kind of multi-year, multi-disciplinary effort that large research organizations are built for. The “one day” in Hassabis’ statement is doing a lot of work. It signals patience and long-term investment rather than a near-term product roadmap.
Still, the statement has implications for the industry’s expectations. When a major AI leader publicly frames the goal as solving all disease, it raises the stakes for everyone else. It encourages more funding, more partnerships, and more competition. It also increases scrutiny: if progress stalls, critics will point to the grand language.
That tension is unavoidable. But it’s also part of how scientific fields accelerate. Big goals attract talent and resources