Y CombinatorDemis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough
At a glance
WHAT IT’S REALLY ABOUT
Demis Hassabis on agents, memory gaps, and AI-driven science breakthroughs
- Hassabis argues current foundation-model paradigms will remain core to AGI, but continual learning, long-horizon reasoning, and better memory systems are still unsolved requirements.
- He frames agents as the practical path to AGI—active systems that plan and act—while noting today’s agent workflows are still early and often don’t justify heavy autonomous “swarms” yet.
- DeepMind’s RL/search heritage (AlphaGo/AlphaZero) is portrayed as directly relevant to modern “thinking modes,” with older ideas like tree search re-emerging at larger scale in more general settings.
- He expects rapid capability diffusion via distillation, making small/edge models nearly as strong as frontier models and enabling fast iteration, privacy-preserving local AI, and robotics assistants.
- He describes AI as a general-purpose scientific instrument: AlphaFold as the archetype, virtual-cell simulations as a ~10-year goal, and a near-term frontier of AI that can generate genuinely novel hypotheses (the “Einstein test”).
IDEAS WORTH REMEMBERING
5 ideasAGI likely needs a few missing capabilities, not a full paradigm reset.
Hassabis expects pretraining + post-training techniques to stay central, but highlights continual learning, robust long-horizon reasoning, and more principled memory as the main gaps, with a “one or two big ideas” possibility.
Context windows are not a real solution to memory.
Even huge windows are brute-force working memory; the hard part is storing the right things, filtering wrong/irrelevant info, and retrieving what matters with low latency—especially for long-running, video-rich personal assistants.
Agents are essential, but the ecosystem is still in the experimentation phase.
He sees agents as required for “active problem solving,” yet observes many multi-agent runs don’t yet produce outputs that justify the compute/time; he expects clearer killer workflows in the next 6–12 months.
Reasoning failures look like missing self-monitoring and control, not lack of raw capability.
He describes models that “overthink,” loop, or choose known-blunder moves (e.g., in chess), suggesting improvements may come from better oversight of intermediate reasoning and interventions during thought, not just larger models.
DeepMind’s game-era planning ideas are re-entering foundation models.
He connects chain-of-thought and “thinking modes” to AlphaGo-style methods and expects more explicit planning/search (e.g., Monte Carlo tree search-like augmentation) to drive near-term progress.
WORDS WORTH SAVING
5 quotesContinual learning, long-term reasoning, uh, some aspects of memory, these are still unsolved. I think all of these are gonna be required for AGI.
— Demis Hassabis
You have to have an active system, uh, that can actively solve problems for you to get to AGI. So agents are that path, and I think we're just getting going.
— Demis Hassabis
It's not enough to come up with move 37, like that's pretty cool, very useful. Um, but can it invent Go? That's what I w- I want a system that can invent Go if you give it a high-level description.
— Demis Hassabis
One was, step one was solve intelligence, i.e., build AGI, and then step two was use it to solve everything else.
— Demis Hassabis
Given Life's very short, and y- you know, you only have so much time and energy. You might as well put your life force into something that will really make a, a, a difference if you hadn't done it, if you hadn't been there to push it.
— Demis Hassabis
High quality AI-generated summary created from speaker-labeled transcript.
