Y Combinator

Paul Graham On Startups, Ambition, and Great Founders

Paul Graham

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

• 0:50 Almost everything about starting a startup is the same whatever the technology of the era — microprocessors, AI or the internal combustion engine.

  startups

• 1:17 The complaint that Y Combinator has jumped the shark is permanent rather than observational — it has been made since 2008, and it concedes that YC was once great.
• 4:23 You do not displace Google by attacking it head-on; you wait until the world has changed enough to make its model obsolete, which is what OpenAI did.

  OpenAIGoogle

• 5:23 A founder needs ambition rather than a sense of duty, because a startup is grueling enough that dutifulness will not carry anyone through it.

  startupsambition

• 5:54 What actually motivates founders day to day is fear of disaster and of looking like a fool, not the prospect of becoming rich.

  startups

• 7:12 Ambition in founders is mostly inborn, which is why stories of a founder acquiring it during a batch are rare.

  startupsambition

All 16 from this episode, along the running time →

Michael Kratsios: Inside the White House's AI Strategy

MKMichael Kratsios

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

• 2:08 There is no place where a technologist can work on bigger problems than the US government — not even the biggest technology companies deal with problems of that scale.

  government

• 4:10 Whatever the industry, the thing standing between a technology and its market is usually a government rule, which makes policy the lever that unlocks technologies.

  government

• 4:52 Technology policy is not decided by one or two people in the White House: the decisions are extraordinarily federated across agencies.
• 7:40 The United States cannot lead in artificial intelligence with closed models alone: it needs a vibrant open-source ecosystem alongside them.

  America

• 7:58 Washington cannot make good technology policy in a vacuum: without startups and industry saying what they believe, the decisions come out worse.

  startupsAmerica

• 10:07 Rules for a technology that reinvents itself every six months must not draw firm red lines, because firm lines ultimately do not work.

All 20 from this episode, along the running time →

Susan Kare: Designing Icons & Graphics For the Original Mac

SKSusan Kare

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

• 1:33 Design is about making things meaningful and memorable: get the idea right and the form follows from it.

  design

• 7:00 A computer meant for anyone has to be understandable without a manual, and it has to be friendly.
• 7:31 You can be creative within any set of constraints, so long as you understand what they are and get excited about them rather than resent them.
• 12:33 An interface should say what it means with animation and sound rather than text, because then one version works internationally.
• 17:20 A plain drawing of a pencil stands for writing better than a glossy realistic pen, because the more detail an image carries the more particular the thing it depicts.
• 17:38 An icon should carry one or two salient details and nothing else.

All 19 from this episode, along the running time →

Chelsea Finn: This is the State of the Art in Robotics

CFChelsea Finn

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

• 4:06 Every machine-learning deployment that has paid off so far left a person making the decision, which is the only reason imperfect models were useful.
• 4:52 Physical AI has to make far fewer mistakes than any AI deployed so far, because it is only useful when it acts autonomously on the world.
• 7:18 Hand-tuning a robot's data set cannot reach very high reliability, because people get tired; the system has to seek out its own missing data.
• 8:56 Reinforcement learning cannot be scaled for robots the way it was for language models, because every attempt spends real robot hours instead of data-centre compute.
• 10:00 A robot going down a wrong path is not producing training data; the efficient move is for a human to intervene and show the recovery.
• 14:26 Whether a robot policy is any good is settled by running it for thirteen hours straight, not by whether it completes the task once.

All 22 from this episode, along the running time →

Peter Steinberger: "Fun Is Velocity"

Peter Steinberger

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

• 4:55 Product-market fit shows up as anger when you tell people they cannot have it yet, not as polite interest.
• 12:48 Anything you build can be forked or cloned, but your name cannot — so your personal brand matters more than any single product you will ever start.
• 14:16 A correction never travels as far as the scare it corrects: the press said a fifth of the project's skills were malicious, and the measured figure was 0.3 percent.

  journalism

• 15:19 Users like security in the abstract and experience it as breakage: hardening the project made things slower, harder to update, and broke what people depended on.
• 16:33 With agents a new feature is one prompt away, which means the whole cost of a feature has moved to what happens after you ship it.
• 16:56 Adding a config option so nobody breaks compounds into permutations no test suite can cover, because it is infinitely harder to evolve software that has users.

All 22 from this episode, along the running time →

Max Hodak: Average Is Not Good Enough

MHMax Hodak

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

• 4:25 Making a highly paid team wait a week to buy a tool at half price destroys far more value than it saves.
• 4:44 Approving purchases one at a time is the wrong place to control burn; spending review belongs earlier, in a budget people can make their own trade-offs inside.
• 6:54 If nobody can say what an experiment costs, everyone behaves as though experiments are free — it does not cost dollars, it costs media, and media comes from the fridge.
• 9:11 The best companies come out of scenes: a moment and a community that nucleate the founding team, and once that moment passes it is hard to get back.
• 10:09 There is no right hiring process, but there is a wrong one: not having something you do religiously as a company.

  hiring

• 11:47 Any one person or small group placed at the top of the hiring funnel will bottleneck the whole organisation, so first-round review should be distributed across the company as a vote.

  hiring

All 22 from this episode, along the running time →

Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work

DDDmitri Dolgov

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

• 2:11 The best moments in physical AI will look like nothing happened at all.
• 3:07 "Move fast and break things" is the wrong mantra when you are dealing with atoms; the job is to move fast and ship safely.
• 4:11 In the physical world a model's mistake is measured in human lives rather than tokens, and there is no retry button.
• 5:43 A physical AI agent has to be safe on day one, because you cannot ship something merely good enough and let users find the edge cases.
• 8:21 A working demo is at best one per cent of the work; the nines that follow it are where the real work is.
• 10:30 Waymo's demo took eighteen months and the product took fifteen years — the gap between the two is the whole job.

  self-driving cars

All 22 from this episode, along the running time →

Patrick Collison: Is AI Breaking the Lean Startup Playbook?

Patrick Collison

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

• 2:46 Knowing something yourself will stay faster than asking a model for it, because a lookup in your own head beats a round trip to an agent.
• 3:23 Renouncing your own cognitive ability because the models are good is premature, because companies still pay an enormous premium for it.
• 4:13 Models can prove monumental theorems and still have never written an essay worth reading.
• 7:23 Dropping out of college is not a trapdoor and the reputational cost is nil: as far as anyone can tell, nobody has ever cared.
• 8:20 Startup opportunities are not fleeting; Silicon Valley has reliably had a surfeit of them for decades.

  startups

• 9:30 This is not the last moment to start a company, and the belief that the window is about to close is a recurring millenarian instinct.

All 20 from this episode, along the running time →

Jeff Dean: The 1% Rule for Building in AI

JDJeff Dean

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

• 0:55 AI models have reached the level of a junior engineer on longer-running, agent-based coding tasks.
• 3:36 Inference, not training, is now the constraint, and specialized low-energy hardware will beat general-purpose GPUs and TPUs on latency.

  energysemiconductors

• 4:51 Agents can already be run for days or weeks on a single hard problem, and almost nobody has internalised that.
• 7:56 A chip that can do nothing but low-precision dense linear algebra is enough for machine learning, even though it cannot run Chrome.
• 10:05 Good ideas come from asking how you would solve a problem from first principles, rather than staying anchored on how it is solved today.
• 11:09 The numbers an AI engineer should know by heart are memory bandwidths and the energy cost of a single operation.

  energy

All 22 from this episode, along the running time →

Alexandr Wang: “This is a Once-in-a-Civilization Opportunity”

AWAlexandr Wang

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

• 6:49 The only way to build something that matters is to develop conviction in a set of beliefs nobody else agrees with yet, long before they become consensus.
• 7:33 A founder who takes their bearings from what everyone around them is saying will end up nowhere; you have to build your own compass for what the future looks like.
• 9:37 The bottleneck in AI is no longer the progress of the models but their diffusion — if the models froze today, the economy would still take decades to absorb what already exists.
• 10:09 This is a once-in-a-civilization opportunity to have a vision and impose it on the world by building something.
• 10:50 AI has ended the David-versus-Goliath era of startups: a startup that properly embraces agents can now easily outcompete incumbents.

  startups

• 11:49 Every one of the billions of people on earth will have their own superintelligence, tailored to them, and its point is to expand their agency.

  AI alignment

All 20 from this episode, along the running time →

Sam Altman: "Never a Better Time to Do a Startup"

Sam Altman

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

• 4:52 Startups are the main thing that keeps an economy from going stagnant, because established companies reliably drift towards being bad.

  startups

• 5:06 Startups matter more the bigger AI turns out to be, because they are what stops the technology's power from pooling in a few companies and models.

  startups

• 9:50 Startups beat incumbents when the technology is moving fast, costs are falling and cycle times are short — and all three are true right now.

  startups

• 10:30 Great startups do not arrive evenly; they cluster at the moments when the ecosystem shifts and incumbents lose most of their advantage.

  startups

• 11:31 In this era fluency with AI tools beats years of accumulated experience, so seniority is now a liability rather than an edge.
• 12:44 The single most valuable thing a founder can find is a belief they can hold with reason that the conventional wisdom says is wrong.

All 20 from this episode, along the running time →

Blake Scholl: How 50 People Built a Supersonic Jet

BSBlake Scholl

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

• 0:14 Half a century after Apollo and Concorde we can neither reach the moon nor fly faster than sound: the future stopped getting faster.
• 0:49 A defence industry that needs more than two years to build a single missile interceptor can neither build the future nor win a war.
• 1:09 Airliner design has not progressed in fifty years — the newest jet is no faster and makes the world no more accessible than the one that opened the jet age.

  design

• 2:26 When air travel doubles in speed the important changes happen on the ground: Hawaii was not a tourist destination until the jet made it reachable.
• 4:52 The incumbent manufacturers will never carry aviation into a supersonic age, because they have stopped building new aeroplanes at all.
• 10:12 The way to settle an argument about how quiet supersonic flight must be is to remove the boom entirely, so there is nothing left to argue about.

All 22 from this episode, along the running time →

Boris Cherny: We Cut 80% of Claude Code's Prompt

Boris Cherny

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

• 2:01 Prompt injection is no longer a live attack on the frontier model: it simply does not follow instructions it reads on the internet any more.

  prompt injection

• 3:01 Prompt injection can be caught by watching the neurons that fire when it happens, so a defence no longer depends on the model reporting the attack.

  prompt injection

• 4:01 Prompt work does not transfer across model generations: what you wrote three months ago for one model may be worth nothing to the next.
• 5:02 A frontier model is measurably more intelligent with no system prompt at all; the prompts that remain are there for the product, not the model.
• 6:55 Every six months you should delete your whole AI configuration — instructions, skills and hooks — and find out how much of it the model never needed.
• 7:49 You cannot predict which instructions a model needs, so a system prompt has to be discovered by running the thing rather than designed in advance.

All 20 from this episode, along the running time →

Jensen Huang: The Mindset That Built NVIDIA

JHJensen Huang

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

• 4:23 The technology a company is founded on can be exactly wrong and still not matter, so long as the founders confront reality and can learn.

  startups

• 6:07 The business is not building a great chip; it is accelerating an algorithm domain.
• 6:17 What makes a great company is a unique perspective about the world that its founders deeply believe in — not the technology and not even the market.

  startups

• 9:12 You do not invest in companies, you invest in people — which is why an honest founder who cannot deliver still gets kept alive.
• 11:26 AlexNet's real breakthrough was not computer vision but the discovery of a universal function approximator that can learn any function.
• 15:21 A CEO is in service of the company, so going deep into the technical weeds is a personality technique rather than a management technique.

All 20 from this episode, along the running time →