Demis Hassabis on AGI, agents, and AI-driven scientific discovery

Executive overview

Current AI systems are missing a small number of capabilities — continual learning, long-term reasoning, and consistent introspection — before reaching AGI. Hassabis puts AGI at roughly 2030 and is confident the existing paradigm (pre-training, RLHF, chain of thought) is part of the final architecture.

The core insight: agents are the path to AGI, and science is the highest-leverage application once we get there.

• Continual learning is the binding constraint holding agents back from autonomous, end-to-end task completion.
• AlphaFold-style breakthroughs follow a pattern: massive combinatorial search space, clear objective function, sufficient data or simulator.
• Deep-tech founders should plan around AGI arriving mid-journey — and design what they build accordingly.

What's missing for AGI

• Continual learning remains unsolved — current context windows are brute-force working memory, not graceful integration of new knowledge.
• Long-term reasoning has gaps: models can solve IMO gold-medal problems but still make elementary errors when questions are posed differently.
• Chain-of-thought is still primitive — models sometimes revisit a move they know is a blunder because they cannot find anything better; a precise reasoning system shouldn't do that.
• Introspection about its own thought process appears absent — the "jagged intelligence" problem.
• Hassabis estimates 50/50 odds that one or two genuinely new ideas are still required on top of scaling existing techniques.

Agents: where we actually are

• Agents need to actively solve problems autonomously; that is the necessary path to AGI.
• Current agents are useful for parts of tasks but don't adapt well to context — they are not yet "fire and forget."
• The technology has only recently crossed from toy demonstrations into genuinely valuable use cases.
• Running dozens of agents for 40 hours has not yet produced output proportionate to the input — the experimentation phase is not over.
• First signal of true delivery: a person in a room like YC produces a hit game or app that clearly used agents throughout — that hasn't happened yet.
• Expected timeframe for agents delivering full value: six to twelve months.

Reinforcement learning and the AlphaGo legacy

• All DeepMind's game-playing systems (Atari, AlphaGo, AlphaStar) were agent systems — designed to accomplish goals, make decisions, and plan.
• Chain-of-thought reasoning and thinking modes in today's frontier models are direct descendants of AlphaGo-era techniques.
• Monte Carlo tree search and related search methods from AlphaGo/AlphaZero are being re-examined at scale in foundation models.
• RL is likely underrated relative to where it will be in two to three years.

Distillation and efficient models

• DeepMind invented the distillation process; flash models are ~95% of frontier quality at ~one-tenth the cost.
• Serving billions of users across Search, Maps, YouTube, and Android creates a structural incentive to compress frontier capability into small models fast.
• No theoretical information-density limit has been reached yet — a frontier model's capability appears in edge-scale models roughly six to twelve months later.
• Edge/nano models are now open (Gemma) partly because they are vulnerable once deployed on-device anyway — open-sourcing unifies the Android/glasses/robotics stack.
• Local models processing personal audiovisual data, orchestrated by cloud frontier models only when needed, is the likely end state for privacy-sensitive applications.

The AlphaFold pattern for scientific breakthroughs

• Three conditions make a domain ripe: massive combinatorial search space, clear objective function, sufficient data or simulator.
• Both Go moves and protein configurations satisfy all three conditions — that's why both were tractable.
• Drug discovery fits the same frame: there exists a compound that solves a disease, the only question is finding it efficiently.
• AlphaFold is used in virtually every drug-discovery program globally; Isomorphic Labs is extending it across the full drug-discovery pipeline.
• Materials science, climate modeling, and mathematics are all approaching an "AlphaFold one moment" — promising results, grand challenge not yet solved.

Virtual cells and the limits of biology modeling

• A full virtual cell is roughly ten years away; a virtual nucleus is the near-term target as a self-contained slice.
• The binding constraint is data: imaging a live cell at nanometer resolution without killing it would convert the problem into a vision task AI already knows how to solve.
• Two solution paths: hardware/data-driven (better electron microscopy) or learned simulators of dynamical systems.

Genuine scientific reasoning versus pattern matching

• No AI system has yet produced a genuine massive discovery — only solved known hard problems.
• True discovery requires analogical reasoning beyond the bounds of the known; current systems do not demonstrably have this.
• Hassabis's "Einstein test": train a system on knowledge up to 1901 and see if it independently produces special relativity and the Annus Mirabilis papers.
• A harder bar: generate a new set of Millennium Prize problems that top mathematicians consider deep and worthy of lifetime study.
• Co-Scientist and AlphaEvolve are experimental systems pushing beyond base Gemini, but none has cleared the genuine-discovery threshold yet.

Advice for deep-tech founders

• Intercept where AI is going, then combine it with a deep non-AI technology domain — the interdisciplinary intersection is defensible.
• World-of-atoms problems (materials, medicine, chemistry) cannot be shortcut by the next foundation model update.
• Ideally, the founding team is expert in both machine learning and the domain being applied to.
• AGI timeline (~2030) means a ten-year deep-tech journey will likely see AGI arrive in the middle of it — plan explicitly for that.
• Specialized models (e.g. AlphaFold) will function as tools called by general models (Gemini, Claude), not be absorbed into one giant brain.
• Think through what physical infrastructure, finance systems, or domain tools would be valuable once AGI can orchestrate specialist systems as subroutines.