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

# Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough **TL;DR:** DeepMind founder Demis Hassabis believes current deep learning components (pre-training, RLHF, chain-of-thought) are part of the AGI architecture, but are missing continual learning, long-range reasoning, and memory. He predicts AGI around 2030, and emphasizes that reinforcement learning, agents, and distillation remain underrated. ## AGI Timeline & Key Missing Pieces Hassabis has been working toward Artificial General Intelligence (AGI) since his teens. He expects AGI to arrive around 2030, so any deep tech startup must consider this timeline. "It’s not necessarily a bad thing, but you have to take it into account." Regarding large-scale pre-training, RLHF, and chain-of-thought in the current paradigm, Hassabis thinks they "will be part of the final architecture for AGI," but "there are probably one or two things still missing." Specifically: - **Continual learning**: Current systems rely on ad-hoc workarounds (e.g., night-time dreaming cycles, experience replay) and cannot gracefully integrate new knowledge into an existing knowledge base. - **Long-range reasoning**: Models still make errors on simple reasoning, exhibiting a "jagged" intelligence—able to solve International Mathematical Olympiad problems but making basic arithmetic mistakes. - **Memory**: Million-token context windows are large, but they crudely store everything (including irrelevant information), and real-time video requires support for even longer time spans. "I think all of these are essential to achieve AGI." ## From Neuroscience to AI: Memory & Dream Cycles Hassabis's PhD research focused on how the hippocampus works, particularly the role of sleep (especially REM sleep) in memory consolidation. DeepMind's early Atari program DQN borrowed "experience replay" from neuroscience—replaying successful trajectories multiple times. "This goes back to 2013, the 'dark ages' of AI." But current methods are still unsatisfactory: "We shove everything into a context window—unimportant, wrong information. That doesn't seem right." Even with million-level context windows, there is a cost to finding relevant information. There is enormous room for innovation in the memory domain. ## Reinforcement Learning Is Underrated DeepMind has historically relied on reinforcement learning and search (AlphaGo, AlphaZero, MuZero). Hassabis believes RL is "somewhat underrated." The "thinking mode" and "chain-of-thought reasoning" in current leading models are actually a return to pioneering work from AlphaGo. "We are revisiting some old ideas at a larger scale and in a more general way, including Monte Carlo Tree Search." He predicts these ideas will become a significant part of progress in the coming years. ## Distillation & Small Models Hassabis emphasizes one of DeepMind's core strengths: first build the largest models to achieve frontier capabilities, then "quickly distill and package that capability into increasingly smaller and faster models." Distillation was invented by DeepMind founders (e.g., Jeff, Oriol), and the team remains world-class experts. Value of small models: - **Cost & Speed**: In 2023, a small model could achieve 95% of frontier model performance at one-tenth the cost. For iterative tasks like coding, the speed gain outpaces the 5% loss. - **Privacy & Security**: Can run locally on edge devices, processing personal audio/video data, only delegating to the cloud when necessary. - **Robotics**: Home robots require efficient, powerful local models. Hassabis believes we haven't seen the information-theoretic limit yet: "A year or half a year after our leading specialist models are released, you can see similar capabilities in very small, near-edge-device models." The Gemma 4 series proves this. ## Reasoning: Introspection & "Overthinking" Hassabis enjoys playing chess with Gemini and observing its thought process: "All leading foundation models are terrible at games." He gives an example where a model considers a move, realizes it's a blunder, but cannot find a better one, and eventually returns to the blunder. "This shouldn't happen in a precise reasoning system." He thinks "one or two tweaks could fix these gaps," but the gaps are "quite significant." An "introspective ability about its own thought process" might be the missing piece. ## Agents: Just Getting Started "You need a system that can proactively solve problems for you to get to AGI. Agents are exactly that path, and I think we are just getting started." Current agents are useful for tasks, but lack continual learning—they cannot be "set and forget." Making a system adapt to a specific context is the missing link. Hassabis believes agents are far from overhyped: "The potential is utterly incredible." --- **Source:** Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough - YouTube (https://www.youtube.com/watch?v=JNyuX1zoOgU)