Skip to content
Dwarkesh PodcastDwarkesh Podcast

Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold

Here is my episode with Demis Hassabis, CEO of Google DeepMind. We discuss: * Why scaling is an artform * Adding search, planning, & AlphaZero type training atop LLMs * Making sure rogue nations can't steal weights * The right way to align superhuman AIs and do an intelligence explosion 𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒 * Transcript: https://www.dwarkeshpatel.com/p/demis-hassabis * Apple Podcasts: https://podcasts.apple.com/us/podcast/demis-hassabis-scaling-superhuman-ais-alphazero-atop/id1516093381?i=1000647410338 * Spotify: https://open.spotify.com/episode/6SWbwjYPs5WevIoCCiSByS?si=nCVFSRr7QGGI_STgbrOBDA * Follow me on Twitter: https://twitter.com/dwarkesh_sp 𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 00:00:00 - Nature of intelligence 00:05:56 - RL atop LLMs 00:16:31 - Scaling and alignment 00:24:13 - Timelines and intelligence explosion 00:28:42 - Gemini training 00:35:30 - Governance of superhuman AIs 00:40:42 - Safety, open source, and security of weights 00:47:00 - Multimodal and further progress 00:54:18 - Inside Google DeepMind

Demis HassabisguestDwarkesh Patelhost
Feb 27, 20241h 1mWatch on YouTube ↗

At a glance

WHAT IT’S REALLY ABOUT

Demis Hassabis explains path to AGI, scaling, safety, and Gemini

  1. Demis Hassabis outlines how large multimodal models plus planning/search (AlphaZero-style RL) are his likely recipe for AGI within the next decade, with current LLMs seen as “unreasonably effective” but still incomplete.
  2. He emphasizes neuroscience-inspired ideas like world models, imagination, and sample-efficient learning, and argues that combining old deep RL advances with new large-model scaling will be crucial.
  3. A large portion of the conversation covers safety: grounding, deception, evals, sandboxes, cybersecurity, responsible scaling, and governance that involves governments, academia, and civil society rather than just private firms.
  4. Hassabis also discusses Gemini’s development, multimodality, robotics, AI-for-science applications like AlphaFold, and how AGI and proto-AGI systems could accelerate future AI research while requiring tight control and oversight.

IDEAS WORTH REMEMBERING

5 ideas

AGI will likely require both scaled multimodal models and explicit planning/search.

Hassabis expects future AGI systems to use large multimodal world models (like Gemini) as priors, with AlphaZero-style planning and search on top to perform deliberate reasoning and goal-directed behavior that current LLMs lack.

Neuroscience remains a powerful guide for AI, especially around world models and imagination.

Concepts like reinforcement learning, attention, experience replay, and mental simulation were inspired by neuroscience and will inform unresolved areas such as planning, world-model construction, and imagination-like mental simulation in machines.

Scaling has gone further than expected, but its ultimate limits are still empirical.

Large models show emergent abstractions and partial grounding from language alone, which even scaling proponents didn’t fully anticipate; Hassabis argues we must push scaling and innovation in parallel to see whether we hit a soft asymptote or a “brick wall.”

Synthetic data and self-play will be central to overcoming data bottlenecks.

DeepMind plans to leverage realistic simulations, self-play between agents, and targeted synthetic data generation to fill gaps in training distributions, using careful data analysis to identify underrepresented regions and reduce bias.

Robust safety requires evals for deception and misuse, secure infrastructure, and sandboxes.

Hassabis stresses the need for better benchmarks (e.g., for deception, code exfiltration), hardened sandbox environments, strong cybersecurity, and even narrow AIs to help inspect more capable “general” systems before wider deployment.

WORDS WORTH SAVING

5 quotes

I wouldn’t be surprised if we had AGI-like systems within the next decade.

— Demis Hassabis

I sort of look at the large models today and I think they’re almost unreasonably effective for what they are.

— Demis Hassabis

The brain is an existence proof that general intelligence is possible at all.

— Demis Hassabis

You don’t want to be live A/B testing out in the world with these very consequential systems.

— Demis Hassabis

This is so consequential, this technology, I think it’s much bigger than any one company or even industry in general.

— Demis Hassabis

Definitions and mechanisms of intelligence, world models, and imaginationScaling laws, large language models, and the limits of the scaling hypothesisCombining LLMs with reinforcement learning, planning, and AlphaZero-style searchData, self-play, synthetic data generation, and multimodal groundingSafety, alignment, evals, deception, model governance, and cybersecurityGemini’s development, compute, efficiency, and DeepMind–Brain integrationAI for science and robotics: AlphaFold, theorem proving, coding, and RT-2

High quality AI-generated summary created from speaker-labeled transcript.

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.

Add to Chrome