Key Moments
DeepMind’s Insane AI Breakthroughs With CEO Demis Hassabis
Two Minute PapersWant to know something specific about what's covered?
We've already dissected every moment. Ask and we will deliver (with timestamps).
Key Moments
DeepMind is building a platform of AI models to discover drugs for all diseases within 10-20 years, but regulatory hurdles and the need for physical verification may slow the process.
Key Insights
Gemini, with its long context capabilities, has been used by individuals to analyze medical scans, providing results later verified by doctors.
Demis Hassabis uses Gemini for brainstorming, creative ideas, and summarizing new research areas, acting as a sparring partner for his work.
AlphaFold has been used by over 3 million researchers, and Hassabis believes similar AI tools could lead to 'second-order Nobel' prizes by enabling new discoveries.
Co-scientist, a fine-tuned version of Gemini, assists in hypothesis generation, data analysis, and literature summarization, acting as a research assistant.
DeepMind is developing a suite of AI models for drug discovery, aiming to replicate AlphaFold's success across various stages of the process, with a goal of curing all diseases in the next 10-20 years.
Automated labs are being developed for material science and potentially other scientific domains, but the complexity of physical verification and safety concerns for 'no human in the loop' systems are significant challenges.
AI's role in health and personal use cases
The conversation begins with personal anecdotes highlighting the impact of AI, particularly Gemini. One story shares how Gemini's long context analysis of a medical scan provided reassurance and was later verified by a doctor, underscoring its potential in healthcare. Demis Hassabis confirms numerous such anecdotes where Gemini has been used for health reasons, sometimes life-saving, affirming its incredible use case. He also mentions Gemma, a free local AI that can offer similar benefits with added compassion. This illustrates AI's growing accessibility and its ability to assist in critical personal situations, moving beyond purely research applications.
AI as a creative and critical thinking partner
Demis Hassabis describes his personal use of AI, particularly Gemini, as a 'sparring partner' for brainstorming project ideas and creative concepts. He also uses it to quickly grasp the key points of new research areas outside his expertise. While not yet used as a confidant in the way Jensen Huang of NVIDIA uses LLMs, Hassabis engages with AI to critically examine his ideas, framing it as a collaborative process to think through steps and uncover potential flaws. This highlights AI's utility in augmenting human creativity and critical analysis, acting as a sounding board for complex thought processes.
The 'second-order Nobel' and AI-driven scientific discovery
The discussion touches upon the profound impact of AlphaFold, with over 3 million researchers utilizing it. The idea of a 'second-order Nobel' Prize—awarded for a discovery *made using* AI tools like AlphaFold—is introduced. Hassabis finds this concept compelling, suggesting it's possible given the widespread use of AlphaFold in impactful research. This points to a future where AI not only assists scientific inquiry but actively enables groundbreaking discoveries that could lead to entirely new levels of recognition and advancement. The potential for AI to accelerate scientific breakthroughs is immense, shifting the paradigm of invention.
Co-scientist: an AI research assistant
Co-scientist is presented as a fine-tuned version of Gemini, enhanced with specific tools for hypothesis generation, data analysis, and literature summarization. It's conceptualized as an advanced research assistant for daily work. The hypothesis generator has shown impressive results, even on niche topics like ray tracing, providing sensible ideas after users narrow down their initial concepts. While currently assisting scientists rather than autonomously discovering, earlier versions of similar systems have improved computer science algorithms, like matrix multiplication, and enhanced tools like AlphaFold. The focus is on improving efficiency and pushing the boundaries of what's possible in scientific exploration. The co-scientist aims to empower researchers by handling complex analytical tasks, freeing them to focus on higher-level conceptualization and innovation.
A platform for curing all diseases
Hassabis reiterates his belief, stated in April 2025, that AI could help cure all diseases within a decade. He views the development of AI tools for drug discovery not as a gradual process, but as a platform-building effort akin to AlphaFold. This platform will comprise numerous 'AlphaFold-level' models addressing different aspects of drug discovery beyond protein structure, such as protein interactions, molecular binding, and predicting *in vivo* properties like absorption and toxicity. The goal is to integrate these models and test them on disease profiles, creating an engine applicable to nearly any disease. He likens it to AlphaFold's ability to accurately fold all 200 million proteins once the accuracy threshold was met. This ambitious endeavor is currently in pre-clinical stages, with potential breakthroughs expected in the coming years.
Accelerating drug discovery and clinical trials
The drug discovery process is complex, involving numerous steps beyond initial predictions. Hassabis explains that AI is being developed to predict protein interactions, molecular binding, and pharmacokinetic properties (absorption, distribution, metabolism, excretion, and toxicity). While AI can significantly speed up drug discovery, he acknowledges that clinical trials still require substantial time. However, AI could also accelerate clinical trials by stratifying patients more effectively and predicting optimal dosages. This dual-pronged approach, addressing both discovery and trials with AI, could lead to a significant step-change in human health within the next 10 to 20 years, provided no fundamental laws of physics are violated. The process is expected to be exponential, similar to the human genome project, with rapid advancements manifesting after initial foundational work.
Challenges in automated scientific discovery
Fully automated discovery loops, where AI generates hypotheses and verifies them without direct human intervention, are discussed. Hassabis notes that while feasible in domains like coding and math due to easily verifiable outputs, physical sciences present greater challenges. Verification in physics, chemistry, or biology often requires automated labs, significantly lengthening the loop. DeepMind is exploring automated labs for material science, holding 200,000 material designs, some potentially superconducting. However, bottlenecks in hypothesis generation versus validation, and critical safety concerns for 'no human in the loop' systems, remain significant hurdles for recursive self-improvement in science. Robotics advancements are seen as key to enabling these automated physical verification processes.
AI in gaming and the future of AI creativity
DeepMind's partnership with the game EVE Online is highlighted as a sandbox for testing AI ideas, particularly concerning game mechanics, economies, and storylines. The potential for AI agents to play with, assist, or even act as game masters for players is explored. This collaboration, leveraging games as a proving ground, aligns with DeepMind's tradition of using this medium for AI development. The conversation also touches on the ultimate Turing Test: an AI inventing Nobel Prize-worthy science, akin to an 'Einstein Test squared,' where an AI with a knowledge cutoff must create new scientific paradigms. This implies AI's potential to eventually move beyond assistance to true autonomous invention.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Drugs & Medications
●People Referenced
Common Questions
Gemini, with its long context capabilities, has been used to analyze health scans, providing accurate initial assessments that were later verified by doctors. It has also offered life-saving advice in some cases.
Topics
Mentioned in this video
A groundbreaking AI technique used by over 3 million researchers, with the potential to lead to prize-winning inventions. It is also a component of a larger platform being built for drug discovery.
A cloud service providing powerful NVIDIA GPUs for running AI models like DeepSeek, used by the host.
Long context AI model that analyzed a health scan, providing accurate and reassuring results verified by a doctor. It's also used for life-saving advice and as a brainstorming partner.
A large AI model with 671 billion parameters that the host uses on Lambda GPU Cloud for running chatbots and experiments.
A free local AI model that can perform similar tasks to Gemini, considered a 'gift to humanity'.
A fine-tuned version of Gemini with extra tools for hypothesis generation, data analysis, and literature summarization, acting as a research assistant.
A spin-out company where they are building tools and a platform for drug discovery, related to AlphaFold and advanced versions of it.
A company whose CEO, Jensen Huang, uses LLMs for decision-making.
The company where Demis Hassabis is CEO, known for AI breakthroughs like AlphaFold. They use games as a safe proving ground for AI ideas.
CEO of NVIDIA, who reportedly uses LLMs as a confidant for decision-making.
His Annus Mirabilis papers of 1905, including special relativity, serve as a benchmark for an AI's creative scientific potential in the AI 'Einstein test'.
CEO of the company behind EVO Night, a friend of Demis Hassabis from back when Hassabis worked in games.
A person who expressed a desire to see someone use AlphaFold techniques to invent something prize-worthy, suggesting a 'second order Nobel'.
More from Two Minute Papers
View all 19 summaries
11 minThe Physics Bug That Stumped Everyone Is Finally Gone!
10 minAdobe & NVIDIA’s New Tech Shouldn’t Be Real Time. But It Is.
12 minThe Most Realistic Fire Simulation Ever
10 minNVIDIA’s Insane AI Found The Math Of Reality
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free