ExpertsFMHow to Build the Future: Demis Hassabis
Y Combinator Startup Podcast · April 29, 2026
5 Selected Lessons:
reasoning_pattern · 2 priority · 1 case · 1 framework · 1
Prioritize 'root node' problems whose solutions unlock whole new branches
Source
Y Combinator Startup Podcast — How to Build the Future: Demis Hassabis (April 29, 2026) · at 28:17 ▸
When this applies
when allocating scarce research/engineering effort across many possible problems and wanting outsized downstream impact
The takeaway
Rank candidate problems by how many downstream efforts/domains depend on them, and prioritize the ones that unlock the most new avenues, not the easiest or most isolated wins.
Why it works
Solving a foundational problem that many other efforts depend on (like protein structure) cascades value across an entire field, far more than solving an isolated problem.
One level up
Leverage comes from solving upstream dependencies, not from maximizing the number of problems solved.
Where it stops holding
This stops working when it's unclear which problems are truly 'root nodes' versus merely appear foundational, risking misallocated effort on a non-pivotal problem.
Detail
hierarchy:
- Root node problems that unlock whole new branches of science (highest priority — e.g. AlphaFold/protein structure)
- Problems with promising but not-yet-solved 'grand challenge' status (materials, climate, math — 'almost like an alpha fold one moment')
- Isolated or narrow problems without broad downstream dependency (lowest leverage)
rationale: specifically, I was meaning to solve other what I call root node problems in science. So areas of science that would unlock whole new branches or avenues of discovery.
In their words
So areas of science that would unlock whole new branches or avenues of discovery. And AlphaFold is the prototypical example of what we want to do.
The three conditions for an 'AlphaFold-style' breakthrough
Source
Y Combinator Startup Podcast — How to Build the Future: Demis Hassabis (April 29, 2026) · at 33:23 ▸
When this applies
when assessing whether any hard search/optimization problem is a good candidate for a data/compute-driven solution versus requiring a different approach
The takeaway
Target problems that have (1) a combinatorially massive search space, (2) a clearly specifiable objective function, and (3) sufficient real or synthetic in-distribution data/simulator — these three jointly make a problem tractable for today's AI methods.
Why it works
Massive combinatorial search spaces defeat brute force or special-case algorithms, but become tractable once you have a clear, hill-climbable objective function and enough real or simulated in-distribution data to train on.
One level up
Look for 'needle in the haystack' problems where the haystack is enormous but the needle's quality is measurable, because measurability is what converts search into optimization.
Where it stops holding
This stops working when the objective function can't be clearly specified (e.g. subjective or multi-faceted goals), or when there isn't enough data/simulator to generate in-distribution training signal.
Detail
procedure:
- Check that the problem has a massive combinatorial search space (no brute force/special-case algorithm works)
- Specify a clear objective function you can hill-climb (e.g. minimize free energy, win probability)
- Ensure enough real data and/or a simulator that generates sufficient in-distribution synthetic data
- If all three hold, apply today's search/RL methods to find the needle in the haystack
rationale: the lesson I've learned from all the alpha projects we've done, specifically AlphaGo and AlphaFold, is I think the techniques we have and the problems I like to look for are great in if the situation can be described as massive combinatorial search space... And then you have a clear objective function.
In their words
And then you have a clear objective function. So, you know, you could think of it as minimizing the free energy in the proteins, or winning the game of go.
Old specialized techniques generalize into new general systems
Source
Y Combinator Startup Podcast — How to Build the Future: Demis Hassabis (April 29, 2026) · at 3:25 ▸ · at 12:33 ▸
When this applies
when a mature specialized technique solved a narrow problem well, and you now face a broader, more general version of the same underlying challenge
The takeaway
Revisit and re-apply proven techniques from a narrow domain at greater scale and generality, rather than assuming they were domain-specific dead ends.
Why it works
Techniques like Monte Carlo tree search and RL, originally built for tractable narrow domains (games), encode general problem-solving principles that can be re-applied at scale to broader domains like language and reasoning.
One level up
Breakthroughs often aren't single-use; their underlying mechanism is reusable once you abstract away the domain it was first proven in.
Where it stops holding
This stops working when the narrow technique's success depended on domain-specific structure (e.g. perfect information, discrete action spaces) that doesn't transfer to the new general domain.
Detail
approach: Treat early narrow-domain systems (games) as proofs of a general mechanism, then re-scale and generalize them to broader domains (language models).
signals:
- technique worked well in a constrained/tractable version of the problem
- the new domain shares the same underlying structure (goal-directed decision making, search)
rationale: 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.
In their words
I actually think there's a lot of work we did back then that is relevant today. And we're sort of re-looking at some of those old ideas at scale today in a more general way, including things like Monte Carlo tree search and other ways of doing, augmenting the RL on top of the reinforcement learning we're ready to do today.
Multimodal-first design pays off later in adjacent domains
Source
Y Combinator Startup Podcast — How to Build the Future: Demis Hassabis (April 29, 2026) · at 22:30 ▸
When this applies
when choosing between a narrower, faster initial build versus a broader foundation that costs more upfront but unlocks future adjacent capabilities
The takeaway
Accept short-term difficulty/slower initial progress for a broader foundational design if it positions you to dominate adjacent, higher-value future domains.
Why it works
Building multimodality in from the start is harder initially (slower early progress vs text-only competitors) but compounds into advantages in downstream domains (world models, robotics) that require understanding the physical world.
One level up
Architectural choices made early can become durable competitive advantages years later, even if they look like a handicap at the time.
Where it stops holding
This stops working when the adjacent future domains never materialize or when competitors catch up on breadth without paying the early cost.
Detail
situation: Gemini was architected multimodal from inception, making early development harder than text-only competitors.
decision: DeepMind stuck with the multimodal-first approach, betting on long-run gains.
outcome: Hassabis claims this now underlies advantages in robotics, Waymo, and world-model building (Genie), with Gemini 'far and away the strongest models on those types of problems.'
lesson: An early architectural investment that's costly can become a long-term moat in adjacent domains.
In their words
I think that's a sort of still a slightly underappreciated aspect of the Gemini series is we started it being multimodal from the start. That made it a little bit more difficult actually to begin with because than just focusing on text, for example. But we believe we're going to gain from that in the long run.
Pursuing AI despite near-universal disbelief, 2010
Source
Y Combinator Startup Podcast — How to Build the Future: Demis Hassabis (April 29, 2026) · at 30:36 ▸
When this applies
when pursuing a contrarian deep-conviction idea that experts and gatekeepers actively dismiss as already-disproven
The takeaway
When pursuing a contrarian idea, ensure your commitment is rooted in genuine passion/conviction strong enough to survive total external disbelief, since validation will not come until after the fact.
Why it works
Deep personal conviction and intrinsic passion sustain effort through long periods with no external validation, which is necessary because truly novel ideas look indistinguishable from failed ones until proven.
One level up
Before success, contrarian pursuits are indistinguishable from failures to outside observers — only intrinsic motivation bridges that gap.
Where it stops holding
This stops working when the contrarian belief is simply wrong, and years of disbelief that it's a dead end turn out to be correct rather than a temporary obstacle.
Detail
situation: In 2010, AI was considered a niche, already-failed subject by investors and academics ('we know it doesn't work').
decision: Hassabis and co-founders proceeded anyway, driven by personal conviction and passion, willing to keep working on it even if it stayed unsuccessful for decades.
outcome: DeepMind went on to build AlphaGo, AlphaFold, and Gemini; Hassabis notes 'once you've done it, you've done it. But before you've done it, people are arrayed against you.'
lesson: Conviction sufficient to work on something 'no matter what happened' is what let the bet survive the period before validation.
In their words
Like for me, I would have worked on AI no matter what happened. I just decided from a very young age, it was the thing that could be the most consequential thing I could think of.