Can AI Really Cure All Disease in 10–20 Years?
Yes — and DeepMind CEO Demis Hassabis believes there are no laws of physics standing in the way. When asked directly whether AI can cure all disease within the next decade, Hassabis was careful but optimistic: "In the next 10 to 20 years, I don't see any laws of physics that prevent that." He pushed back slightly on a specific nine-year timeline, but the direction of his answer was clear. This is not a distant dream — it is an active engineering project happening right now at DeepMind and its drug discovery spinout, Isomorphic Labs.
The key insight Hassabis offered is that the breakthrough will not be gradual. It will look more like AlphaFold: years of quiet, compounding progress followed by a sudden, dramatic leap. He told the interviewer not to expect frequent updates on a progress website — "You may not have any updates for a few years, and then suddenly there will be some big breakthroughs." Think of it like the Human Genome Project, which appeared to be moving slowly right up until it wasn't.
04:12
Hassabis explains why the AI cure-all-disease breakthrough will look sudden, not gradual — like AlphaFold did
Watch at 04:12 →
How Does AlphaFold Actually Help Find New Drugs?
AlphaFold solved one of biology's hardest problems: predicting the 3D structure of a protein from its amino acid sequence. Over 3 million researchers are now using it. But Hassabis was direct about its limitations in the context of drug discovery — protein structure is only one step in a process that has dozens of critical checkpoints.
What DeepMind is now building is best understood as a platform of AlphaFold-level models, each targeting a different stage of the drug discovery pipeline. Here is what that platform is designed to predict:
- Protein-protein interactions — not just static snapshots, but how proteins move and bind together dynamically
- Protein-molecule interactions — how a potential drug compound fits into a specific pocket on a protein
- ADMET properties — absorption, distribution, metabolism, excretion, and toxicity, which determine side effects
- Compound design — what exact molecule should be synthesized and how it can be made
Hassabis described the goal as building "another half dozen to a dozen AlphaFold-level models" covering these different parts of the process, then integrating them into a unified platform. Once that platform is validated in pre-clinical trials — which he estimates will take a few more years — it could theoretically be pointed at almost any disease. The analogy he used is striking: just as AlphaFold 2 eventually folded all 200 million known proteins in a single year, this platform could compress decades of drug development into a fraction of the time.
What About Clinical Trials — Can AI Speed Those Up Too?
Hassabis thinks so. Even after a promising drug candidate is identified, it still has to pass through clinical trials — a process that can take a decade and cost billions. He suggested AI could help here as well, by better stratifying patients, predicting optimal dosages, and identifying which trial participants are most likely to respond. He pointed to mRNA vaccines as a precedent: a new technology met an emergency need, the regulatory process was accelerated, and it worked. A similar dynamic could emerge as the first wave of fully AI-designed drugs moves through trials and generates the evidence base regulators need to modernize their processes.
08:45
Hassabis describes the full stack of AlphaFold-level models DeepMind is building for drug discovery
Watch at 08:45 →
What Is DeepMind's Co-Scientist and What Has It Invented?
Co-Scientist is best described as a fine-tuned version of Gemini built specifically for scientific research. It comes equipped with extra tools for hypothesis generation, literature summarization, and data analysis. Hassabis called it "the beginnings of having a great research assistant helping you in your daily work."
It is not writing papers autonomously — not yet. Today it functions as a highly capable collaborator. But its track record already includes some remarkable results. Earlier versions of similar systems inside DeepMind were used to discover more efficient matrix multiplication algorithms and to improve computer science algorithms more broadly — essentially turning the tools of invention on themselves to become more efficient. Hassabis hinted that announcements about co-scientist's scientific contributions are coming soon.
One of the most interesting moments in the interview came when the interviewer — a computer graphics researcher specializing in ray tracing — described testing co-scientist on global illumination problems, a niche field with very little training data. The system came back with sensible, interesting ideas. That kind of generalization to low-data domains is exactly what makes co-scientist potentially transformative for researchers working outside mainstream AI-saturated fields.
How Does Demis Hassabis Actually Use Gemini Day-to-Day?
Hassabis uses Gemini primarily in three ways. First, as a brainstorming partner — for project ideas, project names, and creative exploration. He described it as a sparring partner he bounces ideas off of rather than a tool he uses for final outputs. Second, for rapid literature reviews — getting up to speed on a new area of research he is not deeply expert in, quickly extracting the key points without reading dozens of papers. Third, for thinking through problem-solving steps — working through the logic of an idea collaboratively.
He also acknowledged the potential for using AI more critically — firing up deep reasoning mode and asking it to find the flaws in an argument — but admitted he tends to use it in a more collaborative frame. He said he might try being harsher with it going forward. He stopped short of calling Gemini a confidant in the way Nvidia's Jensen Huang has described using LLMs, but left the door open: "Maybe at some point I will."
15:30
The co-scientist demo moment — ray tracing researcher shares surprising results from a niche low-data domain
Watch at 15:30 →
Perhaps the most moving moment in the interview came from an audience member who shared that Gemini had analyzed a large medical scan file for his mother while they waited weeks for a doctor's evaluation. Gemini told them not to worry — and the doctor later confirmed it was right. Hassabis responded that he has heard many such stories, including cases where AI may have been life-saving.
What Is the Einstein Test for AI — and Why Does It Matter?
This is one of the most intellectually interesting ideas in the entire conversation. Hassabis described a benchmark he calls the Einstein Test: take an AI model, give it a knowledge cutoff of 1901, and ask whether it could independently derive Einstein's Annus Mirabilis papers from 1905 — the four landmark papers that included special relativity, Brownian motion, the photoelectric effect, and the foundation of statistical mechanics.
If an AI can pass that back-test — producing those breakthroughs from the knowledge state of 1901 — then you have meaningful evidence that the same system, trained on all of modern physics, could potentially produce something genuinely new. Something beyond string theory. The Einstein Test is not just a fun thought experiment. It is a rigorous validation framework for scientific AI. Pass it, and you have earned the right to be taken seriously as a discovery engine.
Is DeepMind Building an Automated Lab for New Materials?
Yes, and it is already underway. DeepMind is constructing an automated physical laboratory in London specifically for materials science. The reason is concrete: DeepMind's AI systems have already generated 200,000 designs for novel materials — including potential superconductors — but there is no way to test them fast enough using conventional lab methods.
This connects to a deeper point Hassabis made about recursive self-improvement in science. In coding and math, the verification loop is fast and cheap — you can check whether code runs or whether a proof is valid almost instantly. But in physical sciences, the verifier lives in the world of atoms. It requires robots, reagents, and time. That is why Hassabis is watching robotics development at DeepMind closely, and why he estimated that a truly closed-loop automated discovery system for chemistry or biology is probably 18 to 24 months away from being practical.
21:10
Hassabis outlines the Einstein Test as a rigorous back-test for scientific AI capability
Watch at 21:10 →
Why Is DeepMind Partnering With EVE Online?
DeepMind has announced a partnership with EVE Online, the massively multiplayer space game known for its player-driven economy, complex faction politics, and emergent storylines. For Hassabis — himself a former game developer — the appeal is obvious. EVE is one of the most complex simulated environments ever created, with a functioning economy, dynamic alliances, and a community that effectively co-authors the game's universe.
DeepMind sees it as a sandbox for testing AI ideas in a safe, rich environment — continuing the tradition that produced AlphaGo and AlphaStar. The specific applications being explored include AI agents playing alongside human players, AI systems assisting players with strategy, and potentially an AI game master that dynamically shapes the narrative. Hassabis knows EVE's CEO personally from his earlier days in the games industry, and described the collaboration as a natural fit for both organizations' cultures of pushing boundaries.
