No Priors

From Restoring Sight to Reimagining the Brain, with Max Hodak

MHMax Hodak

Our page for this episode →Watch on youtube.com ↗

• 6:07 Moonshots succeed more often than Silicon Valley's use of the word admits -- we actually got to the moon.
• 7:13 Humanity is not good enough at molecular biology to make drugs that restore vision or hearing or stop Parkinson's.
• 8:16 Putting a device into the brain works in an hour, which makes it far more amenable to biology than drug discovery.
• 10:12 The brain is very literally a computer, and so is the universe -- this is not a metaphor.
• 12:26 Treating brain-computer interfaces as one investment category is like having a single "drug bet" -- the field is as varied as pharma.
• 13:32 You want to command a device explicitly, not have it act on a thought you happened to have.

All 18 from this episode, along the running time →

How Nuclear Will Unlock Energy Abundance with Valar Atomics Founder Isaiah Taylor

ITIsaiah Taylor

Our page for this episode →Watch on youtube.com ↗

• 0:01:44 Nuclear has never had its Ford moment: reactors have to be manufactured rather than constructed before fission can be cheap.
• 0:04:59 America stopped building reactors because Three Mile Island was mismanaged as a public-relations event, not because anyone was harmed by it.

  America

• 0:06:44 Small modular reactors are the right idea, but you only reach them by iterating on real hardware -- which is what everyone else missed.
• 0:07:53 Most of the nuclear industry is a modelling and simulation business producing paper reactors, not a business that builds them.
• 0:09:56 The legal pathway to test new reactors sat unused in American law for forty years, waiting for an administration to use it.

  America

• 0:16:49 Nuclear is empirically the safest form of energy -- safer even than solar, once you count the people who fall off roofs.

  solarenergy

All 18 from this episode, along the running time →

Really Big Test-Time Compute in AI Changes Benchmarks, Safety and Research with OpenAI's Noam Brown

NBNoam Brown

Our page for this episode →Watch on youtube.com ↗

• 0:13 A model's capability is now a function of how much money you spend on it, so asking what a model can do means nothing until you name a budget.
• 2:27 Model comparisons understate real progress, because benchmark tables do not control for how much test-time compute each answer used.

  benchmarksscaling laws

• 3:31 Running a model until its performance plateaus is no longer a usable evaluation rule, because a well-scaffolded model keeps improving for weeks.
• 4:02 A benchmark result should be reported under a stated budget, or as a curve against test-time compute — never as a single number.

  benchmarksscaling laws

• 7:18 Running a model five times and keeping the best answer buys a higher benchmark score without buying a better model.

  benchmarks

• 11:16 Within a year a model will write an entire state-of-the-art poker solver in one shot — the work of a whole PhD thesis.

All 18 from this episode, along the running time →

Re-engineering the Semiconductor Supply Chain with Intel CEO Lip Bu Tan

LTLip-Bu Tan

Our page for this episode →Watch on youtube.com ↗

• 4:41 A government taking a stake in its chipmaker is ordinary infrastructure policy — Taiwan, Japan and Singapore all did it.

  governmentinfrastructure

• 6:33 A foundry is a trust business before it is a technology business: a customer whose wafers yield badly misses its own revenue.
• 7:01 Selling silicon is no longer enough — customers now ask for the whole rack, so a chip company has to become a full-stack company.
• 8:14 Semiconductor infrastructure has not kept up with AI's growth, and closing that gap needs capacity built in unconventional ways.

  semiconductorsinfrastructure

• 11:10 AI's impact will be bigger than the internet's, and more profound.
• 11:56 The binding constraints on AI's growth are power, helium and memory — and almost nobody has noticed the helium one.

All 16 from this episode, along the running time →

“Curing All Disease by next century is too conservative" - Mark Zuckerberg

MZMark Zuckerberg

Our page for this episode →Watch on youtube.com ↗

• 2:30 Curing, preventing or managing all disease by the end of the century is now too conservative a goal.
• 17:32 This work could have been run as a business; open-sourcing it simply puts the tools in more scientists' hands faster.
• 18:06 Biology models are smaller than language models only because there is less data, not because biology is simpler.
• 18:21 There is no factory you can pay to produce biological data; each new dataset requires inventing a new scientific method first.
• 19:09 Tools this complicated need a ten-to-fifteen-year horizon, which is why they belong in a philanthropy rather than a startup.

  startups

• 19:41 The point is not to cure diseases but to build the tools with which the whole scientific field cures them.

All 15 from this episode, along the running time →

Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI

Andrej Karpathy

Our page for this episode →Watch on youtube.com ↗

• 0:01:45 The way software gets built flipped in December: writing code yourself is now the exception, not the default.
• 0:03:36 What limits you with coding agents is your own skill at stringing them together, not the capability of the models.

  coding agents

• 0:05:54 The engineer's resource constraint has moved from flops to tokens: what matters now is the token throughput you command.
• 0:13:11 The apps built to drive smart devices should not exist at all: the devices should expose APIs and agents should call them directly.
• 0:14:12 Software's customer is no longer the human but the agent acting on their behalf, and the whole industry has to be refactored around that.
• 0:17:20 The goal is not a better prompt but your own absence: arrange it once, hit go, and keep agents running for long stretches without you.

All 20 from this episode, along the running time →

How Until is Engineering Away Time in Organ Transplants | Laura Deming

LDLaura Deming

Our page for this episode →Watch on youtube.com ↗

• 2:42 Aging is neglected not for technical reasons but because it is not yet socially classified as a disease.
• 3:09 The areas worth working on are the ones with a social blind spot around them.
• 8:38 Medicine needs an ambulance of the future: a way to pause a dying patient's biological time until the cure for their disease arrives.
• 11:43 What will limit medical hibernation is not the technology but the social cost -- you cannot take the people around you with you.
• 13:08 Cryopreservation is tractable because ice formation is a probabilistic process you can steer, not a threshold that flips.
• 14:44 We already reversibly cryopreserve whole human bodies, at the few-hundred-cell stage, so the open question is scale and not possibility.

All 15 from this episode, along the running time →

NVIDIA’s Jensen Huang on Reasoning Models, Robotics, and Refuting the “AI Bubble” Narrative

JHJensen Huang

Our page for this episode →Watch on youtube.com ↗

• 0:05:27 AI is software that is not pre-recorded: unlike Excel it generates every token for the first time, every time, which is why it needs factories.
• 0:06:23 AI is infrastructure on a par with energy and the internet, because every application, company, industry and country uses it.

  energyinfrastructure

• 0:08:40 Geoffrey Hinton's prediction that AI would take over radiology came true, and the number of radiologists went up anyway.
• 0:12:30 Robots will close a severe and worsening labour shortage rather than take jobs, because ageing economies already cannot find the workers they need.
• 0:13:24 A world of a billion robots will need maintenance crews and hubs the way robotaxis already do, creating the largest repair industry on the planet.

  self-driving cars

• 0:22:11 No company is anywhere near a single God AI, and no researcher has any reasonable ability to build one.

All 20 from this episode, along the running time →