Apr 29, 2026 · 40m · y-combinator
Demis Hassabis: Agents, AGI & The Next Big Scientific Breakthrough · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
At the Y Combinator x Google DeepMind Startups Day, CEO Demis Hassabis discusses the path to artificial general intelligence, key architectural challenges like continual learning and meta-cognition, the practical evolution of AI agents, and how AI will revolutionize fundamental scientific discovery.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 3.6 Guest teaching 5.5 Guest disagreement 0.6 The partners pushing back 0.1
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 24:11 Reframing the zero-cost inference assumption
Hassabis directly dismisses the host's premise that inference will become free, citing Jevons paradox and long-term hardware bottlenecks.
Hardest push from the partners ▶ 5:12 Tan argues 1M context is already plenty bigTan counters the need for immediate memory innovations by asserting that modern million-token context windows are already large enough for almost any practical use.
Biggest teaching moment ▶ 5:19 Explaining the limitations of brute-force contextHassabis breaks down why context windows are a naive substitute for working memory and episodic replay, pointing out that 1 million tokens covers just 20 minutes of live video.
The partners hold their own ▶ 22:19 Tan showcases his direct-speech Her demoTan brings concrete technical credibility to the conversation by discussing his hands-on experimentation building a real-time speech agent using native multimodal Gemini APIs.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Title Sequence: How to Build the Future | 4 | 5 | 0 | 0 | Tan delivers a glowing biographical introduction and asks a well-structured opening technical question about current AI architectures versus missing pieces for AGI. Hassabis politely outlines the remaining frontiers like continual learning and long-term reasoning. | |
| Continual Learning, Memory, and Context Windows | 4 | 7 | 2 | 1 | When Tan suggests 1 million context tokens feels plenty big, Hassabis gently schools him by drawing on his neuroscience background and noting that 1 million tokens represents only 20 minutes of raw video, arguing that current context windows are brute force rather than true memory. | |
| Reinforcement Learning, Search, and Gemini | 5 | 6 | 0 | 0 | Tan asks an insightful question connecting DeepMind's RL heritage (AlphaGo, MuZero) with modern Gemini architectures and distillation. Hassabis enthusiastically confirms and explains how search and distillation remain foundational at scale. | |
| Fast Models, Developer Productivity, and Edge Deployment | 4 | 5 | 0 | 0 | Tan brings up developer leverage anecdotes from Steve Yegge, while Hassabis expands on the utility of smaller edge models for low latency, privacy, and robotics orchestration. | |
| Y Combinator Startup School Announcement | 0 | 0 | 0 | 0 | Brief promotional voiceover segment announcing YC Startup School. | |
| Reality vs. Hype in Agent Capabilities | 3 | 6 | 1 | 0 | Tan asks about current agent hype, and Hassabis provides a grounded assessment, noting that despite high inputs, the ecosystem hasn't yet produced a standalone breakout hit or AAA game built solely by autonomous agents. | |
| Human-AI Collaboration and Deep Creativity | 4 | 7 | 1 | 0 | Hassabis challenges the capability of current systems by explaining that making move 37 in Go is not the same as having the deep creativity required to invent the game of Go itself. | |
| Open Source Strategy and Gemma Models | 5 | 4 | 0 | 0 | Tan demonstrates his own technical chops by mentioning a custom multimodal 'Samantha' client he built, prompting Hassabis to outline Google's open Gemma and edge AI strategy. | |
| The Economics of Inference and Compute Bottlenecks | 3 | 7 | 3 | 0 | When Tan asks what happens when inference is essentially free, Hassabis directly challenges the premise, invoking Jevons paradox and physical manufacturing bottlenecks to explain why compute will always be rationed. | |
| Transforming Scientific Discovery with AI | 4 | 6 | 0 | 0 | Tan asks about expanding beyond AlphaFold into cellular modeling; Hassabis lays out DeepMind's roadmap toward virtual cells and the experimental imaging limits currently bottlenecking live cell data. | |
| Advice for Science AI Startups and Defensible Deep Tech | 4 | 5 | 0 | 0 | Tan invites Hassabis to advise biotech and science founders on building real defensibility rather than shallow model wrappers. Hassabis urges founders to tackle interdisciplinary deep tech and atoms. | |
| The Formula for AlphaFold-Style Breakthroughs | 4 | 8 | 1 | 0 | Hassabis lays out his exact framework for AlphaFold-scale breakthroughs (massive combinatorial search space, clear objective function, robust simulation/data) and proposes the 'Einstein test' for true scientific reasoning. | |
| Advice for 25-Year-Old Builders and Future of AGI | 3 | 6 | 0 | 0 | Tan concludes with a reflective question for young builders, and Hassabis delivers advice on taking on hard problems while explicitly planning for AGI arriving mid-journey. |
Statements from this episode (32)
Hassabis: Current AI paradigms will be part of final AGI architecture
“The components that you just mentioned, I'm pretty sure will be part of the final architecture for AGI. So I think they've come such a long way now and we've proven out so many things about what they can do. I can't see a world in which we will sort of realize…”
Hassabis: Continual learning, reasoning, and memory are required for AGI
“Continual learning, Long-term reasoning. Some aspects of memory. These are still unsolved. And how to get the systems to be more consistent across the board. I think all of these are going to be required for AGI.”
Hassabis: 50% chance AGI requires one or two new big ideas
“It could be that there's still one or two big ideas left that need to be cracked. I don't think it's more than one or two, if there are out there. And I think You know, my betting is about fifty-fifty if that's the case.”
Hassabis: Information retrieval costs remain non-trivial even in massive context windows
“Even though we're working on machines, not biological brains, and so potentially you could have, you know, millions or tens of millions size context window or memory, and it can be perfect. There's still a cost to looking it up and finding the right thing that…”
Hassabis: A million-token context window holds only 20 minutes of live video
“And then the problem is if you're then trying to try and process live video, and you're just going to naively record all the tokens, then actually a million tokens isn't that much. It's only like 20 minutes.”
Hassabis: Thinking modes and chain of thought revive AlphaGo techniques
“Really you can think of a lot of The things we're doing today, all the leading models with thinking modes and chain of thought reasoning as aspects of what was sort of pioneered with AlphaGo coming back now.”
Hassabis: Next few years of AI advances will come from RL and search
“And I think a lot of those ideas, both from AlphaGo and AlphaZero are really, really relevant to where we are with today's foundation models. And I think a lot of that is what we're going to see of the advances the next few years.”
Hassabis: Frontier model capabilities reach tiny edge models within 6-12 months
“But I think for now there's the assumption we make is that, you know, a year later after one of our leading, you know, pro models or frontier models goes out half a year later, you'll have them in the,
The really tiny, almost edge models.”
Hassabis: AI is far from reaching theoretical limits in model distillation
“So I didn't really see any limit yet in terms of like some kind of theoretical limit.
I think we're still pretty far off of that.”
Hassabis: Lack of continual learning holds AI agents back from full tasks
“I think that's one of the not having continual learning currently is one of the things holding back agents from doing full tasks. You know, I think they're really useful for aspects of tasks right now, and you can patch them together and do some really cool th…”
Hassabis: All leading foundation models are poor at playing games
“That all the leading foundation models are pretty poor at games, which is quite interesting.”
Hassabis: Gemini sometimes identifies chess blunders and plays them anyway
“And so what we see is that, you know, sometimes it will consider a move. It will realize it's a blunder, but it can't find anything better, so it kind of goes back to that move and does it anyway.”
Hassabis: Fixing AI reasoning gaps may only require one or two tweaks
“So there's just sort of huge gaps, I think, still, but it may only be one or two tweaks that are required to fix those kind of gaps, just to be clear, but I think that's pretty, pretty obvious there are there, and that's why you get this kind of jagged intelli…”
Hassabis: Active agent systems are the necessary path to AGI
“You have to have an active system that can actively solve problems for you to get to AGI. That was always clear to us. So agents are that path.”
Hassabis: AI builds Theme Park prototype in 30 minutes vs 6 months
“I can do a prototype of theme park in half an hour now, which took me six months back when I was 17.”
Hassabis: AI agents will deliver full commercial value in 6 to 12 months
“I haven't seen the result yet, which I would expect once this is really delivering that full value, which I think will come in the next six to 12 months.”
Hassabis: Bestselling AI-assisted apps and games will precede full automation
“Games companies or, you know, other types of companies that have built some kind of best selling app, best selling game using these tools. That's what you should see first. And then more of that will get automated.”
Hassabis: Today's AI systems cannot invent a game like Go
“Clearly today's systems, I think can't do that.”
Hassabis: No lab has enough spare compute to train two maximum-scale frontier models
“Like, nobody has enough spare compute to just make two, you know frontier models at maximum size, right, with different attributes, so that's pretty difficult.”
Hassabis: DeepMind makes edge models open because on-device models are vulnerable anyway
“For now, what we've decided is that our edge models, the things we want to use for Android, and glasses, and robotics it's best that they're open models. Because they're vulnerable anyway on the, once you put them out on the surfaces, so they might as well be …”
Hassabis doubts inference will ever be free due to Jevons' paradox
“Yeah, I'm not sure inference will ever be essentially free. I mean, there's sort of Jevon's paradox and other things about like, I think we'll just end up using All of us will end up using whatever we can get our hands on, and you could imagine millions of age…”
Hassabis: Chip manufacturing will bottleneck AI inference for decades
“Certainly the energy, if we solve fusion or, you know, superconductors or, you know, optimal batteries or some set of those things, which I think we will do with material science, energy costs will be essentially zero, but there'll still be the physical creati…”
Hassabis: A full virtual cell simulation is probably 10 years away
“And I think we're about 10 years away probably from something like a virtual cell, like a full virtual cell”
Hassabis: Almost Every Future Drug Will Use AlphaFold During Discovery
“I was told by some of my, you know, pharma executive friends that, you know, almost every drug Discovered from now on will have used alpha fold at some point in it, in the drug discovery process.”
Hassabis: AI startups in physical sciences are safe from foundation model updates
“I think that those kinds of interdisciplinary teams, especially if it involves the world of atoms as well there's not going to be a shortcut to that, at least in the foreseeable future. Those are areas that are pretty safe from just getting swarmed by whatever…”
Hassabis: Three Conditions Enable AlphaFold-Style AI Breakthroughs
“The lesson I've learned from all the alpha projects we've done, specifically alpha go and alpha fold is I think the techniques we have and the problems I look like to look for are great in, if this, if the situation can be described as massive combinatorial se…”
Hassabis: Frontier AI has not yet made a massive scientific discovery
“I've yet to seen anything so far, and we all tinker with same things, you know, some math problems that are a little bit harder than IMO and so on. I haven't seen anything yet that is a true genuine, You know, massive discovery. That's my personal opinion.”
Hassabis: AI might solve a Millennium Prize math problem within years
“When I say just, we're now talking about just like solving the Riemann hypothesis or something. This would be obviously amazing. Well, one of the Millennium Prize problems, and maybe we're a couple of years out from doing that.”
Hassabis: AI will eventually invent new foundational math and science problems
“I do think these systems will be eventually be able to do that. Maybe we were missing one or two things.”
Hassabis proposes the 'Einstein Test' to evaluate true AI scientific creativity
“The way we would test that is, you know, sometimes call it my Einstein test, which is, you know, can you train a system With the knowledge of cutoff of 19 oh one. And then will it come up with the, you know, what Einstein did in 19 oh five, including special r…”
Hassabis predicts AGI will arrive around 2030
“Depending on what your AGI timeline is, You know, mine's like, 20, 30 or something like this.”
Hassabis: Frontier models will call specialized tools rather than subsume them
“One thing I can think see happening is Gemini, Claude, or one of these general systems making use of alpha fold, like specialized systems as tools. I don't think we're going to have it just in one giant brain because it will have too much regression in if I pu…”