Results
142 articlesSort
44:21ENMercor's CEO: Knowledge Work's Next Job Is Writing EvalsHost Jacob EffronBrendan Foody tells Redpoint how Mercor turned AI hiring and expert data into a $100 million round — and why he thinks a huge slice of white-collar work will become creating tests for models.
1:54:08ENWhy Anthropic's first check went to a company building microscopes for neural netsHost Nathan LabenzGoodfire's Dan Balsam and Tom McGrath argue that mechanistic interpretability now has a proto-paradigm, two stubborn gaps, and three products worth shipping today.
38:25ENA thousand airplanes will never train an AI engineerFormer Airbus CTO Paul Eremenko explains why physical-world engineering still has no coding-AI moment, and how P-1 AI's agent Archie is supposed to get there anyway.
45:30ENA $20,000 AI Agent Replacing a $90,000 SDR: Manny Medina on Pricing in the Agent EraA fully loaded human SDR costs $70,000 to $90,000 a year; one AI SDR pitch is $20,000 per agent for 80% of the work, plus a bonus when targets are hit.
41:52ENMercor Hit $100 Million ARR Teaching Models to Hire. Humans Now Write the Exams.Host Sarah Guo, Elad GilBrendan Foody tells Sarah Guo and Elad Gil how Mercor automated recruiting for frontier labs — and why building evals may become the world's most common knowledge job.
43:51ENThe $177 Billion GPU Bet Only Works If AI Becomes a $1.5 Trillion MarketHost Jacob EffronFour Redpoint partners walk their limited partners through where AI value will accrue—and why revenue is now the easiest metric to misread.
1:04:52ENThe Consumer Playbook That Put Medical AI in 100,000 Exam RoomsOpenEvidence CEO Daniel Nadler explains how a free App Store product, trained only on peer-reviewed literature, reached as much as a quarter of U.S. physicians in a year.
40:33ENModerna was only company 18. The real product is the company factoryHost Sarah Guo, Elad GilNoubar Afeyan tells No Priors how Flagship Pioneering turned starting companies into a profession, underwrites scientific uncertainty, and wants AI to run discovery the way Waymo drives.
49:36EN86 Out of 100: Harvey's Playbook for Selling AI Into LawHost Sarah Guo, Elad GilWinston Weinberg on a Reddit blind test, expanding-and-collapsing product design, and why Harvey started with century-old firms instead of the mid-market.
3:45:44ENDecline Is a Choice: Marc Andreessen on Power, Censorship, and a Second Roaring TwentiesGuest Marc Andreessen · Host Lex FridmanThe United States is adding a trillion dollars of national debt every 100 days — and Andreessen says the boom is still there if the country wants it.
31:18ENElias Torres sold a billion-dollar company and called it a failure. Agency is the redo.Host Sarah Guo, Elad GilThe Drift founder wants $1 billion in revenue with 100 people, one email address, and software that works like a text to someone who actually shows up.
38:05ENFrom Zero to $20M ARR in Two Months: Bolt's Moat Lives in the BrowserHost Sarah Guo, Elad GilStackBlitz CEO Eric Simons tells No Priors how a five-year WebContainers bet, an open-sourced prompt, and a community of non-coders turned Bolt.new into an overnight full-stack generator.
1:27:33ENA 70B Model Booked Dinner 20% of the Time. MultiOn Took It to 95% in a DayHost Nathan LabenzMultiOn CEO Div Garg returns to The Cognitive Revolution to recap 18 months of agent hype, the AgentQ stack, and why 2025 might still be the year of vertical agents.
1:03:07ENReasoning Hasn't Budged in Two Years. Dust Still Saved a Customer 8,000 HoursDust co-founders Gabriel Hubert and Stanislas Polu on multi-model agents, refusing to buy GPUs before product-market fit, and the jump from calculator tools to stochastic ones.
1:06:15ENJailbreaks Got Harder. Scale Made One Attack Worse.Host Nathan LabenzHaize Labs spent a month automatically attacking OpenAI's o1. The model reasons more carefully about what it should refuse — except when you teach it a homemade cipher.
51:15ENAGI Has One Problem Left: Spend a Million Times More on a TokenMagic founder Eric Steinberger tells Sequoia’s Training Data why he trains his own models, why a software engineer is the MVP of AGI, and why 96% still is not a colleague.
37:47ENWorking Backwards from AGI: Magic's Five-Million-Token BetHost Sarah Guo, Elad GilEric Steinberger wants a coding colleague you trust like a teammate — and a world that is still a good place thirty years after AGI.
50:33ENThe 40% data-center save started with an email that ended in a question markPhaidra founder Jim Gao tells Sequoia how an AlphaGo-inspired bet at Google became reinforcement learning that runs billion-dollar plants.
3:43:34ENAt 27, Pieter Levels Stared at a Hotel Ceiling. Then He Shipped Twelve Startups in a YearGuest Pieter Levels · Host Lex FridmanLex Fridman talks with the indie hacker behind Photo AI, Nomad List, and Remote OK about viral AI products, digital-nomad loneliness, and why he still deploys PHP to production from a sofa.
1:52:00EN4,096 Numbers Have to Carry Every Concept. Goodfire Wants the KnobsHost Nathan LabenzDan Balsam and Tom McGrath explain how sparse autoencoders unpack a language model's residual stream — and why a public-benefit startup is trying to make that a product.
1:07:34ENGitHub Didn’t Believe Copilot Was Writing 25% of the CodeThomas Dohmke tells Sequoia how an unplanned GPT-3 experiment became Copilot, why GitHub still will not train its own foundation model, and how he wants a billion people writing software by 2030.
1:03:22ENA 40-person team at $35 million ARR: HeyGen made AI avatars a business, not a demoHost Nathan LabenzJoshua Xu and Benchmark’s Victor Lazarte tell Nathan Labenz how a camera-replacement product grew past 40,000 business customers — and why half the faces on this episode are fake.
1:32:45ENHe doubled a company with AI — then asked the labs to pause hyperscalingHost Nathan LabenzOn the Nick Halaris Show, Cognitive Revolution host Nathan Labenz argues for adopting today's models fast and slowing the race to systems more powerful than we are.
52:45ENA $150 million model, $5 million six months later: Benchmark’s AI investment mathHost Nathan LabenzEric Vishria and Quilter’s Sergiy Nesterenko explain why PCB layout is not a Copilot problem, why human board data would be the wrong training set, and which AI companies look like sugar highs.
Load moreShowing 24 of 142









