Y Combinator

Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough

Key Themes

AGI architecturememory gapsagent systemsmodel distillationmultimodal AIscientific discoverydeep tech strategy

41 minApr 29, 2026

Summary

Demis Hassabis argues that AGI is close, but still missing a few crucial capabilities, while AI’s biggest near-term impact may be scientific discovery.

In this conversation, Demis Hassabis lays out a practical view of where AI is headed: the core ingredients for AGI already exist, but systems still lack durable memory, continual learning, robust reasoning, and consistency. He sees agents as an essential next step, views distillation and multimodal models as strategic advantages, and argues that open models matter for edge deployment and real-world use. The second half shifts toward science, where he frames AlphaFold, drug discovery, and a future virtual cell as examples of AI becoming the ultimate tool for research. For builders and investors, the message is to focus on deep technical problems in domains where AI can materially accelerate discovery, while designing for a world where more general systems arrive mid-journey.

Look for startups that combine AI with deep, hard-to-automate domains.

Hassabis repeatedly emphasizes materials, medicine, biology, and other atom-level sciences as especially durable areas where AI can compound value.

Favor founding teams that pair domain expertise with machine learning expertise.

He argues the best teams either contain a founder who is deeply technical in both areas or a team that jointly covers the AI and domain sides.

Build products that will still be useful if more general AGI arrives mid-development.

His advice is to assume AGI may emerge during a long deep-tech cycle and to design tools that can work with future general systems rather than only with today’s models.

Open and efficient models may be strategically important at the edge.

He highlights privacy, security, and cost as reasons smaller open models matter for devices like phones, glasses, and robotics.

Select any chapter text to Deep Dive with AI

01AGI Gaps: Memory, Reasoning, Agents, and Creativity

Demis Hassabis argues that today’s AI stack is close to the eventual AGI architecture, but still lacks a few key pieces: continual learning, long-term reasoning, memory, and better consistency. He connects current models back to DeepMind’s game-playing and RL roots, says distillation is making smaller models surprisingly powerful and useful on-device, and argues agents are still early but promising. The chapter ends on creativity, with Hassabis suggesting current systems can produce impressive outputs, but may still need either new technical breakthroughs or better human/tool collaboration to truly invent at a higher level.

Core AGI components like pre-training, RLHF, and chain-of-thought likely remain part of the final architecture.

Continual learning, long-term reasoning, memory, and consistency are still unsolved and may be required for AGI.

DeepMind’s earlier work on Atari, AlphaGo, AlphaZero, MuZero, and search still informs modern foundation-model reasoning.

Distillation is a major strength: small models can retain much of frontier capability while being faster, cheaper, and better for edge deployment.

Agents are viewed as early-stage but essential to AGI, with current use still more experimental than transformative.

Reasoning systems still show brittle behavior, loops, and elementary errors despite strong performance on hard problems.

Creativity is framed as the next test: not just making novel moves, but inventing whole systems from a high-level idea.

Humans using tools creatively may matter as much as model capability in the near term.

02Open Models, Multimodal AI, and What to Build Before AGI

Demis Hassabis argues that open models like Gemma are strategically important for edge devices and that multimodal systems such as Gemini provide a long-term advantage for robotics and real-world assistants. He says inference will remain constrained by physical bottlenecks even as it gets cheaper, so efficient use and agent swarms will matter. The conversation then shifts to science: AlphaFold, Isomorphic Labs, and the idea of a virtual cell as a future breakthrough in biology. Hassabis frames AI as the ultimate tool for science, recommends founding teams that combine deep technical and domain expertise, and explains that the best frontier problems resemble massive combinatorial search spaces with clear objectives and enough data or simulators. He closes by discussing scientific reasoning in AI, the possibility of AI generating novel hypotheses, and advice for builders to think seriously about AGI arriving mid-journey.

Open models are especially important for edge use cases like Android, glasses, and robotics.

Gemma is positioned as a competitive open model and part of a broader strategic open-source commitment.

Gemini’s multimodal design is presented as a core long-term advantage for robotics and assistants in the physical world.

Even if inference gets much cheaper, physical bottlenecks and Jevons-style demand expansion mean efficient use will still matter.

AlphaFold is presented as the template for AI-driven scientific breakthroughs: combinatorial search, clear objective, and enough data or simulator support.

A virtual cell is described as a long-term goal, with a virtual nucleus as an intermediate step.

Hassabis sees AI as the ultimate tool for science and expects major impact across materials, medicine, climate, and mathematics.

Founders should combine AI expertise with deep domain knowledge; interdisciplinary teams are especially defensible.

Current systems are close to scientific reasoning, but not yet producing genuinely novel discoveries or hypotheses at the level of major scientific leaps.

Builders should account for AGI arriving during a long deep-tech journey and design tools that can work with future general systems.