Just caught the replay of Demis Hassabis talking at Y Combinator about where we actually stand on AGI, and honestly some of the takes are worth sitting with for a minute.

So here's the thing: according to the DeepMind founder, we basically already have most of the architectural pieces we need for AGI. Large-scale pre-training, RLHF, chain-of-thought reasoning—these are probably going to be part of the final architecture. But there are still one or two critical missing pieces. Continuous learning, long-term reasoning, and certain aspects of memory are still unsolved. His personal timeline? Around 2030. If that's even close to right, it changes how you should think about any long-term project you're building today.

What caught my attention was the "jagged intelligence" problem he described. Current models can solve IMO gold medal-level math problems but then make elementary school arithmetic mistakes on a different question. It's like the reasoning process has massive blind spots. He gave a chess example—sometimes Gemini realizes it's about to make a bad move but then makes it anyway because it can't find better alternatives. A truly intelligent system shouldn't work that way. The DeepMind team thinks fixing this might only require one or two specific improvements, but it's a clear gap.

On agents: Hassabis was pretty direct—we're just getting started. Everyone's experimenting, but we haven't really found the killer use cases yet. He mentioned that no one's created a top-ranking AAA game using AI coding tools despite it being theoretically possible with current capabilities. Something's missing in either the tools or the process. He expects to see real breakthroughs in agent applications within 6-12 months.

The memory discussion was fascinating too. Context windows of a million tokens sound huge until you realize that's only about 20 minutes of video streaming. And the current approach is basically cramming everything into those windows—important and unimportant data mixed together. The brain does this elegantly through sleep cycles and memory consolidation. DeepMind's been thinking about this since the DQN days back in 2013, drawing from neuroscience, but we're still using crude approaches.

On the distillation side: their hypothesis is that within 6-12 months of releasing a cutting-edge model, they can compress its capabilities into much smaller models that run on edge devices. They haven't hit any theoretical limits yet. The Gemma models are a good example—Gemma 4 performs exceptionally well for its size. This matters because it means AI that's fast, efficient, and private—running locally on your phone or robot instead of in the cloud.

What really stood out was his point on scientific breakthroughs. AlphaFold was huge—three million researchers worldwide now use it, and he's heard it'll be part of almost every future drug discovery process. But even that's just the beginning. He calls it the "Einstein test": can you train a system on knowledge from 1901 and have it independently derive what Einstein figured out in 1905? Once that works, we're close to systems that can actually invent new things rather than just solve existing problems.

For founders, his advice was direct: pursue problems that only you can solve if you don't. Don't optimize for easy. Also—and this is important—if you're starting a deep tech project today that's meant to be a ten-year journey, you have to factor in the possibility that AGI might show up halfway through. Think about whether your project can work with AGI, how it integrates, whether it remains useful in that world. His vision is specialized systems like AlphaFold working as tools that general-purpose models like Gemini can call on, not everything crammed into one massive model.

The multimodal angle for DeepMind is interesting too. Building Gemini multimodal from the start was harder initially, but it's paying off now—better world models, robotics applications, autonomous driving integration. That's becoming a competitive advantage.

Overall, the conversation painted a picture of AI progress that's rapid but still has specific technical hurdles to clear. We're not just scaling our way to AGI—there are actual problems that need solving. And for anyone building in this space, the timeline matters. Think about what remains valuable when the landscape shifts.