What if the cure for a disease already exists inside the data, but nobody is allowed to look at it? Medical records across thousands of hospitals. Genetic data spread across dozens of countries. Research that can’t be shared because the privacy risks are too high. Andrew Trask believes we don’t have to choose between progress and privacy.
Andrew Trask is a research scientist at Google DeepMind, the founder of OpenMined, a community of over 18,000 researchers building technology that lets AI answer questions using data it can’t see, and a member of the United Nations Privacy Task Force. He’s spent a decade proving that the most sensitive data in the world can be used safely. In this conversation he lays out a paradigm shift that could unlock a million to a billion times more data than current AI systems use, and explains why the biggest problems in AI are not AI problems at all.
What You’ll Discover:
🎨 Taste, Trust, and Rare Data
- Three things AI needs from humans, and why each becomes a bottleneck as AI scales
- Why everyone is a curator of taste, whether they realise it or not
- How the rise of AI-generated music, art, and content invites a return to local culture rather than a flattening of it
🥼 The Breast Cancer Problem
- A billion screening images exist in the world. AI is trained on a few million. Why?
- How siloed data costs lives, and what attribution-based control could change
- The shift from one big model in one place to a network of neural networks distributed across the world
🔐 When AI and Cryptography Collide
- Why Andrew describes cryptography as AI’s yin and yang, and why their collision may be the most important convergence in technology right now
- Secure multi-party computation: how ten banks could trace where a dollar came from without revealing their ledgers to each other
- How OpenMined enabled Anthropic and the UK AI Safety Institute to do encrypted inference, where neither party had to share their model or data
🏛️ Broadcasting vs. Broad Listening
- Why the attention economy exists: there are far more mouths speaking than ears listening
- Why your senator cannot hear you, and what changes when AI lets them
- The four-year US election cycle was set by the speed of the postal service in the 1700s. What governance could look like when listening scales.
🌍 Finishing the Project of Civilisation
- Why our democratic, market, and free speech institutions are broken at the infrastructure level, not the surface level
- The trillion-dollar money laundering problem that could be solved with cryptography, not policy
- Andrew’s end state: a person able to speak to and listen to everyone in the world, instantly, all the time, for free
Key Insights:
“Many of the largest problems in AI are not AI problems. They’re cryptography problems.”
“The attention economy comes from the imbalance in scale between broadcasting and broad listening. There are way more mouths speaking than there are ears that can take it in.”
“We’ve been building civilisation with information technology that is radically incomplete.”
About Andrew Trask:
Andrew Trask is a research scientist at Google DeepMind, founder of OpenMined (a community of over 18,000 researchers working on privacy-preserving AI), and a member of the United Nations Privacy Task Force. He holds a PhD from the University of Oxford and has spent a decade at the intersection of artificial intelligence, cryptography, and data governance. His work focuses on building infrastructure that allows AI to learn from data it can’t see.
🎯 Perfect for: AI founders thinking about the next paradigm shift beyond scaling, builders working with sensitive or siloed data, leaders concerned about the power dynamics of centralised AI, and anyone wondering how we get the upside of AI without trading away privacy, sovereignty, or trust.
⏰ Timestamps:
00:00 – Cold open: a person who can speak to and listen to everyone in the world
01:03 – Introduction: the cure inside the data
02:23 – Do LLMs still believe everything they read?
03:54 – Taste, trust, and rare data: the three bottlenecks of the AI economy
12:36 – Why everyone is a curator of taste
16:44 – The pregnancy reveal song and the unwinding of mass culture
19:27 – Why AI still can’t detect breast cancer well
26:39 – The scaling laws are just getting started
29:32 – The 1.5 trillion dollar money laundering problem
32:22 – When AI and cryptography collide
35:22 – Secure multi-party computation, explained
39:40 – Trust isn’t the bottleneck if you build the right infrastructure
40:38 – Broadcasting vs. broad listening
43:19 – What a senator with AI broad listening could actually do
48:37 – Why this is the thing Andrew is most excited about
49:03 – Finishing the project of communication technology
Subscribe for more conversations that matter, and let us know in the comments: If your data could help save a stranger’s life without ever being seen, would you share it?
#AndrewTrask #OpenMined #PrivacyPreservingAI #Cryptography #BroadListening #DataSovereignty #FounderMindset #ArtificialHappiness