Most important take away

The era of pure scale-driven LLM progress is flattening, and the value of novel ideas (search, multi-modality, agents, new architectures) is about to spike again — creating a fresh window for startups that was effectively closed for the last five years. For builders and operators, the actionable move is to stop trying to out-scale OpenAI/Anthropic and instead bet on the next S-curve, while paying close attention to talent allocation, geographic arbitrage (the UK is dramatically under-leveraged), and building the infrastructure layer that autonomous agents will require.

Summary

Actionable insights and patterns from the conversation:

Career and entrepreneurship advice

• The best predictor of future behavior is past behavior. When evaluating founders, look for evidence of self-initiated action — things the person did without being told to by a teacher, parent, or boss. The form this takes varies by culture (a rural Indian candidate doesn’t need to have started a company; a Stanford grad who never thought about it is a negative signal).
• You can just say yes and figure it out. Matt’s foundational lesson at age 13 (fixing computers despite knowing nothing) is the meta-skill behind everything he’s done since. Treat permissionless action as a learnable habit.
• Talent is fungible. The same ambitious, high-skill person who becomes a Jane Street trader in London would have become a founder in the Bay Area. The “forced entrepreneurs” research shows people pushed into founding during recessions outperform the median voluntary founder — meaning ecosystems, not individual ability, are often the bottleneck.
• Peak performance matters more than average. Entrepreneurship rewards the best day from the best person on the team, not consistent median output. When evaluating co-founders, ask how good the best person is at their best.
• You can start a company with a stranger. The conventional “only co-found with someone you’ve known for years” rule is survivorship bias from a pre-EF era.
• Don’t spend seven years on something that can never be big enough for your ambition. The worst outcome is sunk-cost commitment to an under-scoped idea.
• Fatherhood (and any deep relationship) has no shortcuts — it compounds like other long-term investments. Status earned at work transfers to zero with your kids.

AI tech patterns and where to build

• LLM scaling is on the flattening end of an S-curve. Compute and data have driven nearly all recent progress; the marginal return on more of both is dropping. Ideas are about to matter again.
• Don’t try to compete by building bigger LLMs — that’s a capital game even Mistral-scale raises can’t win. Instead look for the next S-curve: search techniques (AlphaGo-style) layered onto LLMs, multi-modality (especially video as a world-model substrate), and agentic systems.
• Expect divergence, not convergence. GPT-4-class capabilities have converged across labs, but GPT-5 will likely incorporate a non-obvious idea, and labs are getting much more secretive about those ideas.
• Data is the hardest of the three inputs (compute/data/algorithms) to keep scaling — but most models haven’t ingested video at scale yet, so don’t write off data ceilings prematurely.
• The best companies will be defined by their ability to swap models. Build for model portability.
• Agentic infrastructure is the next big opportunity. Think of high-frequency trading as the analog: autonomous agents need protocols, governance, observability, kill switches, and edge-case handling. Whoever builds the “operating system for agents” in a key vertical owns a massive asset. Government should set standards; private companies should build the infrastructure.
• Apply Joel Spolsky’s “commoditize your complements” lens to every layer of the AI stack. Nvidia is eyeing models; model companies are eyeing chips. Decide which adjacent layer to commoditize to protect your core.

Geopolitics and macro

• China is more paranoid about AI safety than any Western government — their regulation is more onerous than the EU’s because the CCP prioritizes stability over speed. This makes a Chinese Sam Altman structurally hard to produce.
• US export controls on semiconductors are a real (non-binary) drag on China. Chinese-made GPUs are roughly 70% of Nvidia’s performance — meaningful friction that compounds.
• China bull case: domestic GPU ecosystem matures and industrial policy unleashes. Bear case: the compute gap holds long enough that Western labs hit the next S-curve first.
• The UK is structurally under-leveraged. DeepMind, Wave, top universities, ARIA, light AI regulation, and deep research talent are all here. The rate-limiting factors are self-imposed: planning permission for data centers, grid capacity, and pension funds with zero VC allocation. These are policy levers, not laws of physics.
• AI Nationalism: countries will become AI makers or AI takers. Takers ship dollars to California forever.
• Defense tech is both a commercial opportunity and a moral necessity. Cheap smart drones are asymmetric weapons that favor non-state actors against expensive assets like aircraft carriers — defensive tech will be critical.
• Nuclear war is underrated as a risk (read Annie Jacobson’s “Nuclear War: A Scenario”). Cybersecurity becomes more important than ever in a deepfake-capable world.

Investing patterns

• Maintain multiple radically different scenarios as a VC; don’t bet on certainty. Adjust the weights, not the worldview.
• Charlie Songhurst’s two superpowers worth copying: (1) rapidly modeling a founder’s peak capability, and (2) pitching the most ambitious version of the founder’s own idea back to them and reading their reaction. European founders self-censor ambition in pitches — force them to react to the ceiling.
• Retention beats hiring in AI research. London’s underrated advantage is that you can keep a world-class research team together longer than in the Bay Area, where Sam Altman will buy out your best people the moment you have a demo.

Chapter Summaries

1. Childhood and the origin of agency — Matt grew up in a small ex-industrial town near Bradford. A neighbor’s lightning-damaged computer at age 13 became his “you can just say yes” moment, leading to a teenage business fixing PCs and building websites. The lesson: permissionless action is the foundation of everything that followed.

2. Evaluating founders — Past behavior predicts future behavior, but the form of “entrepreneurial behavior” is culturally relative. EF interviews look for evidence of self-directed action, not specifically prior companies.

3. The flattening of the LLM S-curve — Most AI progress over the last five years has been scale, not new ideas. Returns to more compute and data on text are diminishing. The value of novel ideas is about to spike, opening a window for startups that don’t try to out-scale incumbents.

4. Where the next S-curve lives — Search techniques layered onto LLMs, multi-modality (especially video for world models), and agentic systems. Expect lab divergence rather than continued convergence as labs guard novel ideas more tightly.

5. Bottlenecks: compute, data, algorithms — Data is hardest to scale obviously, but video and interactive data are largely untapped. Don’t bet against smart people finding new data sources.

6. Buying OpenAI at $90B — Both Harry and Matt would pass. The thesis depends on whether OpenAI continues to lead with novel ideas, not just GTM.

7. China’s AI position — More sophisticated than commonly assumed (the Minister of Science and Technology is a computer scientist). But CCP prioritization of stability over ambition makes producing a Chinese Sam Altman structurally hard. AI regulation in China is more onerous than the EU’s.

8. The bull and bear case for China — Bear: US export controls hold and Western labs hit the next S-curve first. Bull: domestic GPU ecosystem matures and industrial policy unleashes once China has full sovereignty over the stack.

9. Verticalization across the AI stack — Nvidia moving into models, model companies moving into chips. Apply “commoditize your complements” to decide which adjacent layers to absorb.

10. Agents and infrastructure — GPT-5 won’t be agent-ready at scale, but within five years agents will need protocols, governance, and observability layers analogous to high-frequency trading infrastructure. Whoever builds that owns enormous value.

11. UK regulatory and capital environment — Light AI regulation, world-class talent, but blocked by local planning vetoes on data centers, an under-built electricity grid, and pension funds with zero VC allocation. All are policy choices that can be reversed.

12. AI Nationalism and the future of warfare — Countries become AI makers or takers. Cheap autonomous drones are asymmetric weapons favoring non-state actors. Defense tech is both commercially important and a values-defense imperative. Nuclear war risk is underrated.

13. Founders and synergies — Synergies matter, but peak performance from the top person matters more. Most founding teams don’t stay together long-term anyway. Many great founders are first-time founders who can’t predict what the journey will be like.

14. Talent allocation and the UK ecosystem — The “forced entrepreneurs” paper: talent is fungible, and the Bay Area’s advantage is that ambitious people default to founding rather than trading. The most aspirational job for a Cambridge CS grad is still Jane Street. Changing belief capital is the unlock.

15. EF’s lessons — Overreacted to short-term signals (seed round raises) early on; learned to keep funding exceptional 21-year-olds over good-not-exceptional 31-year-olds. About half of accepted EF cohort members pair up and get funded; ~two-thirds of those raise a seed.

16. Quick-fire — Most respected VC: Charlie Songhurst, for talent spotting and the “pitch back the most ambitious version” technique. Contrarian advice: you can start a company with a stranger. Mind-changed in 12 months: it’s not too late to build an AGI-scale company. Fatherhood: compounding investment with no shortcuts. Unasked question: writing immersive historical murder mystery games.