Demis Hassabis on Intelligence Squared
Demis Hassabis and Sebastian Mallaby on The Quest for Artificial General Intelligence (Part Two)
Worth hearing for a comparatively specific glimpse of DeepMind’s architectural experiments: adding AlphaGo-style planning and Monte Carlo tree search on top of transformers. It advances beyond Hassabis’s familiar list of missing AGI capabilities, though the claim of promising results remains vague.
And in terms of the architecture, you talk about Transformers and you talk about, that's 50, 50 in terms of the two different approaches, do you sense you and the team getting closer to an architecture that you're feeling more confident? Could be that pivot point? Yeah, I mean, it wouldn't be so much like, I mean, there could be something out there that's even better than Transformers, but I think that's unlikely. But I'm talking about additional things on top of the system, maybe with the thinking perhaps bringing in more AlphaGo ideas around planning Monte Carlo tree search. So we're trying out lots of these ideas and this, I would say we have a lot of promising results. But we, you know, we've got to see how that unfolds over the next year.
Hassabis discusses quantum computing, the continued importance of current AI systems, and possible additions to transformer-based architectures on the route to AGI. The strongest short exchanges emphasize reciprocal acceleration between quantum computing and AGI, alongside planning as a promising architectural direction.