팟캐스트Hear the voice. See the shape of the thought.

Fiona Fung, who runs engineering and product for Claude Code and Cowie at Anthropic, walks through what broke when agentic coding became the team's default — review, ownership, planning, hiring — and the norms they rewrote to keep shipping. The throughline: when coding stops being the bottleneck, every process built around protecting expensive engineering bandwidth quietly stops working, and the manager's job is to notice and rewrite them fast. ## [00:00] Intro and the five themes Fiona opens with a confession that the room is much fuller than she expected (Boris and Jared's session is still letting out), takes a selfie with the audience, and frames the talk. Background: she grew teams at Meta and Microsoft before Anthropic, and is now responsible for Claude Code and Cowie engineering and product. The deck she's about to walk through has already been rewritten in the past month — routines didn't exist when she first wrote the slides. She previews five threads: bottlenecks have shifted, team norms had to be rewritten, how they rolled them out, what signals say the changes are working, and the open questions she's still sitting with. > *"I did this slide deck maybe like a month ago and already I've had to change some of the content cuz when I started this deck, there were no routines."* ## [02:10] The shift: bottlenecks have moved Fiona's subtitle for the whole talk is *what served you prior may not serve you any longer*. She takes the audience back to shipping Visual Studio 2005 on CD-ROMs — hard deadlines because the manufacturing lab had to print discs — and points out that the move from CDs to online distribution already rewired how teams ship. The new shift is bigger: for years coding throughput and engineering bandwidth were the expensive things, and that's quietly stopped being true on Claude Code. When the bottleneck moves, it doesn't disappear — it relocates to verification, review, cross-functional handoffs, and security. The questions that matter now are "is this code correct?" and "is this safe?", and the old planning and ownership norms quietly stop serving the team. > *"What served you prior may not serve you any longer."* ## [07:40] Rewriting team norms: code review, JIT planning, technical debates Inside Claude Code the team had to rewrite the norms one by one. Code review is the first — human judgment shifts to "who actually needs to look at this." Planning is the second — Fiona calls it JIT planning, like JIT compiling, because prototyping is no longer the expensive step that justifies a six-month roadmap. Technical debates are the third: code wins. Instead of two engineers arguing on a doc, both prototype the API and look at impact on callers, and Fiona made a point of caring about the API's downstream effects as much as the implementation itself. The unifying rule: when building is cheap and arguing is expensive, you don't let the last person who checks in win — you build the routines that get *you* the last word. > *"When building is cheap, arguing expensive, again, how does that shift your team norms a bit?"* ## [13:30] Routines and Claude as a second pair of hands With morning coffee Fiona now reads what a routine produced overnight rather than kicking off the work herself. The team leans on Claude code review heavily — Claude babysits PRs, handles styling, lint, and feedback requests, catches bugs before commit, and adds tests — while humans focus on the calls where trust is still being built. She also stresses product sense in tooling: she themed Claude's terminal output ice blue with snowflakes over the holidays, then pulls back to the bigger point that catching bugs earlier (shift left) and automating the double-click question matter more than any one tool. > *"Where do you trust Claude a lot, but then where do you still want a human?"* ## [16:45] Cross-functional gaps and hiring for the hard parts Fiona walks through a survey-update story: she didn't have a dedicated content designer, so Claude became her partner for terse, terminal-appropriate copy. Meanwhile PMs on the team write code, and engineers lean into PM work. The flip-side conclusion for hiring: non-traditional coders can now do more engineering, so the leader's job is to double down on the hard parts the team is actually missing. When she joined, Claude Code was strong on product generalists and creative folks but thin on distributed-systems expertise — that's where she pushed recruiting. > *"With Claude, you have non-traditional coders now being able to do more engineering, but you also have engineers that we can also now lean in to do other roles."* ## [18:51] Flat org and answering customer feedback yourself Fiona pushed her recruiters into an uncomfortable place: hire managers, but have them start as ICs first. The recruiter thought she was crazy; Fiona's answer is that dogfooding Claude Code is the job, and if a candidate isn't up for it the team is better off finding out early. Flat structure plus Claude as a context-switching aid is what lets her, as a manager, still ship code and answer customer requests directly from her desktop Claude Code — instead of routing every customer question through a triage system, she pulls up the local repository and answers it herself. > *"You want to hire managers and they will start as an IC first. No manager would be interested in that."* ## [25:00] Signals you're trending right and open questions The team's working metric is unglamorous and direct: every commit is cloud-assisted by default, and Fiona hasn't seen a non-Claude commit in roughly four months. But she warns against fetishizing the "X percent of code generated by AI" headline — throughput is one signal, not the goal. The end question is what product you're making more delightful and what problem you're solving, with quality and reliability watched alongside volume. She closes with the section she calls "audit your own effort," opens up the questions she's still asking herself, and hands suggestions back to the audience to take to their own teams. > *"For us, it's by default every commit is cloud-assisted. I don't think I've seen a non-cloud-assisted commit probably in the last 4 months or so."* ## Entities - **Fiona Fung** (Person): Director of Engineering at Anthropic, runs Claude Code and Cowie engineering + product; previously led teams at Meta and Microsoft. - **Boris** (Person): Engineering lead on Claude Code, frequent collaborator referenced throughout. - **Kat (Cat)** (Person): Anthropic colleague who gave a keynote earlier the same day on Claude code review. - **Claude Code** (Software): Anthropic's agentic coding tool that is now the default for the team Fiona runs. - **Cowie** (Software): Sister product Fiona's team also owns engineering + product for. - **Anthropic** (Organization): The company building Claude and Claude Code. - **JIT planning** (Concept): Fiona's term for shifting from a six-month roadmap to just-in-time planning, modeled on JIT compilation. - **Shift left** (Concept): Moving bug-catching and verification earlier — into automation and tooling — instead of relying on review after the fact. - **Routines** (Concept): Repeatable Claude-driven workflows the team relies on so a single human gets the last word on outcomes rather than the last commit timestamp winning.

Ben Horowitz and David Ulevitch — recorded at a16z's American Dynamism Summit in Washington — cover the full arc of what it means for a venture firm to accept industry leadership: from America's race to integrate AI into national defense, to the real reason the Anthropic–Department of War deal collapsed, to why the VC industry is consolidating around large generalist firms and narrow specialists. Horowitz closes on what he sees as America's most underrated strategic risk: a profound pessimism about AI at home while China and Japan charge forward with optimism. ## [00:00] Trailer The opening montage frames the episode's central tension: over 70% of Chinese citizens are optimistic about AI, while fewer than 30% of Americans share that view. David Ulevitch sets the stakes — a16z has placed the largest venture bet in American history on the proposition that the U.S. will win the next century of technology. > *"Over 70% of people in China are optimistic about AI and less than 30% in America were optimistic about AI."* ## [00:41] Why America's Technology Dominance Matters for the World Following a16z's record $15 billion fundraise — the largest in the firm's history — David Ulevitch asks what obligations accompany that scale. Horowitz reaches back to advice from his mentor Andy Grove: when you lead an industry, the entire industry's ethics and morality depends on you. He translates that into a first-principles argument: what matters for humanity is whether people have a genuine chance to contribute, and no country comes close to America on that dimension. Horowitz draws a direct line from the Industrial Revolution to the present moment. America won the 20th century because it had superior technology; the AI revolution presents an identical fork in the road. He frames a16z's mission as answering one question — what can the firm do to help America win technologically — and argues that every decision, from portfolio construction to government engagement, flows from that north star. > *"And so when I think about our role in the industry, it's what can we do to help America win technologically?"* ## [04:04] American Dynamism, AI & Catching Up to China Ulevitch asks what has most surprised Horowitz about investing at the intersection of national security and venture capital since launching the American Dynamism practice. Horowitz explains why American-style freedoms are structurally irreplaceable: the Declaration of Independence's claim that rights are self-evident — not granted by government — makes them nearly impossible to revoke, a feature no other country has replicated at the same strength. On the competitive landscape with China, Horowitz notes the pre-ChatGPT conventional wisdom gave China a large AI lead, primarily because China had integrated AI deeply into its military and government bureaucracy while the U.S. lagged far behind. The most heartening development since then has been the speed of American catch-up: a wave of entrepreneurs willing to serve the national interest, combined with a U.S. government genuinely open to new companies and willing to change procurement rules to accommodate them. > *"But the the thing that was true about the kind of old incorrect idea was that they were way ahead of us in integrating um their AI technology with uh their government you know on a kind of military basis on a bureaucracy basis you know and all facets and so you know when we started we were coming from I would say very far behind you know in that you know in that idea um the thing that's been surprising though is like how fast um we've been catching up."* ## [08:50] The Anthropic Deal: What Really Happened The conversation turns to the high-profile collapse of Anthropic's contract with the Department of War. Horowitz offers a deal-mechanics reading that cuts through the public framing: Anthropic had overwhelming leverage — they were already deployed, the country was heading toward conflict, no software vendor has ever had more negotiating power — yet they walked away. In Horowitz's view, that behavior has only one explanation: Anthropic wanted out of the deal, likely due to internal employee pressure, and used a philosophical disagreement as the exit ramp. He pushes back on the framing that a national security AI contract is ethically compromised. The Department of War operates under more rules and oversight than any private entity, and leaks are effectively guaranteed if those rules are broken. Ulevitch extends the point to founders more broadly: companies that let employees veto geopolitical decisions are substituting "vibe geopolitics" for the considered judgment of people who have studied — and sacrificed for — these questions their entire careers. > *"It fell apart because Ananthropic wanted out of the deal."* ## [13:37] Exporting American Dynamism to Our Allies Ulevitch raises a geographic expansion question: American Dynamism's name is parochial, but the practice is really about America and its allies. Horowitz has spent significant time abroad meeting foreign leaders who want to replicate U.S. startup culture. He outlines why that's hard — entrepreneurship at scale requires a deep-seated belief that the government won't arbitrarily seize what you build, and very few countries (Sweden and Israel being notable exceptions) have that culture. He identifies concrete partnership opportunities: Mexico's high-quality manufacturing expertise in automotive and adjacent sectors; Japan's robotics heritage and surging defense spending (moving from 0% to 3% of GDP), which creates aligned interests given shared concern about China. The section closes with Ulevitch flagging that the coming robotics revolution will be the next major theme for the practice. > *"America does give everybody a chance and entrepreneurs can really count on that."* ## [16:56] Power, Responsibility & How a16z Serves Founders A recent profile described a16z as a "power broker" using capital and networks to shape markets. Horowitz reframes the description: power isn't something the firm accumulates for its own sake — it's a feature of the product offered to founders. Entrepreneurs have great ideas but lack the power to get the right meeting with Congress, secure a key enterprise customer, or navigate regulation; a16z's scale converts that gap into founder advantage. The internal culture is deliberately countervailing. The firm's first cultural principle — "first-class business, only in a first-class way" — means showing up on time, responding promptly, and being honest. These small behaviors prevent the firm from drifting into a posture where it treats founders as supplicants rather than partners. > *"So power is sort of a feature of our offering is the way I think about it."* ## [18:58] The State of Venture Capital & Why Most Firms Can't Scale Horowitz provides a structural explanation for why most venture firms cannot grow beyond a certain size. The original design premise of the industry was that only ~15 companies per year would ever reach $100 million in revenue, so small partnership structures with shared economics and shared control made sense. Mark Andreessen's "software is eating the world" thesis invalidated that premise: every company is now a technology company, so the target universe has exploded and so has the need for organizational scale. Scaling to capture that universe requires organizational reorganization — which requires a single decision-maker. Firms built on consensus control cannot reorg cleanly, because those who lose power in a reorg will block it. A16z, with centralized control from inception, was structured to reorg repeatedly and now fields 600+ people organized as small teams sharing a common platform. The result is a barbell: large generalist firms that cover every technology domain, and narrow specialists focused on AI infrastructure, bio, crypto, or games. The mid-size generalist firm is being squeezed out. > *"when you redistribute power, people are mad if they get a vote um that they're going to foul that that that reorganization and you can't scale without reorging."* ## [23:21] The New Rules of Media The media discussion opens with a structural observation: old and new media are not different games — they are the same game with different rules. Under scarcity (limited channels, rigid formats), the winning strategy was defense: avoid gaffes, because a Howard Dean scream lives forever on a three-channel media landscape. Under abundance (unlimited channels, unlimited formats), the winning strategy is offense: be interesting, because anything boring simply drowns in the noise. Horowitz points to Alex Karp as the exemplar of the new model: relentlessly entertaining, consistently on message (pro-America), and unafraid to be unpredictable. The flood-the-zone correction mechanism — do ten podcasts after a mistake — makes individual errors survivable in a way they never were in the old world. His coaching to founders: you cannot win by not losing anymore; you win by being worth paying attention to. > *"Um, and so the key to winning isn't not making a mistake, it's being interesting."* ## [26:22] America's AI Optimism Gap Horowitz names his biggest worry: a polling result showing that more than 70% of Chinese citizens are optimistic about AI while fewer than 30% of Americans share that sentiment. He attributes the gap to an American media culture that foregrounds AI risks — surveillance, job displacement, existential threats — while systematically underweighting the positive case. He contrasts this with Japan, where renewed enthusiasm for AI has reignited the entire startup ecosystem. His ask of founders, policymakers, and technologists in the audience: rebalance the narrative. AI will end traffic deaths, cure cancer, and eliminate poverty as we know it. These outcomes deserve as much airtime as the dangers. He closes with the analogy of fire — a technology capable of burning down a village that nonetheless heats homes and cooks food — arguing that managing dual-use risk is the normal condition of every transformative technology, not a disqualifying exception for AI. > *"We're going to cure cancer."* ## Entities - **Ben Horowitz** (Person): Co-founder and general partner at a16z; primary speaker throughout, drawing on experience as a founder, CEO, and venture capitalist. - **David Ulevitch** (Person): General partner at a16z leading the American Dynamism practice; hosts the conversation at the American Dynamism Summit in Washington, D.C. - **Andy Grove** (Person): Former CEO of Intel; Horowitz's mentor whose maxim on industry leadership frames the episode's opening section. - **Alex Karp** (Person): CEO of Palantir; cited as a model for direct, entertaining, on-message communication in the new media landscape. - **Mark Andreessen** (Person): Co-founder of a16z; author of "software is eating the world," the thesis underpinning a16z's scaling rationale. - **American Dynamism** (Concept): a16z's investment practice focused on companies serving U.S. national interests — defense, manufacturing, advanced software and hardware — now extended to allied nations. - **Anthropic** (Organization): AI safety company whose contract with the U.S. Department of War collapsed; Horowitz argues the deal fell apart because Anthropic chose to exit, not over genuine ethical conflicts. - **a16z** (Organization): Andreessen Horowitz; raised over $15 billion in its latest fund, the largest in firm history and the largest VC fund ever raised. - **Department of War** (Organization): U.S. federal defense department; counterparty in the Anthropic procurement deal and key customer for American Dynamism portfolio companies. - **Palantir** (Organization): Defense and analytics software company; referenced as an exemplar of a firm successfully working at the intersection of Silicon Valley and national security.

Dan Shipper interviews Angela Jiang (head of product) and Katelyn Lesse (head of engineering) for the Claude platform at Anthropic, recorded at the Code with Claude developer event. The conversation unpacks how Claude's platform has grown from a simple completion API into a fully managed agent infrastructure, why the harness and the model are increasingly inseparable, and what the "outcome + budget" vision means for the future of agent development. Together the three trace every stage of the agent lifecycle — from spinning up a first session to retiring stale agents — and share candid war stories from Anthropic's own internal deployments. ## [00:00] Where the platform will be in a year Dan opens with a question the rest of the episode keeps circling back to: a year from now, where is the platform? Angela's answer — Claude understands itself well enough to pick its own sub-agents and write its own harness on the fly. Katelyn picks up the other half: an infrastructure layer that can keep up with agents that continually rewrite themselves. This exchange actually comes from late in the interview; the show puts it up front because the whole conversation is about how today's primitives get you there. > *"We'd want to experiment with directions where Claude actually gets so good at understanding itself, it figures out what model you should be using, it figures out how to spin up all the sub agents."* — Angela Jiang ## [01:48] How the Claude platform evolved from API to agents Angela traces the arc from early LLM APIs — stateless, exploratory, maximum surface area — through session-based chat, and now into fully autonomous agents. The through-line is always the same: raise the abstraction layer high enough that customers can get the best outcome from Claude with as little work as possible. Early adopters wanted every raw knob; today, most teams arriving at Anthropic want a substantial set of things "out of the box." The platform's job is to keep shrinking the distance between intention and outcome. > *"It probably ends up just being like whatever it's like the set of primitives and infrastructure that enables you to basically get the outcome as fast as possible with actually as little of work as possible."* — Angela Jiang ## [04:09] The primitives that make up Claude Managed Agents Katelyn explains that Claude Managed Agents is assembled from the same primitives available to anyone on the Messages API — code execution sandboxes, web search, and built-in tools — but wrapped in a curated harness Anthropic has already battle-tested internally. Angela adds that the team is opinionated about two primitives in particular: file systems and skills. These are treated as load-bearing choices that shape how Claude behaves across all agent tasks. The platform is designed to be modular so developers can plug in custom pieces where the standard harness does not fit, and Anthropic publishes reference implementations for teams that want to stay on the Messages API directly. Dan describes his team running Claude via the `claude -p` command on Mac Minis and worries about lock-in and divergence from Claude Code. Katelyn responds that Anthropic's internal first-party products run on the same platform as external customers, which means divergence between Managed Agents and Claude Code will shrink over time. > *"We've taken what we see as all the most powerful of those things and put them together into a harness and a set of infrastructure that is just the way to get what we think is the best outcomes out of Claude."* — Katelyn Lesse ## [10:37] Why the harness and the model are becoming a single unit Angela challenges the conventional wisdom that a generic, model-swappable harness is the right architecture. As models diverge in technique across labs, the alpha is in tight harness-model co-design rather than hot-swapping. Internally, Anthropic tested multiple harness variants for the memory feature and found they performed "drastically differently." The implication: treat the agent (harness + model) as the unit of redundancy, not the model alone. Dan pushes on whether this creates path dependence in the model itself. Angela acknowledges that the primitives chosen really do shape the model's trajectory, and that being wrong about them is hard to undo. She cites models that over-indexed on reasoning versus those that went deep on computer-use as two diverging paths that are difficult to reverse. > *"The harness and the model get very paired. You still need redundancy, and you still might want to use other models for things, but you probably do it at the layer of like the agent, meaning like the harness plus the model."* — Angela Jiang ## [18:49] The infrastructure wall that kills most agent projects in production Katelyn identifies the real blocker for most agent projects: not harness engineering, but the infrastructure wall hit when teams try to move from prototype to production. Keeping a persistent server alive, managing sandbox failures, storing transcript data, and handling secure credential injection — these mundane concerns kill projects that technically "work" on a Mac Mini. Anthropic's own repeated experience of hitting this wall internally was the primary motivation for building Managed Agents. Angela describes the vaults primitive as an early step toward one-click agent deployment: once agent identity and credentials are handled securely at the platform layer, adding a Slack integration should eventually be as simple as telling Claude to "add Slack" and watching the bot appear. > *"Everyone hits the same problem of like, oh wow, I either need to like keep a server constantly running or I need to use infrastructure that will spin up and spin down, and I need to store the transcript data, and I need secure sandboxing, and all these sorts of things."* — Katelyn Lesse ## [24:49] Why team agents need a different shape than individual productivity tools Angela explains why individual productivity tools like Claude Code do not simply scale to team use. The moment three people want a shared agent that automates an end-to-end process across roles, a laptop-resident tool breaks down in availability, access control, and coordination. She cites Guillermo Rauch of Vercel's framing of an internal "AI software factory" as the right mental model: not individual augmentation, but a full organizational stack of agents that continuously produces high-leverage output for every function in the company. > *"When you get to the team layer suddenly everything gets like massively more complex. Like number one obviously it can't like sit on your laptop."* — Angela Jiang ## [26:36] How Anthropic's legal team uses an agent to review marketing copy Katelyn walks through one of Anthropic's own internal deployments: a legal-review agent that accepts marketing copy submissions and performs a first-pass review before anything reaches a human lawyer. The agent can approve copy outright or escalate for human review, eliminating low-value ticket-queue work. The form factor is a thin app layer on top of Managed Agents with shared visibility across both teams. Angela and Dan dig into why this is an agent rather than a skill: human-in-the-loop requirements, the need to spin up separate sessions, and multi-team collaboration all exceed what a single skill invocation can handle. The governance model that emerged was notable: rather than gating changes behind the platform team, end users discovered they could self-serve small improvements via Claude Code. Angela describes the end-state user experience as simply "talking to Claude," even when the underlying system is "many many Claudes engaging with each other." > *"Under the hood it's many many Claudes engaging with each other to get to the part where then they the Claudes themselves are doing the more complex work that the human doesn't really necessarily need to interpret."* — Angela Jiang ## [34:24] Using multi-agent orchestration for advisor strategies, adversarial pairs, and swarms Angela highlights three multi-agent architecture patterns people are assembling with the newly launched orchestration primitives: an advisor strategy that separates execution from advice; adversarial pairs where one agent generates and another critiques; and swarms that split a problem into many small parallel pieces and recombine results. Each pattern suits a different problem class — swarms excel at bug hunting, while wide-research tasks benefit from advisor or parallel-decomposition architectures. LEGO-like primitives let practitioners hill-climb at the architecture level, not just the prompt level. > *"If we can make the primitives very LEGO-like, then people can put them together to solve things at a slightly higher form factor, which is more like an architecture or like a strategy."* — Angela Jiang ## [35:50] How to measure agent success with outcome and budget as the end state Angela frames the long-term measurement philosophy: compress everything to an outcome and a budget, and let the platform resolve all intermediate decisions. Domain-specific evals (e.g., PR-merge rate for coding agents) remain useful today, but the target is a verifiable outcome spec that Claude can grade itself against repeatedly. Katelyn addresses the adjacent problem of agent staleness: Anthropic has built skills to help teams upgrade agents when new models ship, and the most forward-leaning teams already run meta-agents that monitor other agents for degradation and trigger upgrades automatically. > *"Our kind of principle of like maybe the end state of some of these things is that everything should kind of compress down to an outcome and like a budget. And that's probably like about it."* — Angela Jiang ## [39:11] What the platform looks like a year from now, when Claude writes its own harness Angela envisions a world where users supply only an outcome and a budget, and Claude self-selects models, spins up sub-agents, and writes its own harness on the fly — eliminating harness engineering entirely, just as today's platform has already eliminated much of manual tool construction and prompt engineering. She is cautiously optimistic that the "outcome" half of the equation may be achievable within a year with some budget error bars. Katelyn adds the infrastructure corollary: such a world requires a platform capable of supporting agents that continuously recreate themselves, handling arbitrarily shaped long-running requests without ever becoming the bottleneck. > *"Claude is actually able to understand itself enough that it can come almost like write itself on the fly to figure out what is necessary in that kind of like two-parameter world of like outcome and budget."* — Angela Jiang ## Entities - **Angela Jiang** (Person): Head of Product for the Claude platform at Anthropic; co-architect of the Managed Agents product vision. - **Katelyn Lesse** (Person): Head of Engineering for the Claude platform at Anthropic; focuses on infrastructure reliability and scale. - **Dan Shipper** (Person): Host of AI & I on Every; CEO of Every; building internal agent products on the Claude platform. - **Claude Managed Agents** (Software): Anthropic's hosted agent infrastructure — a harness plus cloud compute that wraps the Messages API with built-in memory, sandboxing, vaults, and skills. - **Messages API** (Software): Anthropic's core API; the underlying primitive on which Managed Agents and all first-party products are built. - **Anthropic** (Organization): AI safety company that builds and operates the Claude model family and its associated platform. - **Every** (Organization): Media company producing AI & I; an early Managed Agents customer building internal editorial agents. - **Stripe Minions** (Software): Stripe's internal end-to-end software development platform built on agent infrastructure; cited as a model for company-wide coding agent deployment. - **Vercel** (Organization): Developer infrastructure company; CEO Guillermo Rauch's "AI software factory" framing used as the mental model for team-level agent adoption. - **Outcome + Budget** (Concept): Anthropic's long-term design principle that the final form of agent interaction should require only a verifiable outcome and a cost ceiling, with the platform resolving all intermediate decisions.

In one of their most consequential episodes, the All-In besties dissect SpaceX's surprise compute lease to Anthropic — the deal that may cement Anthropic as AI's dominant platform — and debate whether David Sacks's "Rockefeller" framing is prophecy or paranoia. The group then wrestles with a White House trial balloon about an "FDA for AI," ultimately concluding it was mostly media spin, before closing with a bullish-but-cautious read on the AI-driven market boom. Brad Gerstner fills in for David Friedberg, bringing investor perspective from both public and private markets across the episode's 82 minutes. ## [00:00] Bestie intros! Thoughts on the LA mayor election Jason Calacanis opens with the full crew: Chamath Palihapitiya, David Sacks, and fifth bestie Brad Gerstner joining in for David Friedberg, who is out sick. The warm-up quickly turns to the LA mayoral race, where Spencer Pratt is mounting a surprisingly effective challenge to incumbent Karen Bass. The group praises Pratt's viral debate performance — evisceration of the city council candidate over homeless policy — and Chamath notes the power of a sharp social-media team in modern politics. Brad flags a California ballot initiative that would constitutionally protect retirement savings and ban a wealth tax, reading it as a potential seismic signal. Jason observes that New York City hedge-fund titan Ken Griffin publicly announced he is pulling investment from New York after NYC councilman Zohran Mamdani targeted his home in a campaign video, underlining the tension between aggressive progressive politics and capital flight. > *"If California effectively passes a constitutional amendment protecting retirement savings and personal assets and banning the wealth tax and [Spencer Pratt] gets elected, the message that would send to the country — that's a very non-consensus view that I'm becoming increasingly optimistic about."* — Brad Gerstner ## [04:38] SpaceX-Anthropic deal, Elon Web Services, SpaceX IPO valuation, Anthropic's insane growth trajectory Jason leads with the blockbuster news: SpaceX has leased all of Colossus 1 — its H100-based Memphis data center — to Anthropic, adding over 220,000 Nvidia GPUs and 300 megawatts to Anthropic's supply-constrained capacity. The deal immediately doubled Claude Code's rate limits and removed peak-usage caps for paid users. Chamath frames Anthropic's explosive growth as purely supply-constrained: if unlimited power existed, revenues would be "even more parabolic." He sees the deal as Elon strategically de-risking SpaceX's valuation story — blunting bear cases around delayed orbital data centers while generating near-term revenue to subsidize Grok training. Brad estimates the arrangement adds $4–5 billion in incremental 2026 revenue for SpaceX, calling EWS (Elon Web Services) a genuine fourth hyperscaler alongside AWS, Azure, and GCP. He also warns that organized activists — not organic local opposition — are using the same playbook that stalled nuclear construction in America to delay data-center permitting. David Sacks notes that Anthropic grew from $10B ARR on January 1 to $44B ARR by April — a trajectory he calls unlike anything Silicon Valley has ever witnessed. > *"Nobody in Silicon Valley has ever seen anything like it. Forget about the rest of the country. I mean, all we do in Silicon Valley is deal with exponentials. And still, people have never seen that kind of growth at that level of scale."* — David Sacks ## [26:48] Is Anthropic the next great monopoly? Early signals or major overreaction? David Sacks draws an extended analogy between Anthropic and John D. Rockefeller's Standard Oil, arguing that safety-first rhetoric can function as regulatory capture — building a moat that locks in the emerging duopoly of Anthropic and OpenAI while blocking competitors. He notes that if Anthropic sustains its 10× annual growth for just 18 more months it could become "the most powerful monopoly ever created in human history," dwarfing the combined Mag-7 revenue. Brad pushes back hard: Anthropic and OpenAI are still fledgling startups on a GAAP basis, Google and Amazon are producing hundreds of billions in free cash flow to fund competing models, and pre-emptive antitrust action at the starting line of AI would be "a disaster." Jason translates Brad's position as "don't mess with my paper," since Altimeter holds positions in several of these companies. Sacks clarifies his northstar is vigorous competition — but he flags Anthropic's banning of OpenClaw from using its API as a concrete anti-competitive act worth scrutiny. > *"Unless something about their current trajectory changes, Anthropic will be the most powerful monopoly ever created in human history — a trillion dollars of ARR growing at some rate. Dario calls it AGI. I call it the biggest monopoly in human history."* — David Sacks ## [35:21] "FDA for AI" freakout, how the White House thinks about AI safety Reports surfaced that the White House was considering an executive order to create an AI working group that could require pre-release safety reviews for new frontier models — triggered, according to the New York Times, by Anthropic's classified "Mythos" model reportedly alarming national-security officials. NEC Director Kevin Hassett appeared on Fox Business drawing an FDA analogy, while Treasury Secretary Scott Bessent spoke more carefully about balancing innovation and safety. Sacks calls much of it "fake news" amplified by Andrew Ross Sorkin's DealBook column, noting that Susie Wiles, the White House Chief of Staff, issued a statement walking back the FDA framing. He reveals he spoke with Hassett directly and confirms no senior official actually supports a pre-approval regime. He points to the White House's March 20 National AI Regulatory Framework as evidence the administration favors specific solutions over broad regulatory capture. The group converges on one concrete measure: KYC (Know Your Customer) requirements before frontier model API access during preview periods, plus rapid deployment of cyber-capable AI to companies like CrowdStrike and Palo Alto Networks. > *"There is a substantial faction of AI ideologues or doomers who are basically employing the classic 'never let a crisis go to waste' strategy. Yes, we do have this cyber issue that is real — everyone needs to harden their systems now. But what they're trying to do is use that issue to try and create a permanent new infrastructure in Washington."* — David Sacks ## [52:01] Flipping AI's negative perception: Giving, healthcare and education innovation Jason shifts from regulatory defense to offense: how should the tech industry proactively counter negative public perception of AI? He proposes that companies going public — Anthropic, OpenAI, SpaceX — could dedicate 1–5% of IPO proceeds to every American via "Invest America" accounts, creating tangible shared upside. He also calls for serious engagement on minimum wage and universal healthcare, arguing that a financially healthier consumer base is structurally good for capitalism itself. Brad endorses the "Invest America" concept, adding that data center host communities should receive direct benefits like free local electricity. David pivots to political salience data: AI ranks 29th out of 39 voter issues — well below cost of living and economic growth, two metrics where AI is actively deflationary and expansionary. The industry's real message should be economic delivery, not safety governance. Chamath gives tech leaders a "D-minus trending to F" for communications and calls for tangible reinvestment in America at scale. > *"I think that there's a pretty profound vibe shift with respect to tech, tech oligarchs, Silicon Valley, and particularly AI. That vibe shift has already happened on Main Street, and I think that's starting to seep into Washington."* — Chamath Palihapitiya ## [60:04] Trading the AI market, state of the economy Brad leads a comprehensive market check: AWS on a $150B run rate (28% growth), Azure at $108B (39%), Google Cloud at $80B (63%). The S&P 500 is at all-time highs, the 10-year sits at 4.3%, and inflation is under control — far better outcomes than the doom scenarios predicted around tariffs and geopolitical conflicts. S&P 500 operating margins improved from 11% in 2023 to 13% in Q1 2026, and the Mag-5's combined headcount grew only 3% over three years while revenues surged. Chamath urges caution: there is still no direct evidence AI is lifting enterprise profit margins in aggregate, and a reckoning arrives in roughly 500 days when the fork between opex reduction and revenue growth will determine whether the AI boom is real or a mirage. Jason counters that for startups the ROI is already "fait accompli" — AI-generated ad creative at Nike and DoorDash, portfolio companies shipping product at half the headcount. David credits Trump administration policies — rescinding Biden's chip-export licensing and AI-approval regime, unleashing energy permits — for creating the conditions that enabled the boom, and notes that the unemployment rate for recent college graduates has actually improved, contradicting the entry-level-job-loss narrative. > *"I think we have kind of call it 500 days where you just got to be net long. But I think it's literally in the hundreds of days from now that you're going to have to have an important reckoning moment. The people that are paying for all these tokens need to see an actual benefit."* — Chamath Palihapitiya ## Entities - **Jason Calacanis** (Person): Host and moderator; angel investor and podcast co-founder - **Chamath Palihapitiya** (Person): General partner, Social Capital; co-host; contrarian macro voice on AI ROI and market cycles - **David Sacks** (Person): Co-host; former White House AI & Crypto Czar; framed Anthropic as a potential historic monopoly using the Rockefeller analogy - **Brad Gerstner** (Person): Founder & CEO, Altimeter Capital; fifth bestie; bullish on compute stocks and AI market structure - **Dario Amodei** (Person): CEO of Anthropic; referenced as "Daario D. Rockefeller" by Sacks; party to the SpaceX compute deal - **Elon Musk** (Person): CEO of SpaceX and xAI; architect of Elon Web Services and the Colossus 1 compute lease strategy - **Anthropic** (Organization): AI lab behind Claude; grew from $10B to $44B ARR in four months; center of monopoly and FDA debates - **SpaceX / xAI** (Organization): Lessor of Colossus 1 data center to Anthropic; emerging fourth hyperscaler under EWS branding - **Elon Web Services (EWS)** (Concept): SpaceX's compute-leasing business positioned as a hyperscaler competitor to AWS, Azure, and GCP - **Mythos** (Software): Anthropic's classified cyber-capable frontier model that reportedly alarmed White House national-security officials - **KYC for AI** (Concept): Proposal to require identity verification before granting API access to frontier models during preview periods - **Invest America** (Concept): Proposal for IPO-stage tech companies to dedicate a share of proceeds to universal investment accounts for US citizens

A short Anthropic walkthrough of Claude Code hooks — the deterministic escape hatch for things that absolutely must happen on every edit, every tool call, every commit. The pitch: if you find yourself writing "always run prettier" into claude.md and hoping, you've already lost; move it to a hook. ## [00:02] What hooks are and why they're deterministic Hooks fire at fixed points in Claude Code's lifecycle, and the narrator's whole framing is that unlike prompt-level instructions, they always run. Telling the model in claude.md to run prettier after every edit works most of the time — but "most of the time" is exactly the gap a hook closes. Same intent, but enforced by the runtime instead of suggested to the LLM. > *You can tell Claude in your claude.md file to run prettier after every file edit and most of the time it will do that, but sometimes it won't. It's not perfect. But a hook makes it happen every single time with no exceptions.* ## [00:37] Common use cases Four representative examples set the scope: auto-format after file edits, log every executed command for compliance, block dangerous operations such as touching production files, and ping yourself when Claude finishes a long task. > *Common use cases could include auto formatting after file edits, logging all executed commands for compliance, blocking dangerous operations like modifying production files, and sending yourself notifications when Claude finishes a task.* ## [00:52] Configuring hooks and the five lifecycle events Configuration lives in `settings.json`: pick an event, optionally narrow it with a matcher for which tool it applies to, then provide a shell command. Five events cover the loop — `UserPromptSubmit` before Claude even sees a prompt, `PreToolUse` and `PostToolUse` wrapping each tool call, `Notification` when Claude pings the user, and `Stop` when Claude finishes responding. > *Pre-tool use which runs before a tool call, post-tool use runs after a tool call completes. Notification runs when Claude sends a notification, and stop runs when Claude finishes responding.* ## [01:22] Auto-formatting with a post-tool-use hook The canonical example: a `PostToolUse` hook with a matcher of `Edit` or `MultiEdit` fires whenever Claude mutates a file. The command checks the extension and routes to the right formatter — prettier for TypeScript, gofmt for Go, ruff for Python, whatever the project standardizes on. > *You set a post-tool use hook with a matcher of edit or multi-edit, right? So, it fires whenever Claude modifies a file. The command checks the file extension and runs the appropriate formatter.* ## [01:49] Blocking tool calls with pre-tool-use and exit codes `PreToolUse` hooks receive the tool name and input as JSON on stdin and decide via exit code: `0` proceeds, `2` blocks. When a hook blocks, whatever it wrote to stderr gets fed back to Claude as feedback, so the model knows why and can adjust its plan. This is where you enforce hard rules — block writes to a production config dir, refuse bash commands containing `rm -rf`, block commits to main. The narrator's framing: things your team needs guaranteed, not suggested. > *If it exits with code two, the action is blocked and the STD error message gets fed back to Claude's feedback so Claude knows why it was blocked and can adjust.* ## [02:26] Project-level hooks and team sharing Hooks in `.claude/settings.json` are project-scoped and can be committed to the repo, which means the whole team inherits them automatically on clone. Reference scripts via the `CLAUDE_PROJECT_DIR` env var so commands resolve correctly no matter where Claude's cwd happens to be. The closing rule of thumb: if something needs to happen every time without fail, don't put it in a prompt — put it in a hook. > *If something needs to happen every time without fail, don't put it in a prompt. Put it in a hook.* ## Entities - **Anthropic Tutorial Narrator** (Person): Anthropic's official voice-over for the Claude Code 101 tutorial series. - **Claude Code** (Software): Anthropic's agentic terminal coding tool that hooks plug into at lifecycle events. - **Hooks** (Concept): Deterministic commands that fire at fixed points in the Claude Code loop — the runtime-enforced alternative to prompt-level instructions. - **settings.json** (Configuration): Where hooks are declared; `.claude/settings.json` at the project root is checked into the repo so teams share the same rules. - **PreToolUse / PostToolUse / UserPromptSubmit / Notification / Stop** (Events): The five lifecycle events a hook can attach to. - **CLAUDE_PROJECT_DIR** (Environment variable): Used inside hook commands to reference project-relative scripts regardless of Claude's current working directory.

Matt Pocock joins swyx at AI Engineer Europe to argue that the old software design canon — DDD, deep modules, ubiquitous language — matters more, not less, in the AI coding era. The thesis: code is not just a compile target; a codebase that is easy for humans to change is easy for AI to change. Along the way they cover course-making, why traditional lectures still beat AI-native learning, and TypeScript's quiet takeover of AI engineering. ## [00:04] Opening at AIE Europe and the Cursed Course swyx welcomes Matt to the AI Engineer Europe podcast booth in London. Matt jokes that AIE is "the worst" event he has ever attended (the location is in fact astonishing) before turning to his Claude Code course, which is just wrapping up its two-week cohort. He explains why he runs short cohorts: AI moves so fast that self-paced courses cannot guarantee updates, and the "curse" of releasing into breaking changes — AI SDK v5 dropped on day two of his AI SDK v4 course, and the Claude Code source leaked during this one — is now baked in. The conversation then turns to teaching as a craft. Matt rejects the "pundit" branch of YouTuber identity — he is not trying to predict the future, only to teach durable material — and notes that being a teacher first is what differentiates his content. > *I'm not a guy who's trying to predict the future. I'm just trying to teach.* ## [02:51] Why Engineering Fundamentals Matter More with AI Matt previews his AIE talk. The popular narrative says code no longer matters because English plus an AI compiler can produce applications. Every time he tried to ignore the code, he ended up with "a terrible mess." So he went back to the classics — *Extreme Programming*, *The Pragmatic Programmer*, *A Philosophy of Software Design*, DDD — and discovered they ported directly into prompts. Keeping the architecture in your head, even when you delegate implementation, yields outsized dividends. > *If you have a code base that's easy to change for humans, it's going to be easy for AI to change, too.* ## [04:23] Narrow Waist and Deep Modules swyx introduces the "narrow waist" concept from internet architecture (TCP/IP, HTTP at layers 3–4) as a way to contain AI-generated slop: define rigid interfaces, delegate the inside. He extends it to running AIE as a nine-person business — "model-view-claw" instead of MVC, where coordination across people and AI is the real systems problem. Matt maps this onto John Ousterhout's notion of *deep modules*: a large amount of functionality behind a simple interface, ports and adapters style. This is, in his experience, the best way to use AI for coding — be intentional about the interface as a human, then delegate the implementation. > *Deep modules basically — a large amount of functionality with a simple interface. Kind of ports and adapters, right?* ## [06:37] Domain-Driven Design Meets AI DDD is having a moment, and Matt argues it works *because* the framework has been around long enough to sit in the latent space of these models. You do not have to invent new vocabulary; you can bolt on a system that is composable and that the model already understands. The deeper point: DDD is fundamentally about aligning code with language, which is exactly what you want when speaking to an AI. He makes it concrete with the `mattpocock/skills` repo (≈13k stars) and its "ubiquitous language" skill — a Claude Code skill that scans your codebase, surfaces the arcane jargon, and refines it with you into a markdown file he keeps open while prompting. He references it from `agents.md` but does not paste it wholesale, so the agent finds it when searching for those terms. > *Essentially, you're trying to create a unified domain model so that the AI and you are speaking the same language.* ## [10:05] Teaching as an Overpowered Skill swyx asks how Matt got so good at explaining things. Matt credits six years as a voice coach before becoming a developer — communication felt like an unfair advantage when he started as a junior. He has since narrowed his focus: split time between learning material and finding the right phrases for it. The old texts help because they give him pre-built mental models to explain new ideas through. He walks through his course-making process: an "explore and exploit" phase, a Zettelkasten-style Obsidian vault, a custom planning app, P1/P2/P3 prioritization, and the rule that *each lesson teaches exactly one thing* with dependencies made explicit. Most of what he produces ends up on the cutting room floor. > *The ability to communicate always just felt like a ridiculous overpowered skill that I had in my locker that no one else had.* ## [13:20] How People Actually Learn AI Engineering The conversation turns to whether AI has changed how people learn. Matt distinguishes knowledge (lectures), skills (interactive exercises), and wisdom (small-group discussion — and now, talking to an AI). Counterintuitively, the more he leans into AI-experimental teaching, the more it turns his audience off. Most learners still want traditional lectures; swyx recalls Maven's cohort-based education arc landing in the same place. Matt's compromise is to force the work without forcing the form: in his TypeScript material he throws learners into a problem first and gives them the knowledge afterwards. > *The more I lean into the kind of AI experimental stuff, the more it actually turns people off my materials.* ## [15:04] TypeScript Overtaking Python swyx flags that TypeScript overtook Python in the GitHub survey this year — a shift he did not see coming, particularly in AI engineering where Python's expressiveness has been dominant on the backend. Matt's echo chamber is 100% TypeScript, but his real argument is ecosystem: when you care about UX and shipping chat-style applications, the framework gravity is in TypeScript (Vercel's Next.js, Cloudflare's variants). swyx admits this would meaningfully change which frameworks he promotes. > *If you're concerned about UX, concerned about shipping great stuff, you're mostly doing it in TypeScript.* ## [16:45] Inversion of Control and Composable Skills Matt looks ahead. His TypeScript-evals bet (Everlight) stalled — "no one's excited to do evals." The next frontier is *inversion of control*: as coding agents converge on similar architectures (Firebase-style backends, small tool sets), the interesting axis becomes how much control sits with the developer versus the harness. Claude Code's opacity buys ease of use but loses observability; Pydantic AI ("Pi") swings the other way — total control, total maintenance burden. He closes by pointing past coding agents entirely. Software engineers are a step ahead because AI produces quality output in their domain, but the composable skills he authors — like his three-sentence "grill me" skill that makes the AI interrogate you until you reach a shared understanding — generalize to any domain where you want the AI aligned with you. > *The inversion of control is going to be really important — you put more control in the hands of the developer and less in the harness.* ## Entities - **Matt Pocock** (Person): Creator of Total TypeScript and AI Hero; teaches TypeScript and AI Engineering through two-week cohort courses. - **Shawn Wang / swyx** (Person): Host; founder of AI Engineer and the AIE conference series. - **AI Engineer Europe (AIE)** (Organization): The London conference where this conversation was recorded; Matt's talk hit 1M views in 13 days — fastest in AIE history. - **AI Hero** (Organization): Matt's AI engineering education platform (aihero.dev). - **Claude Code** (Software): Anthropic's coding agent; subject of Matt's just-finished course and a recurring example throughout. - **Domain-Driven Design (DDD)** (Concept): Software methodology centered on aligning code with the language of the business domain; Matt argues it ports cleanly into AI prompting. - **Ubiquitous Language** (Concept): DDD practice of maintaining a shared vocabulary doc; Matt's namesake Claude Code skill scans a repo and refines this with the user. - **Deep Modules / Narrow Waist** (Concept): Architectural pattern (Ousterhout / internet protocols) of large functionality behind a small interface — Matt's preferred shape for AI-assisted codebases. - **mattpocock/skills** (Software): Matt's open-source repository of Claude Code skills; ≈13k stars at recording time. - **Pydantic AI (Pi)** (Software): Python agent framework built from low-level primitives; cited as the high-control counterpoint to Claude Code's opaque harness. - **Obsidian** (Software): Note-taking app reportedly run by a team of four; the example for non-engineering domains where AI leverage compounds.

Dan Shipper and Austin Tedesco, Every's head of growth, discuss why the Codex desktop app has become their primary interface for all knowledge work — from drafting go-to-market plans to building live KPI dashboards — displacing Claude Code after months of side-by-side use. Dan frames the shift as the emergence of a new "agent management interface" operating system, while Austin walks through his live Codex setup in a screen-share session that covers automations, specialized agent suites, and recruiting workflows. The episode doubles as a practical field guide for non-engineers who want to run the same playbook. ## [00:00] A new operating system for knowledge work Dan opens cold: three months ago Codex was trash. Now Austin is the one firing it up before anything else each morning and routing 80 percent of his working time through it. Dan reads what changed structurally: a general-purpose coding agent that can reach into your filesystem, browser, and connected apps is becoming the operating system for knowledge work, and every major lab is racing for that surface. > *"There's a new operating system for how and where you're going to get your work done and it's this kind of agent management interface."* — Dan Shipper ## [00:57] How Codex went from a tool for senior engineers to a daily driver for knowledge work Dan traces the arc of Codex from its original positioning as a sandboxed pair-programming tool for senior engineers — one that "would argue with you, it would make you feel stupid" — to today's desktop app built on GPT-5.5. He attributes the pivot to OpenAI watching Anthropic prove with Claude Code that an emotionally intelligent, fast, computer-native agent creates a step-change experience for programmers and knowledge workers alike. The race is now between model companies to own the agent management desktop: Anthropic has Claude Code and Claude.ai desktop, OpenAI has Codex, and xAI has effectively acquired Cursor. ## [02:42] How Claude Code proved that a great coding agent works for any knowledge work Dan explains the insight that changed everything: if an agent can write software autonomously, it can do any kind of knowledge work autonomously. Claude Code demonstrated this first, drawing non-engineers — including Austin — into an agent-first workflow. OpenAI's hard pivot on Codex over the last three months is a direct response to that proof point. Dan describes the new paradigm as one where your agent is your interface to software, the internet, and daily tasks, not just a code co-pilot. > *"If it can write software on its own, it can do any kind of knowledge work on its own."* — Dan Shipper ## [07:24] Austin's switch to Codex Austin recounts his agent-pill moment: spending a December week inside Claude Code CLI, hooking it up to every tool he uses for work and personal life, and finding it indispensable for strategic thinking, data analysis, and drafting marketing copy. His initial Codex trial two months later felt alienating — the model was condescending, asking "Why?" when he requested clearer explanations. He kept Claude Code for 80 percent of knowledge work while tolerating Codex for engineering. The turning point was getting early access to GPT-5.5: at model parity, the decisive edge was the Codex desktop app itself — faster, better-organized, and with sub-agents that "just work." > *"So the idea that the codeex app is maybe 30 to 40% better is like that's a lot of work."* — Austin Tedesco ## [13:48] How Austin set up Codex with folders, keys, and reviewer agents Austin shares his screen and walks through his "Every Growth OS" folder inside the Codex app: a directory containing API keys for every tool the company uses (Gmail, Slack, Notion, Stripe), a CLAUDE.md project context file synced to GitHub, and a set of custom reviewer agents forked from Kieran Classen's Compound Engineering plugin. Where the standard Compound Engineering reviewers focus on security and front-end design, Austin's fork — publicly available as "Compound Knowledge" — reviews for strategic alignment with company goals and data accuracy, making it fit for knowledge-work plans rather than code PRs. The folder architecture lets Austin move seamlessly from a go-to-market draft to shipping a code PR without switching apps. > *"It's connected to everything we use for every and then some project instructional files that explain what the every business is, what we care about, how we like to work together."* — Austin Tedesco ## [18:24] Using Codex to brainstorm automations across Gmail, Slack, and Notion Austin demos his recommended on-ramp for new Codex users: open a fresh chat inside the Growth OS folder, run the Compound Engineering brainstorm workflow, and prompt the model to look at Gmail, Slack, and Notion and suggest automations. Codex surfaces a "follow-up radar" that triages incoming communications across sources, a command-center view for events and camps, and a recruiting pipeline automation — all calibrated to Austin's actual work context. Within the session, Codex writes automation scripts that require almost no tweaking and begins scheduling them; Austin highlights a nightly draft-reply routine that compiles unanswered messages and prepares replies for a quick thumbs-up approval. > *"They require very little tweaking to be like this is a thing I would and do use every day of there's this set of instructions that it comes up with based on what it knows about me."* — Austin Tedesco ## [22:42] How Austin manages the human review step when Codex is drafting communications A live audience question from Margaret prompts Austin to describe his human-in-the-loop review discipline. All drafting and orchestration happens inside Codex, but the final review intentionally lives in the native app: Slack draft replies are reviewed in Slack's drafts tab; email drafts are reviewed in Gmail; strategic plans are reviewed in Notion or the Proof markdown viewer. Stepping out of the agentic interface "freshens up my brain" before anything goes to a human. A second question from musician Alex about protecting high-value client emails leads to a discussion of how Austin uses Every's Kora email assistant together with Codex-managed rules, including having the agent interview the user to derive email rules rather than asking the user to specify them manually. > *"I just like for like the last pass before humans engage with it to step away from this agentic space and have a final check in another surface."* — Austin Tedesco ## [28:54] Using Codex to build specialized agents inspired by product executive Claire Vo Austin describes being inspired by a Claire Vo interview with Lenny Rachitsky in which Vo credited a suite of six specialized OpenClaw agents — rather than one overloaded master agent — as the key to unlocking leverage. Austin pasted the transcript of that interview directly into Codex and prompted it to propose six agents tuned to the Every growth function, provisioned into the company Slack. The agents occasionally break, but debugging is straightforward: screenshot the broken output or @-mention the Slack thread inside Codex and ask it to fix the agent's architecture. The result is a self-correcting loop where agent failures become Codex tasks. > *"Um I I actually just sent it the transcript of Claire's interview with Lenny and said like I want to do this too given everything you know about me and my work."* — Austin Tedesco ## [31:09] Synthesizing meeting transcripts and Slack threads into a go-to-market plan Austin walks through his most time-saving workflow: assembling a go-to-market plan for Every's upcoming Plus One product launch using nothing but Codex running the Compound Engineering brainstorm step against all existing meeting transcripts stored in Notion and Slack threads. With only five-minute windows between meetings, Austin prompted Codex to check the scheduled content calendar (a step it skips unless reminded), generate a proof doc, and push the final plan to Notion. The result was 80–90 percent complete. Dan adds the normative point: he prefers reading AI-written documents because they're easier for colleagues to produce, and the standard at Every is that you stand fully behind whatever your agent writes. > *"It's that I'm relying on the model to um look at all of the things that we've already said and thought about the go to market strategy, piece it together, and then review it, right?"* — Austin Tedesco ## [40:15] Building a live KPI tracker in Notion that agents can read Austin shares a more technical workflow: rebuilding Every's KPI tracker as a Notion database that updates every six hours by pulling from Stripe, social platforms, and other data sources via Notion's Workers tool. The tracker is explicitly designed to be both human-readable and agent-readable, so any team member's agent can query it and take autonomous actions — such as spinning up landing pages if an SEO keyword is underperforming. The challenge: the model can't one-shot the full tracker because even a 3–5 percent error in the MRR number is unacceptable for business decisions, so Austin is validating it column by column. Dan notes the philosophical complexity of defining revenue metrics consistently. > *"And so I have been doing this big kind of like to me complex uh workflow problem in codeex of let's build this sheet together, let's have it live in a notion database that all of our agents can point at."* — Austin Tedesco ## [44:54] Using Codex for recruiting Dan describes using Codex for outbound recruiting: he asked Codex to compile a list of General Assembly alumni and then filter it for people who had subsequently moved into AI, targeting candidates for an L&D director role. The first name on the resulting list was someone Dan considered a perfect fit who already followed him on Twitter, allowing an immediate DM. The section expands into a broader Q&A: Austin discusses when to fork Compound Engineering versus using it out of the box, how the team uses a shared Notion "compound" database to capture session learnings and turn them into reusable skills, and how Every's "Think Week" — a bi-annual week with no day-to-day work — creates organizational space for deep AI exploration. > *"Especially for any kind of like outbound effort, it can kind of find that needle in the haststack that you're looking for really really well."* — Dan Shipper ## Entities - **Dan Shipper** (Person): Co-founder and CEO of Every; host of the AI & I podcast; author of essays on AI and vibe coding - **Austin Tedesco** (Person): Head of growth at Every; Codex power user who manages the Growth OS project and suite of specialized agents - **Claire Vo** (Person): Product executive whose interview about specialized agent suites inspired Austin's multi-agent setup at Every - **Kieran Classen** (Person): Engineer at Every; creator of the Compound Engineering plugin used as the basis for Austin's knowledge-work fork - **Codex** (Software): OpenAI's desktop agent app, the primary tool discussed; runs on GPT-5.5 and supports sub-agents, folder-scoped projects, and plugin integrations - **Claude Code** (Software): Anthropic's CLI-based coding agent; Austin's previous daily driver before switching to Codex - **Compound Engineering** (Software): Plugin workflow framework by Kieran Classen; provides structured brainstorm, plan, and review steps used across Claude Code and Codex - **Every** (Organization): AI-focused media and software company publishing essays, courses, and tools; runs the AI & I podcast - **OpenAI** (Organization): Creator of Codex and GPT-5.5; provider of the ChatGPT Pro subscription whose credits were offered to camp attendees - **Notion** (Software): Primary knowledge-management and document platform at Every; used for meeting transcripts, the KPI tracker, and agent-readable databases - **GPT-5.5** (Software): OpenAI model powering the current Codex desktop app; reached parity with Claude Opus for Austin's knowledge-work tasks

Lex Fridman sits down with Jean-Baptiste Kempf, president of VideoLAN and lead developer of VLC, and Kieran Kunhya, longtime FFmpeg contributor and the voice behind the infamous FFmpeg account on X, for a four-hour deep dive into the invisible machinery behind virtually all video on the internet. Together they trace the full arc from raw bytes and container formats through hand-written assembly and codec reverse-engineering, confronting the open-source sustainability crisis along the way. The conversation is both a technical masterclass and a meditation on why brilliant volunteers—many of them teenagers—quietly build infrastructure that powers billions of devices every day. ## [00:00] Episode highlight The episode opens with a rapid-fire highlight reel that captures the spirit of what follows. Kempf distills the FFmpeg community's core value: code quality is the only credential that matters—"Maybe you're a dog. I don't care. I need to look at your code." Kunhya adds the scale: FFmpeg is running on roughly 100 million CPUs at any moment, with three billion devices continuously decoding video, and FFmpeg's x86 assembly hand-optimization runs 62 times faster than equivalent C. The segment also previews the CIA-VLC spy story, the intelligence-agency backdoor request Kempf flatly refused, and Kieran's "no regrets" Twitter philosophy. > *"We care about excellent code. We don't care who you are. Like maybe you're a dog. I don't care, right? I need to look at your code."* — Jean-Baptiste Kempf ## [02:17] Introduction Lex sets the scene: FFmpeg is the invisible backbone behind YouTube, Netflix, Chrome, VLC, Discord, and nearly every platform that touches video or audio. VLC has been downloaded more than 6.5 billion times. Both projects are built entirely by volunteers. Lex frames the episode not merely as a technical discussion but as a tribute to engineers who work for the craft rather than for fame or money—"one of the great examples of human beings quietly collaborating across borders to build something useful, durable, and elegant." > *"It is one of the most incredible software systems ever developed, and it's all done by volunteers."* — Lex Fridman ## [05:35] Weirdest things VLC opens The conversation lightens up with playful examples of VLC's legendary tolerance for exotic formats. Kempf describes users capturing VHS tapes via capture cards, support for DVD-Audio with custom encryption, and the Lucasfilm Star Wars game codec that FFmpeg implemented for a single 10-second opening sequence. At a VideoLAN conference, a competition to create the most broken file ever—an MKV where every frame changed resolution, aspect ratio, and rotation—ended with VLC playing it perfectly. The orange traffic-cone logo is discussed: so recognizable that 25% of VLC's website traffic arrives from people searching "cone player." > *"There was a file that's a valid ZIP and a valid MP3 at the same time or something like that—and VLC opened all of the stupid files."* — Kieran Kunhya ## [09:59] How video playback works Kempf and Kunhya walk through what happens the moment you press play: the player fetches a byte stream from a URL, the demuxer separates audio, video, and subtitle tracks, entropy decoding removes mathematical compression, intra prediction reconstructs still-image frames (I-frames), motion-compensation handles temporal redundancy (P- and B-frames), and the final raw pixels are handed to the GPU or audio card. Video compression achieves 100x to 200x reduction by exploiting how human eyes perceive luminance versus color—working in YUV space rather than RGB—and by reusing unchanged background regions across frames. Kunhya warns that every single sentence in this pipeline represents someone's lifetime of work. > *"Everything we've just said in the past couple of minutes, every sentence is someone's lifetime's work. There are books about every sentence."* — Kieran Kunhya ## [19:20] Video codecs and containers The hosts clarify the often-confused distinction between containers and codecs. A container (MP4, MKV, MOV) multiplexes audio, video, and subtitle tracks; the codec (H.264, AV1) compresses the content inside. VLC and FFmpeg deliberately ignore the file extension and probe the actual bytes—because in the real world, extensions lie. The segment covers how AVI was Microsoft's format, MOV became MP4 via Apple, and the Matroska/MKV format emerged from the open-source community. Modern codecs like AV1 are not single algorithms but collections of tools that adapt to content type—screen share, animation, live video—each requiring different coding strategies. > *"We discard the file format. We look into the file to understand what's in it because so many people say, 'Oh, it's a video, it must be MP4,' but technically it's an MOV or maybe it's a MKV."* — Jean-Baptiste Kempf ## [30:07] FFmpeg explained FFmpeg is described as a low-level library suite—libavcodec, libavformat, libavfilter—plus a command-line tool so expressive that Kempf calls it a full programming language. Every person watching a YouTube video, recording with OBS, or editing in a professional broadcast box is likely touching FFmpeg. Kunhya notes that trillion-dollar corporations and grandmothers with home videos operate on exactly the same technology stack. The segment dives into open-source licensing—MIT, GPL, LGPL, AGPL—as "social contracts" that define community norms. Kempf recounts the painstaking process of re-licensing VLC's core from GPL to LGPL, requiring him to track down more than 350 contributors, including visiting the factory-worker father of a deceased contributor to obtain permission for two lines of code. > *"From a philosophical level, it's incredible that your grandmother's home videos and trillion-dollar corporations are on a level playing field using the same technology stack."* — Kieran Kunhya ## [51:07] Linus Torvalds Kempf offers a nuanced defense of Linus Torvalds's legendary harshness. The Linux kernel's core community is tiny—as is FFmpeg's (10–15 active maintainers)—and those few people must maintain every line of code forever. "We cannot compromise on quality because the core community of FFmpeg is ten to fifteen, and we are the ones who are going to maintain your code." Kunhya adds that terseness is often simply fatigue: volunteers arrive home after a full day of work and review patches without the bandwidth to hand-hold. Kempf also points out that most community members are non-native English speakers, and cultural misreadings amplify perceived hostility. > *"We cannot compromise on quality because the core community of FFmpeg is ten to fifteen, and we are the ones who are going to maintain your code."* — Jean-Baptiste Kempf ## [55:46] Turning down millions to keep VLC ad-free Kempf traces VLC's unlikely origin: a French engineering school (École Centrale Paris) whose student-run campus built a satellite video-streaming system in 1995—a decade before YouTube—just to enable faster networks for video games. From that Network 2000 project grew VideoLAN, and VLC emerged as its client. Kempf joined in 2003 when the project had nearly died, grew it from hundreds of thousands to billions of installs, and along the way repeatedly refused "obscene" offers to bundle toolbars, change search engines, or insert advertisements. His reasoning: "I need to go to bed at night and be happy about what I've done. If I had sold out, I would have betrayed so many other people who work here." > *"I refuse dozens of millions of dollars, yes, several times. Yes, I could be a multimillionaire and be somewhere on the beach. But I did not do it because I thought it was not moral and it was not the right thing to do."* — Jean-Baptiste Kempf ## [70:04] FFmpeg & Google drama Kunhya recounts a public controversy in which Google's security team used AI to auto-generate bug reports for FFmpeg, filing them under tight 90-day deadlines—with some vulnerability reports going to the press before patches could be written—without contributing corresponding fixes or meaningful funding. Kunhya compares it to "a denial of service by AI-generated bug reports" on obscure 1990s game codecs. The saga escalated via spicy FFmpeg tweets (a "rap battle" in Kunhya's words), but produced concrete results: Google began sending patches and established a financial reward system for fixes. A parallel incident saw Microsoft Teams engineers file a high-priority bug on the volunteer tracker, name-dropping their product's scale, and offering a one-time payment of a few thousand dollars in response to a request for a long-term support contract. > *"Google uses FFmpeg at a scale probably you or I couldn't even contemplate—millions of CPU cores. And yes, they contribute in areas mostly regarding their own products. But in a wider sense, there's a disproportionate level of contribution."* — Kieran Kunhya ## [89:18] FFmpeg developers What motivates FFmpeg's volunteer engineers? Kempf identifies three drivers: passion for the subject matter (many contributors arrived because they loved anime), excellence of the craft ("this is the best school ever of programming"), and pride in impact ("you can tell your grandma: I do this so you can play video on your laptop"). Kunhya adds that Andrew Kelley, creator of the Zig programming language, was an FFmpeg developer who credits his time there as his real-world education. Teenagers have written thousands of lines of hand-optimized assembly for FFmpeg. Kieran's favorite quote, from John Collison: "The world is a museum of passion projects." > *"If you're good in C, if you know how to write assembly in FFmpeg, I assure you you're going to be one of the best programmers ever—even if you're working on writing TypeScript."* — Jean-Baptiste Kempf ## [95:55] VLC and FFmpeg Kunhya frames the FFmpeg-VLC relationship as a "binary star system": VLC is to FFmpeg as Android is to Linux—they depend on each other and succeed because of each other. Roughly 80% of FFmpeg pipelines depend on at least one VideoLAN project (most often x264). VLC gives FFmpeg exposure to a vast zoo of real-world broken files. When compiled for Windows, VLC links against about 16 million lines of code, of which only 1 million live in the VLC repository itself. The two projects share many developers and collectively demonstrate that complex software ecosystems can be built entirely from interdependent open-source components. > *"VLC is to FFmpeg as Android is to Linux. They depend on each other, but they coexist because of each other."* — Kieran Kunhya ## [100:29] History of FFmpeg The "eras tour" of FFmpeg begins with Fabrice Bellard creating the initial concept, followed by the Michael Niedermayer era of the early 2000s—exhaustive support for DivX, Xvid, Windows Media, and RealMedia, eliminating the need for bloated, spyware-ridden codec packs. The late 2000s brought H.264 maturity and the rise of high-definition video. Throughout, VLC served as FFmpeg's field test: millions of users exposing edge cases that no lab could anticipate. > *"At the time you needed a new player to play every different type of file format. Having a single library that was fast and open source—that was a massive achievement."* — Kieran Kunhya ## [103:46] Reverse engineering codecs The segment showcases the art of reverse engineering proprietary codecs. Kostya Shishkov—described as "borderline genius"—reverse-engineered 20–30 megabyte binary blobs (each megabyte representing roughly a month of normal work) for fun, producing decoders for Windows Media, RealMedia, and GoToMeeting formats. Kunhya explains the methodology: hook into the proprietary player to dump raw YUV data, open a disassembler, step through machine code instruction by instruction to infer the entropy coding, prediction, and IDCT stages, then validate bit-exactness against sample files. For months, the work produces no visible output—pure debugging in memory. > *"He looked at the world as a binary specification. He didn't need documentation or anything. He would go away and come back and do interesting stuff."* — Kieran Kunhya ## [117:01] FFmpeg testing FFmpeg's FATE (FFmpeg Automated Testing Environment) system runs a pivot table of test combinations: dozens of compilers (GCC, Clang, MSVC, Apple Clang, Intel Compiler), operating systems (Linux, macOS, Windows, BSD, Solaris), and CPU architectures (x86, ARM, RISC-V, PowerPC). All test machines are volunteer-hosted. The system catches compiler miscompilations—rare but devastating, since even a single wrong bit in a frame dependency chain can cascade into major visual corruption. Kunhya notes that the Macs at the top of the FATE dashboard are hosted in his own office. > *"It's not just a matrix at this point. It's like a pivot table of different combinations—all run by volunteers."* — Kieran Kunhya ## [121:08] Assembly code (handwritten) This extended chapter is the technical heart of the episode. Handwritten x86/ARM SIMD assembly in FFmpeg and x264 runs up to 62 times faster than equivalent C—a gap that modern compilers and auto-vectorization cannot close despite years of trying. VLC still supports Windows XP through Windows 11, macOS 10.7 through macOS 26, iOS 9 through the latest, BSD, Solaris, and even OS/2. Understanding assembly forces programmers to internalize CPU pipeline stages, SIMD registers, L1/L2/L3 cache, and memory bus constraints. Kempf and Kunhya introduce the x86inc framework built by Loren Merritt for x264 and JB's Assembly Lessons tutorial series, which have attracted contributions from teenagers learning directly from the source. > *"I believe it's necessary to understand assembly language, even if you don't do it much, to understand what's going on inside your computer. That will make you a better programmer."* — Jean-Baptiste Kempf ## [145:26] Rust programming language Kempf and Kunhya hold divergent opinions on Rust. Kunhya respects the memory-safety goal but finds the community self-important—"It has a very big Esperanto vibe"—and argues that Rust rewrites reaching only 85–90% of required feature coverage are insufficient; "the last 1% takes 99% of the time." Kempf has written Rust VLC modules and sees genuine value, but notes that the lack of training data for low-level SIMD work means AI tools cannot yet assist meaningfully. The discussion broadens to the two assembly wizards of the community: Henrik Gramner, whose knowledge of Intel x86 cycle counts exceeds Intel's own engineers, and Martin Storsjö, who writes ARM Neon assembly on a virtual keyboard while watching his kids play in the playground. > *"Rust reminds me of the Sinclair C5. In order to get people to move, you have to build something as good as, if not better than, what you have now."* — Kieran Kunhya ## [154:42] FFmpeg and Libav fork In 2011, FFmpeg split into FFmpeg and Libav, primarily over governance and leadership style rather than technical disagreements. Several Linux distributions temporarily shipped Libav instead of FFmpeg. Kempf describes open-source forks as healthy—they force projects to confront structural weaknesses. Eventually most of Libav's developers returned to FFmpeg, and the projects merged back. Kempf draws a parallel to the XZ Utils attack, where a lone maintainer, exhausted by coordinated social engineering, granted commit access to an attacker—highlighting how burnout creates the very single-point-of-failure vulnerabilities that make critical open-source infrastructure fragile. > *"Forks are important because they change the status quo of a community. FFmpeg today is better than it was before the fork."* — Jean-Baptiste Kempf ## [163:04] Open source burnout Kempf and Kunhya confront the mental health crisis among open-source maintainers. Kempf has received physical death threats—including a letter containing powder—over decisions such as dropping PowerPC support. The security community's habit of filing alarming CVEs for hobby-project edge cases adds psychological load without providing patches. Kempf now maintains several libraries whose original maintainers burned out. The conversation broadens to the systemic problem: critical infrastructure like libxml and XZ is maintained by one or two people, unknown to the trillion-dollar enterprises that depend on them. > *"The mental health of the open source maintainers is something that large corporations don't care or don't see."* — Jean-Baptiste Kempf ## [170:51] x264 and internet video H.264 transformed internet video by arriving exactly when Intel Core 2/Nehalem CPUs made real-time software decoding practical. The key innovation of x264 was psychovisual rate-distortion optimization—encoding decisions driven by visual quality metrics rather than mean squared error, producing sharper, more natural-looking video. This was driven by the anime community's high standards for perceived sharpness. AV1 offers 40–60% bandwidth savings over H.264 at the same quality, but encoding costs two orders of magnitude more CPU. YouTube therefore re-encodes only popular videos in AV1, making the extra compute worthwhile by amortizing it over millions of viewers. > *"Thirty percent of the video from Netflix is now in AV1, fifty percent of YouTube."* — Jean-Baptiste Kempf ## [184:07] Video compression basics The chapter clarifies I/P/B frame structure: I-frames are complete still images, P-frames reference only previous frames, and B-frames can reference both past and future frames. ProRes is an intra-only codec designed for nonlinear editing—no temporal dependencies, fast seeking. The segment also covers constant-bitrate versus constant-quality encoding, group-of-pictures length, and the thousands of engineers at Netflix, YouTube, and Meta whose entire job is tuning FFmpeg parameters for specific content types. A historical curiosity: Google Video originally used VLC as an ActiveX plugin inside Internet Explorer; today VLC is compiled to WebAssembly to run inside browser JavaScript engines. > *"You have I-frames that are complete frames, P-frames that depend only on I-frames, and B-frames that can depend on frames in front."* — Jean-Baptiste Kempf ## [191:04] CIA and fake VLC WikiLeaks' Vault 7 release revealed that the CIA built a modified version of VLC with an additional DLL (psapi.dll) that silently encrypted and exfiltrated documents while the victim watched a movie, using the expected high CPU load of video playback as cover. VideoLAN issued a press release directing users to download only from the official website. A separate incident involved Chinese state hackers distributing a fake VLC using legitimate signed VideoLAN DLLs to target Indian users, causing India to ban VLC until Kempf fought a successful legal battle to reverse the ban. The segment also surfaces a hidden feature: VLC can render movies as ASCII art in a terminal, useful for diagnosing multicast network paths via SSH. > *"If we had to compromise our software, we would shut it down. This is clear."* — Jean-Baptiste Kempf ## [201:39] Ultra low latency streaming Kempf explains adaptive streaming (HLS, DASH): the player downloads segments, times the download, and adjusts quality tier accordingly. The real engineering frontier is live broadcasting with strict CBR constraints—satellite uplinks cannot burst even for one second. Kempf describes his company Kyber, an open-source (AGPL dual-licensed) ultra-low-latency streaming stack targeting robotics and XR, streaming compressed video feeds to devices without onboard compute. The segment ends with a discussion of teleop for robots, where latency directly determines safety. > *"Kyber is open source. Everything on Kyber is open source. If you want to use it in your product and not open source it, you pay the commercial license."* — Jean-Baptiste Kempf ## [219:07] AV2 codec and video patents AV2, the successor to AV1 within the Alliance for Open Media (of which VideoLAN is a member), promises a further 30% bandwidth reduction. VideoLAN's dav1d decoder will be followed by "dav2d." The Alliance exists specifically to escape the HEVC/H.265 patent thicket: HEVC's three separate patent pools demanded fees so large that HP removed HEVC support from new laptops, and streaming giants calculated they could build a new royalty-free codec for less than the annual licensing cost. France's rejection of software patents means Kempf has never paid codec licensing fees—if he had to, the bill would exceed 200 euros per user. > *"At a hundred million per year, you know, I could create my own codec—and this is what they did."* — Jean-Baptiste Kempf ## [228:59] VLC backdoors Intelligence agencies from two different countries approached Kempf asking him to insert backdoors into VLC. He declined both, in terms he describes as "a lot less polite" than a simple no. The chapter broadens into a discussion of European entrepreneurship: Kempf argues that French startup culture has transformed over 15 years—failure stigma has fallen, AI companies are proliferating—while acknowledging that over-regulation remains a real drag. He closes by reflecting on his strategy for remaining calm under legal and political pressure: always ask "am I dying? Am I hurting someone?" If not, move on. > *"If we had to compromise our software, we would shut it down. Also because what we do is good and it's done for everyone."* — Jean-Baptiste Kempf ## [239:14] Video archiving Kieran profiles the archiving preservation community, led in part by Dave Rice of CUNY, which relies on FFmpeg as a "Rosetta Stone" for playing future-proof multimedia. The community funded FFV1, FFmpeg's lossless codec, to guarantee that archived footage loses no information—critical because lossy compression could destroy forensic or historical details visible only on close inspection. A famous cautionary tale: the BBC's 1986 New Domesday Book project archived content on BBC Micros, and within 20 years no one had working software to read it. There are now more historical video tapes in archives than functional tape heads in the world to digitize them, forcing painful triage decisions about what human history to preserve. > *"C will be like Latin. It will be a thing you learn from the past, but it will still be usable in certain contexts."* — Kieran Kunhya ## [245:51] Future of FFmpeg and VLC The closing chapter surveys where multimedia is heading: volumetric video, point-cloud codecs for robotics, RGBD depth streams, XR/VR streaming, and—speculatively—neural interfaces that may one day require codecs for compressed brain data. Kempf is confident FFmpeg will exist in 100 years; VLC he rates as "maybe." He closes with his personal philosophy: "Regrets are a tax on your mind. Learn from your mistakes, but don't regret." The episode ends with Lex reading Linus Torvalds: "Most good programmers do programming not because they expect to get paid or get adulation by the public, but because it is fun to program." > *"Regrets are a tax on your mind. Learn from your mistakes, but don't regret. Because you've done it, so unless you have a time machine, don't regret."* — Jean-Baptiste Kempf ## Entities - **Jean-Baptiste Kempf** (Person): President of VideoLAN, primary maintainer of VLC, founder of Kyber and several other companies; declined tens of millions of dollars to keep VLC ad-free. - **Kieran Kunhya** (Person): Veteran FFmpeg contributor, codec engineer, founder of Open Broadcast Systems, the voice behind the FFmpeg account on X. - **Lex Fridman** (Person): Host of the Lex Fridman Podcast, AI researcher, longtime VLC and FFmpeg advocate. - **Fabrice Bellard** (Person): Creator of FFmpeg, QEMU, and tcc; foundational figure of the project. - **Michael Niedermayer** (Person): Long-time FFmpeg maintainer who drove exhaustive codec support through the 2000s. - **Kostya Shishkov** (Person): Legendary FFmpeg reverse engineer who decoded proprietary binary blobs for Windows Media, RealMedia, and GoToMeeting codecs. - **Henrik Gramner** (Person): Assembly wizard with deeper knowledge of Intel x86 cycle counts than Intel's own engineers. - **Linus Torvalds** (Person): Creator of Linux and Git; referenced as a model of uncompromising code quality standards in open-source communities. - **FFmpeg** (Software): Open-source multimedia framework providing codecs, muxers, filters, and command-line tools; the invisible backbone of nearly all internet video. - **VLC** (Software): Open-source media player with 6.5+ billion downloads, built on libVLC and FFmpeg; plays virtually any format on any platform. - **x264** (Software): VideoLAN's open-source H.264 encoder; the dominant software encoder for internet video, famous for psychovisual optimizations. - **dav1d** (Software): VideoLAN's fast open-source AV1 decoder; widely deployed in browsers and streaming clients. - **VideoLAN** (Organization): French nonprofit that stewards VLC, x264, dav1d, and related open-source multimedia libraries. - **Alliance for Open Media** (Organization): Industry consortium including Google, Netflix, Apple, Amazon, and VideoLAN that created AV1 and is developing AV2 as royalty-free codec standards. - **FATE** (Software): FFmpeg Automated Testing Environment; volunteer-hosted CI grid testing hundreds of compiler/OS/architecture combinations. - **Kyber** (Organization): JB Kempf's startup building an ultra-low-latency open-source streaming stack for robotics and XR, dual-licensed AGPL/commercial. - **H.264 / AVC** (Concept): The dominant internet video codec standard; open-source implementation is x264; basis of Blu-ray and most MP4 files. - **AV1 / AV2** (Concept): Royalty-free next-generation video codec standards from the Alliance for Open Media; AV1 saves 40-60% bandwidth vs H.264; AV2 adds another 30%.

Anthropic의 공식 Claude Code 안내서——Claude Code가 무엇인지, Claude.ai와 어떻게 다른지, 그리고 LLM이 코드베이스에서 명령을 실행하기 전에 알아야 할 세 가지를 설명합니다. 터미널 도구를 처음 설치하려는 개발자를 대상으로 합니다. ## [00:04] Claude Code의 정의와 실행 환경 Claude Code는 에이전트형 코딩 도구로 포지셔닝됩니다. 코드베이스를 이해하고, 파일을 편집하며, 명령을 실행하고, 이미 사용 중인 개발자 도구와 통합됩니다. 터미널, VS Code, JetBrains IDE, Claude 데스크톱 앱, 웹 등 여러 환경에서 동작하지만, 이 안내서에서는 터미널을 기본 환경으로 다룹니다. > *Claude Code is an agentic coding tool that understands your code base, edits your files, run commands, and integrates with your existing developer tools to help you get things done faster.* ## [00:34] Claude.ai와의 차이점 핵심 차이는 모델 성능이 아니라 접근 방식에 있습니다. Claude Code는 터미널과 전체 코드베이스에 직접 접근하므로, 채팅창에 복사-붙여넣기하는 반복 작업이 사라지고 도구가 제자리에서 작업을 완료합니다. "AI 에이전트"라는 표현은 이 직접 실행 방식을 함축적으로 표현한 것입니다. > *Unlike Claude AI, Claude Code has direct access to your files in your terminal and your entire code base.* ## [00:51] AI 에이전트와 Claude Code로 할 수 있는 것들 여기서 AI 에이전트란 환경과 상호작용하고 정해진 목표를 달성하기 위해 행동을 취하는 소프트웨어를 의미합니다. 가장 기본적인 형태는 도구, 외부 서비스, 다른 에이전트에 접근할 수 있는 실시간 루프 속의 LLM입니다. Claude Code에서는 이것이 구체적인 기능으로 나타납니다. 코드베이스 읽기 및 설명, 파일 전체에서 버그 추적, 빌드 스크립트 및 테스트 실행, 패키지 설치, 그리고 다음 행동을 결정하기 위한 최신 API 문서 웹 검색 등입니다. > *An AI agent is a software that can interact with its environment and perform actions to complete a defined goal.* ## [01:45] 시작 전에 알아야 할 세 가지 개념 나레이터는 일상적인 사용에 영향을 미치는 세 가지 속성을 강조합니다. 첫째, **컨텍스트 윈도우**는 Claude의 작업 메모리로, 크지만 유한합니다. 그래서 에이전트는 코드베이스를 전부 불러오는 대신 전략적으로 탐색해야 합니다. 둘째, Claude Code는 명령을 실행하거나 파일을 변경하기 전에 **허가를 요청합니다**. 모든 단계를 직접 제어하고 싶든, 대부분 자율적으로 실행하게 하고 싶든 제어권은 항상 사용자에게 있습니다. 셋째, **틀릴 수 있습니다**. 의도를 잘못 파악하거나, 버그를 도입하거나, 수정을 과도하게 설계할 수 있습니다. 출력물은 다른 도구의 결과물과 마찬가지로 다루고, 무조건 신뢰하지 마십시오. > *By default, Claude Code will ask you before running commands or making changes to your code base.* ## [02:34] 요약 Claude Code는 코드베이스를 읽고, 파일을 편집하며, 명령을 실행하고, 외부 도구에 연결하여 더 빠르게 결과물을 만들 수 있도록 돕는 에이전트형 코딩 도구입니다. 현재 터미널, VS Code, JetBrains, Claude 데스크톱 앱에서 사용할 수 있습니다. > *Claude Code is an agentic coding tool. It reads your code base, edits your files, runs commands, and connects to external tools to help you ship faster.* ## 엔티티 - **Anthropic Tutorial Narrator** (Person): Claude Code 101 튜토리얼 시리즈의 Anthropic 공식 내레이터. - **Claude Code** (Software): Anthropic의 에이전트형 터미널 기반 코딩 어시스턴트로, 코드베이스에 직접 작동합니다. - **Claude.ai** (Software): 채팅 기반 Claude 제품으로, Claude Code의 환경 내 실행 방식과 대조됩니다. - **AI agent** (Concept): 정해진 목표를 추구하기 위해 도구, 외부 서비스, 다른 에이전트에 접근하며 실시간 루프에서 실행되는 LLM. - **Context window** (Concept): Claude의 작업 메모리. 유한하기 때문에 에이전트는 전체 코드베이스를 불러오는 대신 전략적으로 탐색합니다. - **VS Code / JetBrains IDEs** (Software): Claude Code가 터미널 및 Claude 데스크톱 앱과 함께 통합되는 에디터.

Alex Lupsasca — 2024 New Horizons Breakthrough Prize winner and OpenAI resident scientist — recounts how GPT-5 resolved a year-long open problem in quantum field theory: proving that single-minus gluon tree amplitudes are non-zero and finding their compact closed form. He then describes how the publicly available GPT Pro, given the gluon paper as a seed, independently generalized the result to graviton amplitudes in under three days of human clock time. Throughout the conversation, Lupsasca reflects on what this trajectory means for how physics is done, how the next generation of physicists will be trained, and where the remaining bottlenecks — verification, creativity, and publishing infrastructure — still lie. ## [00:00] Introduction to AI's impact on physics research Lupsasca opens in medias res, framing the episode's central claim before the formal introduction: AI has crossed a threshold where it can resolve questions that stumped human experts for over a year. He describes this not as a curiosity for theoretical physicists but as a profound, if underappreciated, change in the nature of scientific discovery itself. > *"That's a certain milestone that we've passed, and I think maybe for the average person on the street who doesn't care about theoretical physics, this is not very noticeable, but I think it's a very profound change and we've really passed some kind of a threshold."* ## [00:43] Guest introduction: Alex Luposka The hosts — Brandon (Atomic AI) and RJ Honicky (Miro Omix) — introduce Lupsasca as a Vanderbilt professor and OpenAI fellow who holds both the 2024 New Horizons in Physics Breakthrough Prize (often called the "Oscars for science") and the IUPAP Young Scientist Award. Lupsasca immediately sets the narrative arc: a year ago, AI was useful for email but not for his work; ChatGPT o3 was the first model that genuinely helped with research math; then GPT-5 reproduced one of his hardest published results in 30 minutes. > *"When GPT-5 came out it was able to reproduce one of my best papers that took me a very long time to come up with in like 30 minutes. And that's when I really became AI pilled."* ## [02:49] Alex joining OpenAI and the shift in physics research After GPT-5's release, Lupsasca began evangelizing the shift to colleagues who were skeptical. Finding OpenAI equally excited, and being on sabbatical, he joined as resident scientist — the person physicists around the world now email when something astonishing happens. He describes receiving an inbound that week about Codex simulating the Sachdev-Ye-Kitaev (SYK) model in 10 minutes, a feat that many research groups had struggled to achieve due to the narrow Venn diagram of physicists with strong coding skills. > *"I talked to OpenAI. They were also really excited and I thought I have to get in on this and to understand that this is happening and not be a part of it is a huge mistake so I have to go to OpenAI."* ## [04:08] The release of GPT-5 and the shift in capabilities Lupsasca contrasts the lukewarm Twitter reception of GPT-5 (complaints that it was not better at writing email) with what he observed at the science frontier. He notes GPT-5.4 is another significant jump, and describes how AI capabilities for physics have been accelerating rapidly since o3, the first reasoning model strong enough for research-grade mathematics. He uses this as a bridge to the central technical story of the episode: a pair of new papers on gluon and graviton scattering amplitudes. > *"At the science frontier the capabilities were really taking off."* ## [10:05] Explaining Quantum Field Theory and amplitude calculations Lupsasca gives an accessible primer on quantum field theory (QFT), the framework that reconciles special relativity and quantum mechanics. The key objects in QFT are scattering amplitudes — complex-valued functions that encode the quantum probability for a set of incoming particles (with given energies, momenta, and polarizations) to scatter into a set of outgoing particles. These amplitudes are computed at particle colliders like the LHC, and knowing the n-point amplitude (for any number n of particles) encodes essentially the full content of the theory. > *"If you have a particular force and you're able to compute the n-point amplitudes... you know everything about the theory."* ## [14:20] Overview of gluons and the strong force Gluons are the force-carrying particles of the strong nuclear force — the force that, despite like-charge repulsion between protons, holds the atomic nucleus together. They are the QFT analog of photons for electromagnetism and gravitons for gravity. Like photons, gluons carry a polarization (helicity): positive (right-handed) or negative (left-handed). This helicity structure is central to the paper discussed next. > *"The strong force is mediated by the exchange of the particles of the strong force, which are called gluons, because they're what glues together the nucleus of the atom."* ## [14:38] Discussing the first research paper on single-minus gluon tree amplitudes Lupsasca unpacks the paper's title — "Single-Minus Gluon Tree Amplitudes Are Non-Zero" — piece by piece. Tree amplitudes are the leading-order (no-loop) contributions to scattering. All-plus-helicity amplitudes are exactly zero by a symmetry argument. Single-minus amplitudes — where all but one gluon have positive helicity — were assumed in textbooks to also be zero by the same argument. The paper proves they are not. The result involves collaboration with Alfredo Guevara (IAS), David Skinner (Cambridge), Andrew Strominger (Harvard), and Kevin Wheel. > *"If you look at the lecture notes and textbooks that have been written on this, the same argument that rules out the all-plus amplitudes also appears to rule out the single-minus amplitudes."* ## [20:56] How ChatGPT helped solve a year-long physics puzzle Strominger, Guevara, and Skinner had understood for about a year that the textbook argument has a loophole: when particles are collinear (exactly aligned in momentum), the standard dimensional-analysis reasoning fails, and single-minus amplitudes can be non-zero. But computing what those non-zero amplitudes equal had eluded them. Lupsasca invited Strominger to visit OpenAI and work on it with AI. The week before Strominger's flight, Lupsasca began using ChatGPT Pro. By the time Strominger landed, they had the answer. > *"Using ChatGPT we solved the problem before he even got off the plane."* ## [23:02] Complexity of manual calculations in physics Lupsasca shows the audience a concrete illustration of the difficulty: the six-point single-minus amplitude, worked out by hand by Alfredo Guevara, is a sum of 32 terms each of which is itself a product of four complicated factors. The number of terms grows factorially with the number of particles n — super-exponential growth. This is the messy representation that the group had been staring at for a year, seeking the analog of the elegant Parke-Taylor formula. > *"By the time you get to six terms, it explodes in your face."* ## [26:12] The history and mechanics of Feynman diagrams Feynman diagrams are a visual language introduced by Richard Feynman to organize perturbative QFT calculations: diagrams represent possible intermediate histories of a scattering process, and the full amplitude is a sum over all of them. Diagrams are organized by number of vertices (interaction points); each additional vertex is suppressed by the coupling constant, so tree diagrams (fewest vertices) dominate. Loop diagrams — where intermediate particles are created and annihilated — contribute smaller corrections. The combinatorial explosion of tree diagrams is the root cause of factorial growth. > *"In principle, there are infinitely many pictures to sum over."* ## [27:44] The Parke-Taylor formula and the quest for simplification In the 1980s, Parke and Taylor computed the "maximally helicity violating" (MHV, or double-minus) gluon amplitudes through a heroic Feynman diagram expansion. Despite the factorial number of terms, everything canceled to leave a single compact formula — the Parke-Taylor formula — that fits in half a line. Strominger, Guevara, and Skinner spent a year looking for the analogous compact formula for the single-minus case. Their search stalled at the level of the messy Feynman representation. > *"Andy, Alfredo and David spent the last year chasing the analog of the Parke-Taylor formula, the very simple answer that was obtained in the '80s for the double minus amplitudes."* ## [31:26] Using ChatGPT to find the simplification in the special phase space region When the five-point single-minus amplitude was fed to ChatGPT Pro, the model identified a special subregion of phase space (where one particle's frequency has opposite sign) in which the amplitude simplifies from eight terms to a product of just three. This appears not to have been a known fact; the model wrote Python code and tested thousands of possibilities to deduce it. Moving to the six-point amplitude (Guevara's hand calculation), ChatGPT simplified 32 terms to a product of 4. It then conjectured the general n-point formula — with only linear growth in the number of terms, the best possible behavior. GPT-5.2 Pro guessed the formula but could not prove it. > *"The formula that it proposed, instead of having this factorial growth... here it's actually linear. So if you double the number of particles, you only double the number of terms."* ## [38:07] Proving the formula from scratch to ensure validity To obtain a proof, Lupsasca used an internal OpenAI model with extended reasoning. He gave it the problem cold — without the conjectured formula — and asked it to find the general answer in the special phase-space region. After 12 hours of computation, the model independently rediscovered the same formula and produced a complete three-step proof. The proof constitutes the bulk of the published paper. The team kept the AI attribution to one paragraph, framing the paper as a physics result that stands on its own merits. > *"We gave it the whole problem from scratch... and it came back with the same formula which we had not given it. So it rediscovered the correct formula. But this time it also found the proof."* ## [41:00] Determining the scientific impact and future research Asked to compare the result to the Parke-Taylor formula, Lupsasca is candid that scientific impact is only assessable decades later, but argues the result is genuinely surprising and should open a line of attack toward deeper questions in quantum gravity. The conversation pivots naturally to the second paper. > *"I think the true value of a paper can only be assessed decades into the future based on how much future work it leads to and what developments it opens up."* ## [42:27] Introduction to the second paper on graviton amplitudes Gravitons are the hypothetical quanta of gravity — the spin-2 force carrier analogous to the spin-1 photon (electromagnetism) and gluon (strong force). Unlike gluons, gravitons have never been directly detected, but they are central to quantum gravity theory. The second paper, "Single-Minus Graviton Tree Amplitudes Are Non-Zero," shows the same loophole applies to gravity and that a compact formula extends there too — despite gravitons being mathematically more complex than gluons. > *"We wrote this paper which is called single minus graviton tree amplitudes are non-zero. So it's the same title almost, except with graviton instead of gluon."* ## [45:41] Defining particles, irreducible representations, and symmetry Lupsasca sketches the modern QFT definition of a particle (an irreducible representation of the Poincaré group, classified by Wigner according to mass, spin, and charge) and explains why gravitons are spin-2 while gluons and photons are spin-1, making graviton polarization data twice as rich. Crucially, the second paper was complete within three days of the first going public — most elapsed time was spent verifying correctness, not computing. > *"Most of the time was spent verifying the answer, not writing, which is insane, actually, if you take a step back."* ## [47:46] How GPT Pro generalized the research to gravity For the graviton paper, no internal model was needed — the publicly available ChatGPT GPT-5.2 Pro sufficed. Lupsasca provided the gluon paper as context plus two paragraphs describing the key mathematical changes, then said "Good luck. You're a brilliant theoretical physicist." Over a 110-page exchange, the model worked through the graviton calculation — applying the directed matrix tree theorem, a piece of known combinatorics that neither Lupsasca nor collaborators had thought to invoke — produced correct intermediate results, and wrote a draft paper very close to the final arXiv version from section 3 onward. > *"It's a real solid result in quantum gravity that was done pretty much completely by an AI with human steering it and asking kind of the right questions."* ## [53:57] The epistemological shift: Is this a new way of doing physics? The hosts raise the central epistemological question: if an undergraduate with domain knowledge and good prompting could have done this, what does graduate training mean now? Lupsasca agrees this is the hardest open question facing academia. He notes that arduous calculation trains not just skill but self-confidence, that the gap between coursework and the research frontier is growing, and that many "easy" problems professors once assigned to students are now solvable by AI in minutes. He offers two concrete ways AI has already changed his own workflow: dramatically reducing time spent confused between steps, and enabling parallel AI scouts that explore multiple research directions simultaneously. > *"With AI, actually, you can launch 10 instances of chat and have each one try a different route and send it as a scout that moves very fast into the unknown."* ## [59:27] The use of AI as a 'scout' for research directions Lupsasca elaborates on the scout metaphor: rather than carefully mapping a route from A to C before committing to it, a researcher can now dispatch many AI "scouts" in parallel, get rapid feedback on which directions are promising, and redirect human attention accordingly. Even when a scout makes errors, its signposts reduce orientation cost for the human following. This constitutes a qualitatively new mode of research — one where the bottleneck shifts from calculation to judgment about which directions matter. > *"Even if ChatGPT doesn't always get everything right, just kind of having a scout that signposts some key steps along the way that you can use to anchor your own movement is extremely helpful."* ## [61:44] The role of 'taste' and collaboration with AI The hosts push on the problem of "taste" — the ability to identify which questions are at the productive edge of knowledge. Lupsasca argues that working effectively with ChatGPT requires the same skill a professor develops advising students: knowing what question to give, at what level of detail. "Taste" — knowing where the frontier is and which questions there are tractable — is the last skill to develop and the one AI currently lacks. AI is, he says, like an extremely technically skilled graduate student: given a sharp, well-posed question, it can do incredibly hard computations correctly, but it does not yet know which question to ask. > *"The difference between a good physicist and a great physicist is knowing what is the right question to ask — that is actually the hardest part of being a scientist."* ## [70:23] Personal evolution from AI skeptic to resident scientist Lupsasca recapitulates his personal arc: skeptic → converted by o3 (which solved in 11 minutes a calculation that would have taken him days) → "AI-pilled" by GPT-5 (which reproduced, in 30 minutes, his best published result on black hole Love numbers and tidal symmetries — a paper whose training cutoff predated its arXiv release) → now resident scientist at OpenAI. He notes that no competing model at the time could match GPT Pro on that calculation. > *"In under 30 minutes, with one hint... it completely solved this problem, which is one of the nicest calculations that I've ever done."* ## [72:46] Solving a black hole perturbation problem with GPT-5 Lupsasca details the "Move 37" moment that converted him: his paper "Why Is There No Love in Black Holes?" establishes new symmetry generators for perturbations of a Kerr black hole (explaining why black hole Love numbers — tidal response coefficients, named after mathematician Augustus Love — are exactly zero). When GPT-5 Pro was first given the full problem cold, it failed. But after being primed with the simpler flat-space warm-up (a 200-year-old known result), it then solved the full Kerr black hole problem in 18 minutes. > *"GPT-5 was able to reproduce one of my hardest calculations, which I think the number of people in the world that could do that you could count on your hands."* ## [76:34] Discussing whether AI can make original, conceptual leaps The hosts ask whether AI is doing genuine recombination versus true creative leaps. Lupsasca cites Terry Tao, who has not yet seen an AI proof that cannot be traced to an obscure reference. But Lupsasca has been impressed and frames the distinction as one of degree rather than kind — humans may also be recombination machines. He believes continued scaling will produce feats of insight that look like creativity, and notes OpenAI is actively working on enabling models to take bigger, more out-of-distribution leaps suited to scientific discovery. > *"I'm not sure there's a qualitative difference. I think it's just a matter of degree — as we continue scaling the capabilities, I don't see why it's going to stop."* ## [80:09] Challenges of 'AI slop' and the future of academic publishing With models now capable of turning out a physics paper in 30 minutes when properly steered, the arXiv preprint server is being flooded with submissions. Lupsasca distinguishes legitimate use (expert steering + careful verification) from "AI slop" — poorly prompted outputs submitted without adequate checking. His proposed response: raise the bar rather than increase volume. The single-minus amplitude papers open a clear line of attack toward genuine quantum gravity questions; the goal should be to pursue harder problems, not to publish incrementally. > *"Instead, I think now that we have this new tool that gives us AI superpowers, I think we should just raise the bar for what it means to write a good paper."* ## [83:13] The bottleneck of writing academic papers Asked what single bottleneck he would remove, Lupsasca nominates the paper-writing process itself — finding it increasingly strange that researchers use AI to do calculations, compress results into a static paper, and then readers feed that paper back into AI to understand it. He envisions interactive, LLM-embedded papers as a plausible future. He also identifies two missing capabilities in current models: (1) the spark of creativity to identify the next important question, and (2) reliable self-verification, so that the onus of checking long AI-generated proofs does not fall entirely on humans. > *"Maybe some kind of interactive paper which lives in some LLM. Maybe your whole paper is some ChatGPT page... I think we're going to head in that direction."* ## [90:19] Final takeaways and looking ahead to the next year Lupsasca's closing message: pay attention. The trajectory from "useful for email" to "solves open problems in quantum gravity" has taken roughly 18 months. Models are solving open problems that expert communities spent years on. Extrapolating forward, with more scaling already in the pipeline, the next 6 to 12 months should bring further surprises. The right posture is excitement, careful verification, and a commitment to pursuing harder problems. > *"If you just extrapolate that into the future, imagine where we're going to be in 6 months or a year — I think it's kind of surreal to live through this time, but it's really happening."* ## Entities - **Alex Lupsasca** (Person): Theoretical physicist, Vanderbilt University professor and OpenAI resident scientist; 2024 New Horizons Breakthrough Prize and IUPAP Young Scientist Award winner; expert in black hole physics and scattering amplitudes. - **Andrew Strominger** (Person): Harvard professor and Lupsasca's former PhD advisor; pioneer of celestial holography; co-author of both single-minus amplitude papers. - **Alfredo Guevara** (Person): Postdoctoral researcher at the Institute for Advanced Study (IAS); performed the foundational hand calculations underpinning the AI-assisted breakthrough. - **David Skinner** (Person): Professor at Cambridge University; co-author of the single-minus gluon amplitude paper. - **Terry Tao** (Person): Fields Medal-winning mathematician at UCLA; referenced regarding the question of whether AI proofs involve genuine creativity. - **Scattering Amplitudes** (Concept): Complex-valued functions in quantum field theory encoding probabilities for particles to scatter; the central mathematical objects of both papers discussed. - **Single-Minus Gluon/Graviton Amplitudes** (Concept): Tree-level scattering amplitudes where all but one particle have positive helicity; previously assumed zero in textbooks but shown non-zero in a collinear phase-space region. - **Parke-Taylor Formula** (Concept): Compact closed-form result for maximally helicity violating (MHV, double-minus) gluon amplitudes derived in the 1980s; the template whose analog was sought for single-minus amplitudes. - **Feynman Diagrams** (Concept): Diagrammatic technique to organize perturbative QFT calculations; individual diagrams represent distinct intermediate-particle histories whose amplitudes are summed. - **Love Numbers** (Concept): Coefficients encoding tidal deformability; famously vanish for black holes, a fact connected to hidden symmetries studied in Lupsasca's "Why Is There No Love in Black Holes?" paper. - **Celestial Holography** (Concept): Research program exploring symmetries of quantum gravity via scattering amplitude structure; motivates studying graviton amplitudes. - **OpenAI** (Organization): AI research company where Lupsasca serves as resident scientist; developer of GPT-5 and the internal extended-reasoning model used for the amplitude proof. - **arXiv** (Organization): Open-access physics and mathematics preprint server; mentioned in the context of AI-generated "slop" flooding submissions. - **GPT-5 / ChatGPT Pro** (Software): OpenAI's frontier language model used as the primary AI tool in both amplitude papers; capable of extended reasoning steps of 20-34 minutes per prompt.

Boris Cherny, the creator of Claude Code, sits down with Sequoia's Lauren Reeder at AI Ascent 2026 and makes a blunt claim: for the code he writes, coding is already solved. He hasn't typed a line by hand in 2026, runs dozens of agent "loops" at once, and ships most of his work from his phone. The throughline is a bet that writing code is becoming so cheap that the interesting questions move up a level — to what teams look like, what happens to software products, and whether the printing press is the right way to think about what's coming. ## [00:00] Introduction A Sequoia emcee opens the AI Ascent session by asking the room for a show of hands: who uses Claude Code, and who has "Claude Code psychosis." She introduces Boris Cherny as the creator of the tool and hands the interview to Sequoia's Lauren Reeder. > *"We know that the entirety of software development kind of rests on your shoulders."* ## [00:55] Claude Code Crowd Check Reeder frames the conversation for a room full of builders and fills in Boris's background: a career writing code, a TypeScript textbook, an engineer's engineer. The detail that lands hardest is recent — as of early 2026 he hasn't written a single line of code by hand, a sharp reversal for someone who built his reputation on craft. > *"Last time we chatted you hadn't written a single line of code in the last year, or at least so far in 2026, which is quite the change."* ## [02:39] Origin Story of Claude Code Boris explains that Claude Code started almost by accident inside Anthropic Labs, a small incubator he joined in late 2024 that also produced MCP and the desktop app. The team built what it wanted, disbanded, and has since reunited under Mike Krieger for a second round. The motivation was a sense of "product overhang" — capability sitting unused because no product had caught up to it yet. > *"The reason that I started to work on coding is we felt like there was this product overhang."* ## [03:35] From Typeahead to Agents In late 2024 the state of the art was typeahead — press tab, complete one line — which Sonnet 3.5 had just made viable. Boris bet the model was nearly ready to skip that step and write all the code as an agent. It didn't work for the first six months; even after release, Claude Code wasn't a hit. The exponential growth only arrived with Opus 4. > *"I built it, and it just really didn't work for the first 6 months. It was barely usable."* ## [05:07] Is Coding Solved Reeder presses on Boris's on-the-record claim that coding is solved. He polls the room — hand-written code versus fully agent-written — and lands the audience around "50% solved," then says for him it's 100%. He points out the Claude Code codebase itself is unglamorous TypeScript and React, chosen deliberately because that stack is heavily represented in the model's training distribution. > *"For me it's just solved."* ## [06:50] Boris Personal Workflow Boris walks through a setup he first shared on Twitter six months ago and didn't expect to surprise anyone. It has since changed: most of his work now happens from his phone, through the Claude app's code tab, where he keeps five to ten sessions each running a few hundred agents. The tool he reaches for most is the loop — fire-and-forget agents that grind on a task and report back. > *"I sort of feel like loops are the future at this point."* ## [08:51] Future Teams and Generalists Asked what teams will look like, Boris predicts a shift toward generalists. Today a generalist still means an engineer who spans iOS, web, and server; tomorrow it means people who are cross-disciplinary, blending engineering with product and design rather than staying in a single lane. He notes the Claude Code team already skews this way. > *"There's going to be a lot more generalists... generalists that are cross-disciplinary."* ## [10:26] SaaS Apocalypse Predictions Reeder asks the question Boris calls his favorite: if AI makes writing code 10 to 100x cheaper, does the value of software products collapse — a SaaS apocalypse? Boris argues the two things that will actually happen aren't the ones people keep predicting, and uses his guest spot on the Acquired podcast as a detour into why he thinks the conventional framing misses the point. > *"I think there's two things that are going to happen and I don't think either of them is the thing that people have been talking about."* ## [12:57] Audience Q&A Deep Dive The floor opens to the room. An audience member, Dan, asks how much of Claude Code's success Boris attributes to the model versus product decisions — Boris says a mix, roughly 50/50, and won't forecast two years out because the team plans a week at a time. The richest answer is his printing-press analogy: before the press, about 10% of Europe was literate; in the 50 years after, more was published than in the prior thousand, and literacy eventually climbed toward 70%. He uses it to argue that building software is on track to become a near-universal skill. Later questions probe the engineering-versus-world gap, local versus cloud models, and how to parallelize agents with loops, batches, and sub-teams. > *"In the 50 years after the first printing press, there was more literature published in Europe than in the thousand years before."* ## [23:35] Closing and Whats Next For the last question, Boris is asked what kind of product he'd build today that gets more interesting as models improve. He points to Claude Design as a good example — decent now, much better soon — and teases features landing for Claude Code in the coming weeks, plus more work on loops, batch, and massively parallel agents, with computer use in the mix. > *"I think loop and batch and things like this around like massively parallelizing agents, that's going to get better."* ## Entities - **Boris Cherny** (Person): Creator of Claude Code at Anthropic; former Anthropic Labs member, now back on the team under Mike Krieger. - **Lauren Reeder** (Person): Sequoia Capital partner; interviewer for this AI Ascent session. - **Mike Krieger** (Person): Chief Product Officer at Anthropic and Instagram co-founder; leads the reunited Claude Code team. - **Anthropic** (Organization): AI lab behind Claude and Claude Code. - **Claude Code** (Software): Anthropic's agentic coding tool, originated in Anthropic Labs alongside MCP and the desktop app. - **Anthropic Labs** (Organization): Internal incubator where Claude Code, MCP, and the desktop app were built. - **Product overhang** (Concept): Model capability that outpaces the products built on it — the gap Boris set out to close. - **The loop** (Concept): Fire-and-forget agent runs that work a task continuously and report back; Boris's most-used workflow. - **SaaS apocalypse** (Concept): The thesis that cheap AI-written code collapses the value of software products — which Boris pushes back on. - **Printing press analogy** (Concept): Boris's frame for AI coding — literacy went from ~10% to ~70% over centuries; software-building may follow.

NYU Stern professor and serial entrepreneur Scott Galloway delivers a two-hour reality check on artificial intelligence: the doom-and-gloom predictions from AI CEOs are largely fundraising theatre, yet the technology poses a genuinely insidious risk that almost nobody is discussing — an epidemic of loneliness. Galloway argues that AI primarily benefits the already-wealthy, that tech leaders should not be trusted to self-regulate, and that the most valuable human skill in the AI era is not coding or Mandarin — it is the ability to endure rejection. The conversation weaves through geopolitics, investing, the masculinity crisis, and what it means to find purpose, closing with a raw reflection on grief and fatherhood. ## [00:00] Intro Host Stephen introduces Scott Galloway against a backdrop of rapid AI development and unsettling quotes from tech CEOs predicting total job replacement. Galloway opens with his central thesis: the two greatest brand collapses of the past 18 months are the United States' global reputation and artificial intelligence itself — both victims of overpromising and poor trust management. He signals that he is an AI optimist at the macro level, but insists the people building it do not have the public's best interests at heart. > *"These techs, they do not have our best interests at heart."* ## [02:45] What's Actually True About AI Galloway reveals a striking data point: approval of AI is directly correlated with income. Only households earning over $200,000 per year hold a net-positive view of the technology, because they benefit through rising portfolios and are the heaviest users. Everyone else sees higher electricity bills, no equity stake in the companies, and dismissive comments from leaders like Sam Altman telling people to stop complaining about energy costs. The AI brand, he argues, has shifted in 18 months from "scary but optimistic" to "scary and only good for the already rich." > *"Your view of AI is directly correlated to your wealth. The only cohort that has a positive rating of AI is people making over $200,000."* ## [05:00] Are AI CEOs Exaggerating The Future To Raise Billions? Galloway lays out the economic logic behind AI catastrophizing. These companies sit on astronomical valuations that can only be justified if either (a) a trillion dollars in incremental revenue materialises from AI-powered products, or (b) there is a massive wave of labour cost savings. Because option (a) is not yet visible — he sees no AI-driven products at meaningful scale — the CEOs amplify option (b), painting vivid pictures of job destruction to justify the efficiency gains enterprises need to believe in. He calls some of the doom talk "thinly veiled fundraising," noting that founders catastrophize and then take secondaries and leave for Santorini. > *"The catastrophizing is nothing more than a thinly veiled attempt to say my technology is so devastating that it's going to shift society and you should invest at this crazy valuation."* ## [09:00] What Would Prove The AI Skeptics Wrong? Asked where he could be wrong, Galloway is specific: if unemployment rises to 20% even temporarily, history shows civil unrest follows regardless of eventual job recovery. He points to radiologists and coders as cases where AI has augmented rather than eliminated roles — new coder job listings are up 11% year-on-year. His benchmark for being wrong is sustained destruction outpacing creation fast enough that the recovery "V" triggers social breakdown before the other side is reached. > *"At 20% unemployment, especially among youth, especially young men tend to get very angry and take to the streets."* ## [11:05] Could AI Move Too Fast For Society To Handle? The conversation turns to pace of change. Galloway uses the host's own media empire — 220 hires in 24 months — as a live counter-example to the apocalypse narrative. He notes a structural inversion: for the first time in decades, unemployment among non-college graduates is lower than among college graduates because AI data centres are driving a boom in trades. He praises the entrepreneurial wave unlocked by AI tools and flags Denmark's 2% GDP commitment to retraining versus America's inadequate equivalent as the real policy failure. > *"AI is not going to take your job. Someone who understands AI is going to take your job."* ## [16:05] What Happens When AI Combines With Robots? Galloway addresses Elon Musk's Optimus robot predictions and the convergence of physical automation with AI cognition. His 2026 stock pick is Amazon, which already holds more industrialised robots than the rest of the US combined and plans to double its retail operation by 2032 without additional headcount. He is sceptical of domestic humanoid robots but takes seriously the military application of weaponised autonomous systems as a genuinely dark unknown frontier. > *"Amazon is saying they're going to double their largest business, which is their retail business by 2032 without an incremental hire using robotics, industrialised robots."* ## [19:05] Is Elon Musk Selling Vision or Reality? Galloway separates Musk the innovator from Musk the stock promoter. He calls Starlink the best tech product of the past several years and credits Musk with inspiring the EV race. But Tesla should trade at 30x earnings, not 150x, and capital will migrate to SpaceX when it IPOs at a projected 90–110x revenues. The core insight: the modern CEO's job has inverted from underpromise-and-overdeliver to overpromise-and-underdel in order to access cheap capital and pull the future forward. > *"The key attribute of an innovator right now is storytelling — to make sure the promise is way ahead of the performance such you can access cheap capital and pull the future forward."* ## [24:05] Which Jobs Are First To Disappear In The AI Shift? Long-haul trucking is Galloway's clearest near-term casualty: autonomous trucks can run the 10 pm to 4 am window and trucking is the largest single employer of non-high-school-graduate males in America. Legal work at the junior associate level is already being displaced — he now routes contracts through two competing LLMs rather than a $400–$2,000 law firm review, projecting a third reduction in his annual legal spend. The pattern he observes is multiplication: one AI-fluent analyst replaces five, yet the resulting EBITDA funds expansion that creates new jobs elsewhere in the ecosystem. > *"AI is not going to take your job. Someone who understands AI is going to take your job. So have a second screen — always have a second screen open that has nothing but AI on it."* ## [30:05] What Skills Will Actually Matter In The Future? Storytelling tops Galloway's list — the ability to look at data, construct a narrative arc, and communicate it compellingly across every medium. He holds up Jeff Bezos's 1997 shareholder letter, Jensen Huang's stadium keynotes, and Alex Karp's walk-and-talk earnings calls as models. Relationships are the second pillar: as technology converges and products commoditise, the differentiator is whether people want to work with you. He is honest that predicting specific skills is unreliable — private schools doubled down on computer science and Mandarin a decade ago, and neither bet has paid off as expected. > *"The enduring skill is storytelling — your ability to look at data, create a narrative arc and then communicate that story in a compelling way via all the different mediums."* ## [33:45] Are Young People Losing The Ability To Handle Rejection? Galloway identifies the erosion of rejection-tolerance as the most underrated threat facing young people, especially young men. Frictionless online relationships offer a simulacrum of connection without the emotional labour of real-world risk. He mentors young men by assigning deliberate rejection exercises: approach a stranger for friendship, ask someone out for coffee. The goal is not the yes; it is learning that a no is survivable. He argues his own superpower is simply the willingness to mourn failure and try again. > *"The secret to my success is rejection. I ran for sophomore, junior, and senior class president of my high school. I lost all three times."* ## [39:55] Can You Trust The People Building AI? A sharp cultural critique: America has replaced declining religious institutions with tech idolatry, crowning each new CEO as a secular Jesus Christ. Steve Jobs, then Zuckerberg, then Sam Altman, now Dario Amodei — each is briefly positioned as the good guy before completing the villain's journey. Galloway's argument is not that these people are evil but that they are doing exactly what capitalism demands: maximising earnings regardless of wider harm. The answer is not more trustworthy tech founders; it is competent elected officials who regulate them. > *"Can we trust Sam Altman? No. But we shouldn't need to trust him. We should be able to trust that we have smart elected officials that will regulate these companies."* ## [44:50] Are Tech Leaders Quietly Preparing For The End? Galloway reveals that roughly one in three billionaires maintain a "go bag" — a fully funded escape plan, typically a private jet to Auckland and a fortified New Zealand bunker. He calls this nihilism: the ultra-wealthy have sequestered themselves so completely from ordinary infrastructure — private aviation, concierge medicine, private security, elite schools — that they are no longer invested in the health of society. Their disproportionate political donations are therefore not directed at making the system work for everyone. > *"The problem is the 0.1% are not invested in the health of America. They don't have to put up with TSA lines. They fly private."* ## [52:00] Do Some AI Leaders Believe The Risk Is Worth It? A secondhand but chilling account: a source with direct access to an AI CEO described someone who genuinely believes there is a roughly 7–10% chance their work ends in catastrophe, but considers being the person who summoned this new intelligence consequential enough to proceed regardless. Galloway connects this to widening inequality — the delta between middle-class and ultra-wealthy life has expanded so dramatically across healthcare, travel, and security that the incentives of the 0.1% are structurally misaligned with the rest of society. > *"The bottom 99% of Western societies are essentially being optimised and monetised to make the life of the 1% just unbelievable."* ## [58:04] Ads Sponsored segments for LinkedIn Hiring Pro and Function Health. ## [60:05] Could AI Make Us More Human? Galloway offers a surprising positive: unlike social media algorithms that push users toward political extremes, AI models appear to moderate views by seeking statistical medians. He sees genuine value in AI companionship for isolated elderly users. But he returns to his central fear: the biggest downside of AI is not weapons, not election contamination, not even income inequality — it is loneliness. Men aged 20 to 30 are spending less time outdoors than prison inmates, and 42% of men aged 18 to 24 have never asked a woman out in person. > *"The biggest downside of AI in my view is loneliness. AI is convincing people they can have a reasonable facsimile of life on a screen with an algorithm."* ## [65:00] What Happens When AI Becomes Your Closest Companion The conversation shifts to the Iran conflict as a case study in what happens when strategic incompetence meets operational excellence. Galloway credits the initial military strike as tactically credible but argues the absence of Congressional briefing, Gulf ally coordination, and clear exit objectives has produced a quagmire — and notes Iran's IRGC-produced propaganda is outperforming US information operations in the global war of memes. > *"The problem with wars is that the enemy has a say. And all the enemy needs to do — whether it's the Viet Cong or the Taliban or the IRGC — is survive, and they win."* ## [70:00] The Hidden Trade-Off Between Convenience And Real Relationships Galloway diagnoses America's Iran strategy as a product of a gutted diplomatic corps. When senior officials fly to Islamabad expecting a deal, 97% of the preparatory work that career diplomats would normally complete simply has not happened. The IRGC understands the game better: all they need to do is survive, and every day the conflict continues they look like the underdog who stood up to the superpower. His most optimistic scenario is a multinational force enforcing freedom of navigation through the Strait of Hormuz. > *"Do you know what we have done in the US to our diplomatic corps? We've absolutely gutted it."* ## [75:00] Why Loneliness Could Explode US stock markets hit an all-time high during active Middle East conflict — a sign that the wealthy are so insulated from geopolitical risk that war no longer registers in asset prices. The top 10% account for 50% of consumer spending, and that cohort does not care if gasoline hits six dollars a gallon. The pain is outsourced to lower-income households and oil-dependent nations. Galloway frames this dissociation from shared risk as one of the most dangerous structural features of contemporary inequality. > *"We've outsourced the downside of war to less wealthy nations who are very oil dependent, to the Gulf, which is incurring damage here."* ## [79:26] The Real Reason Human Connection Might Become More Valuable Extended discussion of AI market valuations and the historical pattern of infrastructure overbuild. Every great infrastructure boom — railroads, electrification, the internet — ended in a crash, and AI capex now constitutes a significant share of US GDP growth. Galloway argues there is a one-in-three chance AI ends up like jet aviation or vaccines: transformative for humanity but impossible to monetise exclusively for a small group of companies, because open-weight Chinese models could commoditise the entire stack through "AI dumping." > *"AI puts AI out of business. And that is if you look at the convergence of the technologies, all the models are converging."* ## [85:00] What This Means For The Next Generation Galloway argues that a market correction might actually benefit younger generations by making assets affordable again. He flags GLP-1 drugs as his technology pick over AI in terms of real-world human impact. His personal investment philosophy at age 61: aggressive diversification, no single position above 3% of net worth, rotation out of overheated US markets into Europe and Latin America. For young people, the only wealth-building path he trusts is compound interest through low-cost index funds, with money automatically invested before it can be spent. > *"The only answer I have is slowly — find out a way to start saving when you're a teenager, 25 bucks a month, then in your 20s 100, then 500."* ## [90:00] How Power, Politics, And AI Are Becoming Intertwined Drawing on his experience losing 70% of New York Times ad revenue in 60 days during 2008, Galloway warns that younger entrepreneurs have never experienced a true recession. He argues that the political class has systematically bailed out asset-owning baby boomers — COVID relief, corporate bailouts, perpetual market support — while denying younger generations the chance to buy assets at distressed prices. Recessions historically created entry points; that mechanism is now deliberately suppressed. > *"Your generation really doesn't know what a recession looks like. Like, everything stops."* ## [95:00] The Dangerous Gap Between Technology And Regulation Personal finance advice combined with a reflection on the limits of prediction. Galloway's investment rule for young people: put money in yourself first, then in relationships, then in diversified index funds. He is honest that picking winning sectors is largely futile, and that anyone claiming certainty does not know. His own investment in Pokemon cards with his son illustrates that the best investments compound in non-financial ways — relationships and shared experience accrue value that conventional ROI cannot measure. > *"The only answer I have is slowly and it requires some discipline. Save money, diversify, compound interest, invest in relationships early."* ## [100:00] What Happens If Governments Can't Keep Up With AI Asked what a 33-year-old should know that a 61-year-old has learned, Galloway offers three lessons: be humble in success because much of it is luck; forgive yourself in failure because much of it is also circumstance; and invest aggressively in relationships in your 30s, because he spent his prime years professionally focused and nearly ended up isolated. He frames every major disappointment as something people later regret not the thing itself but how upset they allowed themselves to be. > *"Nothing's ever as good or as bad as it seems. Be humble when you're successful. And forgive yourself and realise this will pass."* ## [105:00] The Future Of Work, Power, And Who Really Wins Fatherhood as purpose. Galloway confesses he did not want children and did not fall in love with his sons immediately after birth. What changed his view was discovering that fatherhood is the one investment where a positive financial return is structurally impossible — and that is precisely what makes it purposeful. The same logic applies to any cause large enough to demand more than you can ever get back: veterans, activism, caregiving. He closes with frank advice on partnership, timing, and the liberation of having no choice but to lean into your children's interests. > *"Finding your purpose is finding that thing that you can never get a real positive return on. I will never get a positive return for my children."* ## [110:00] Why The Biggest AI Risks Aren't What You've Been Told The final chapter opens with Galloway's emotional description of his sons' contrasting personalities — one a mirror of himself, one a "different species" he observes with fascination. He discusses his book *Notes on Being a Man*, framing it as letters he hopes his boys will read in 30 years. The closing question — the biggest setback and its lesson — draws the most emotionally raw answer of the episode: his mother's death. He says he has not gotten over it and does not want to, because grief is the receipt for love, and he hopes his sons will one day feel the same about losing him. > *"My mother dying. And you can never tell your parents how much you love them too much. The reverse of love is grief."* ## Entities - **Scott Galloway** (Person): NYU Stern Professor of Marketing, serial entrepreneur, author of *The Four*, *The Algebra of Happiness*, and *Notes on Being a Man*; host of the Prof G Pod and Pivot podcast - **Sam Altman** (Person): CEO of OpenAI; used as the primary case study in the recurring tech-leader idolisation and disillusionment cycle - **Elon Musk** (Person): CEO of Tesla, SpaceX, and xAI; discussed as visionary storyteller whose real products (Starlink, SpaceX) are transformative but whose timelines consistently overshoot - **Dario Amodei** (Person): CEO of Anthropic; cited as the current tech industry "good guy" before the inevitable villain turn - **Jensen Huang** (Person): CEO of Nvidia; held up as a model of storytelling-driven CEO performance via stadium keynotes - **OpenAI** (Organization): Developer of ChatGPT; primary subject of fundraising-hype and overvaluation critique - **Anthropic** (Organization): AI safety company; referenced as beneficiary of the "latest hero" investor narrative - **SpaceX** (Organization): Musk's rocket company; flagged as likely destination for capital migrating away from Tesla at IPO - **Amazon** (Organization): Galloway's top large-cap stock pick for 2026 due to robotics leadership and warehouse automation scale - **Tesla** (Organization): Great car company trading at an unjustifiable multiple that will correct when SpaceX IPOs - **GLP-1 drugs** (Concept): Weight-loss and metabolic medications (Ozempic/Wegovy class) that Galloway argues will create more real-world human impact and shareholder value than AI - **AI dumping** (Concept): Galloway's term for China flooding the US with cheap open-weight AI models to undermine American AI valuations and destabilise the economy - **Go bag / billionaire nihilism** (Concept): The practice among roughly one-in-three billionaires of maintaining funded escape plans as a symptom of disengagement from shared societal wellbeing - **Rejection tolerance** (Concept): Galloway's candidate for the most underrated skill of the AI era — the willingness to hear no, mourn briefly, and try again

Jim Fan, lead of Nvidia's embodied AI research, outlines the transition from language-centric models to World Action Models (WAM) that simulate physical reality. He details a roadmap toward the 'Physical Turing Test' and autonomous factories by 2040, driven by video pre-training and human egocentric data scaling. ## [00:00] Introduction Host Sonya Huang introduces Jim Fan, who leads Nvidia's embodied autonomous research group. Fan reflects on his early days as an intern and the excitement surrounding the future of robotics. > *robots are just one of the most thrilling things that's going to happen.* > *[0, 12]* ## [00:30] DGX One Origin Story Jim Fan recounts the 2016 delivery of the first DGX-1 by Jensen Huang to Elon Musk and the OpenAI team. He highlights how this moment catalyzed the deep learning revolution that led to current AI breakthroughs. > *If you believe in deep learning, deep learning will believe in you.* > *[1, 26]* ## [01:42] The Great Parallel Fan proposes 'The Great Parallel,' applying the successful LLM scaling playbook to robotics. Instead of predicting the next token in a string, the goal is to predict the next physical world state through simulation and alignment. > *instead of simulating strings can we simulate next physical world state?* > *[2, 56]* ## [03:31] Robotics Endgame Setup The strategy for achieving the robotics end game is divided into two primary pillars: model strategy and data strategy. Fan notes that while LLMs are in their final 'boss fight,' robotics is just beginning its scaling journey. > *It boils down to two things, model strategy and data strategy.* > *[3, 32]* ## [03:39] Why VLA Falls Short Visual Language Action (VLA) models are criticized for being 'head-heavy' on language while lacking a fundamental grasp of physics and verbs. Fan argues they are better at encoding static knowledge than dynamic physical interaction. > *VLAs are great at encoding knowledge and nouns, but not so much at physics and verbs.* > *[4, 8]* ## [04:32] Video World Models Fan explains how video models like VEO3 learn internal physics—such as gravity and buoyancy—simply by predicting pixels at scale. These models act as simulators that can solve mazes and plan visual sequences internally. > *Physics emerge by predicting the next blob of pixels at scale.* > *[5, 15]* ## [06:09] DreamZero World Action Nvidia introduces 'Dreamer' and World Action Models (WAM), which jointly decode future world states and motor actions. This allows robots to perform zero-shot tasks by 'dreaming' the correct motion sequence before executing it. > *Dreamer jointly decodes the next world states and next actions.* > *[6, 29]* ## [07:46] Scaling Data Collection To overcome the physical limits of teleoperation, Fan discusses Universal Manipulation Interfaces (UMI) and exoskeletons like Dex-UMI. These tools allow humans to collect high-dexterity data directly without the robot being in the loop. > *we're able to break the curse of 24 hours per robot per day* > *[10, 6]* ## [11:06] EgoScale And Scaling Laws Fan introduces Ego-Exo, a policy trained on 21,000 hours of human egocentric video. This research uncovered a neural scaling law for dexterity, showing a mathematical relationship between pre-training volume and robot performance. > *we discovered this neural scaling law for dexterity.* > *[12, 39]* ## [15:39] DreamDojo And The Roadmap Fan outlines the roadmap to 2040, including the Physical Turing Test and 'lights-out' factories. He introduces Dream Dojo, a neural simulator that replaces classical physics engines with data-driven world models. > *I can say with 95% certainty that we'll get to the end of the end game... by 2040.* > *[19, 19]* ## Entities - **Jim Fan** (person): Lead of the embodied autonomous research group at Nvidia. - **Nvidia** (organization): The technology company developing the hardware and software for the robotics end game. - **Jensen Huang** (person): CEO of Nvidia, mentioned for delivering the first DGX-1 to OpenAI. - **OpenAI** (organization): The research lab that received the first DGX-1 for deep learning development. - **DGX-1** (product): The world's first deep learning supercomputer delivered in 2016. - **VEO3** (model): A video world model capable of simulating physics and visual planning. - **Dreamer** (model): A policy model that predicts future world states and actions simultaneously. - **Ego-Exo** (project): A robotics pre-training framework using large-scale human egocentric video data.

Andrej Karpathy explores the paradigm shift from traditional programming to Software 3.0, where LLMs act as programmable computers driven by context. He details the transition from 'vibe coding' to 'agentic engineering,' emphasizing that while AI handles execution, human taste and understanding remain the ultimate bottlenecks. ## [00:00] Introduction Stephanie Zhan introduces Andrej Karpathy, highlighting his foundational work at OpenAI and Tesla. She notes his unique ability to simplify complex AI shifts and introduces the concept of vibe coding. > *He has a rare gift of making the most complex technical shifts feel both accessible and inevitable. [00:22]* ## [00:44] Feeling Behind as a Coder Karpathy describes a turning point in December 2023 when agentic tools began producing perfect code without manual intervention. This shift led him to adopt vibe coding, trusting the AI to handle complex workflows autonomously. > *I just start to notice that with the latest models the chunks just came out fine. [01:29]* ## [02:28] Software 3.0 Explained Karpathy defines Software 3.0 as a paradigm where the LLM acts as a programmable computer and the context window serves as the primary programming lever. This follows Software 1.0's manual rules and Software 2.0's data-driven weight training. > *Software 3.0 is kind of about your programming now turns to prompting and what's in the context window is your lever. [03:20]* ## [03:44] Agents as the Installer Using the installation of OpenClaw as an example, Karpathy explains how agents replace rigid bash scripts with intelligent, environment-aware execution. This approach allows the AI to debug and adapt to specific system requirements autonomously. > *The agent has its own intelligence that it packages up and then it kind of like follows the instructions. [04:29]* ## [04:49] Menu Gen vs Raw Prompts Karpathy contrasts his custom-coded MenuGen app with raw prompts to models like Gemini, concluding that many traditional software layers are now redundant. He emphasizes that AI can now perform general information processing that was previously impossible with structured code. > *The software 3.0 paradigm is a lot more kind of raw. It just your neural network is doing more and more of the work. [06:11]* ## [07:37] What’s Obvious by 2026 Looking toward 2026, Karpathy envisions neural computers that process raw video and audio directly. These systems would use diffusion models to generate dynamic user interfaces, potentially making traditional UI code obsolete. > *You could imagine completely neural computers... a device that takes raw videos or audio into basically what's a neural net. [08:22]* ## [09:41] Verifiability and Jagged Skills AI models develop 'jagged' capabilities, peaking in verifiable domains like math and code due to reinforcement learning rewards. Karpathy notes the paradox where a model can refactor a massive codebase yet fail simple logic. > *state-of-the-art models today will tell you to walk [to a car wash] because it's so close... This is insane. [11:36]* ## [13:39] Founder Advice and Automation Model performance is heavily dictated by the specific data distributions chosen by frontier labs. Karpathy advises founders to explore the 'circuits' of these models to understand their strengths or use fine-tuning to fill gaps. > *we are slightly at the mercy of whatever the labs are doing, whatever they happen to put into the mix. [12:57]* ## [15:46] From Vibe Coding to Agent Engineering While 'vibe coding' raises the accessibility floor, 'agentic engineering' focuses on maintaining professional quality. This discipline involves coordinating powerful but stochastic agents to accelerate development without sacrificing the engineering bar. > *agentic engineering is about preserving the quality bar of what existed before in professional software. [16:07]* ## [25:17] Agents Everywhere and Learning Karpathy advocates for agent-native infrastructure, expressing frustration with human-centric documentation. He argues that while thinking can be outsourced to AI, human understanding remains a critical bottleneck for directing agents. > *You can outsource your thinking, but you can't outsource your understanding. [28:10]* ## Entities - **Andrej Karpathy** (person): AI researcher and former Director of AI at Tesla and founding member of OpenAI. - **Stephanie Zhan** (person): Partner at Sequoia Capital and host of the discussion. - **Software 3.0** (concept): A paradigm where LLMs act as programmable computers via prompting and context. - **Agentic Engineering** (concept): The professional discipline of coordinating AI agents to maintain software quality. - **MenuGen** (project): An app Karpathy built to OCR and visualize restaurant menus, used as a case study. - **OpenAI** (organization): AI research company co-founded by Karpathy. - **Gemini** (ai-model): Google's LLM used in Karpathy's software comparison. - **Vercel** (organization): A cloud platform used by Karpathy to deploy projects.

이방카 트럼프가 유명한 부모와 극심한 미디어의 관심 속에서 형성된 독특한 유년 시절부터 비즈니스와 공직에서의 영향력 있는 경력까지 솔직하게 들려줍니다. 어머니에게 배운 교훈, 신뢰를 쌓는 데 따르는 어려움, 부모의 이혼과 아버지에 대한 암살 시도 같은 결정적 경험이 어떻게 회복력을 키워주었는지 이야기합니다. 또한 의도적으로 사는 철학, 과소평가의 힘, 그리고 모성과 심리 치료를 통한 개인적 성장의 여정을 공유하며, Planet Harvest를 통한 사명 중심의 활동으로 대화를 마무리합니다. ## [00:00] 신뢰가 쉽지 않은 이유와 그것이 드러내는 것 이방카 트럼프는 특히 아홉 살 때 부모의 대대적으로 보도된 이혼을 겪으며, 끊임없는 미디어의 관심과 공격적인 파파라치로 인해 진심 없는 관계를 일찍부터 경계하는 법을 배웠습니다. 어머니는 과소평가당하는 것의 힘과 압박 속에서 외부의 "잡음"을 걸러내는 중요성을 가르쳐주었습니다. 처음에는 타인을 신뢰하지 않는 강력한 방어 기제를 발달시켰지만, 이후 더 깊은 유대를 위해 의식적으로 신뢰하는 태도를 기르며 그에 따르는 위험을 받아들이고 있습니다. > *어머니는 과소평가당하는 것이 나쁜 게 아니라고 가르쳐주셨어요. 사실 매우 강력한 거라고요 [00:22]* > *저는 실제로 스스로에게 더 신뢰하는 법을 가르쳤어요. [05:48]* ## [03:32] 자신이 다르다는 것을 깨달았을 때 벌어지는 일 이방카 트럼프는 끊임없는 미디어의 관심과 대중의 시선으로 인해 어린 시절부터 자신의 삶이 평범하지 않다는 것을 깨달았으며, 이를 오늘날 소셜 미디어로 더욱 증폭된 아이들의 노출과 대비합니다. 부모가 그녀와 형제들을 이 강렬한 대중의 시선으로부터 보호하려 노력했다고 합니다. 그녀는 잦은 인터뷰보다 깊이 있는 대화를 선호합니다. > *항상 미디어의 관심과 감시가 많았던 것 같아요. 그걸 보고, 아주 일찍부터 경험하게 되죠. [06:24]* > *모든 아이들이 그런 건 아니지만, 어디를 가든 사람들이 손에 녹화 장치를 들고 있는 경험을 우리 아이들은 하고 있어요 [06:40]* ## [05:44] 닫힌 문 뒤에서 어머니의 진짜 모습 이방카 트럼프는 어머니 이바나를 전직 국가대표 스키 선수로서 스포츠의 가치를 심어주고 이방카가 발레를 하게 이끈 규율 있는 사람으로 묘사합니다. Michael Jackson이 자신의 호두까기 인형 공연에 참석한 특이한 어린 시절의 기억을 회상합니다. 이런 비범한 경험에도 불구하고 일상은 외할머니 "버비"가 무조건적인 사랑을 베풀고 요리로 사랑을 표현하며 안정감을 주었습니다. > *어머니는 놀라운 스키 선수였어요... 규율을 기르기 위한 스포츠의 중요성을 정말로 믿으셨어요 [07:07]* > *외할머니가... 정말로 우리를 키우셨어요... 무조건적인 사랑과 다정함을 가르쳐주셨어요 [08:44]* ## [11:47] 그녀를 만든 결정적 차이 이방카 트럼프의 성장은 무조건적인 사랑과 매일의 돌봄을 제공한 외할머니 "버비"와 선구적인 롤모델 역할을 한 어머니 이바나 모두에 의해 깊이 형성되었습니다. 이바나는 강인함, 야망, 회복력의 본보기를 보여주며 전문적 목표를 추구하면서도 사랑하는 어머니가 되는 법을 보여주었습니다. 이방카는 바쁜 부모의 커리어에도 불구하고 그들이 곁에 있었고 자신이 최우선이라는 느낌을 주었으며, 외할머니가 전통적인 양육자 역할을 채워주었다고 말합니다. > *어머니는 놀라운 선구자였어요... 강인함과 회복력, 우아함, 결단력, 야망의 놀라운 본보기였어요. [11:57]* > *제가 아버지의 최우선 순위이고, 아버지가 저에게 언제든 함께할 수 있다는 것에 한 번도 의심한 적이 없어요. [14:42]* ## [15:43] 도널드와 이바나 트럼프의 이혼이 그녀에게 의미한 것 도널드와 이바나 트럼프의 대대적으로 보도된 이혼은 이방카가 아홉 살 때 신문에서 알게 되었고, 그녀에게 깊은 영향을 미쳤습니다. 극심한 미디어의 관심에 두려움을 느꼈고 부모의 별거 중 아이로서 겪는 정상적인 공포를 경험했습니다. O.J. Simpson 재판보다 더 많은 헤드라인을 장식한 이 힘든 시기는 형제들과의 독특한 유대를 만들어주었습니다. 어머니가 세상을 떠난 후, 이방카는 공산주의 체코슬로바키아에서 자란 환경이 형성한 이바나의 복잡한 성격을 더 깊이 이해하게 되었고, 어머니가 살아 계실 때 더 많은 질문을 했으면 좋았을 것이라고 말합니다. > *이 이혼은 O.J. Simpson 재판보다 더 많은 헤드라인을 장식했대요. [20:04]* > *저와 형제들에게 좋은 점은 함께 겪고 있었기 때문에 다른 방식으로 정말 유대감이 생겼다는 거예요. [23:21]* ## [18:27] 트럼프의 딸로 산다는 것의 현실과 사람들의 오해 도널드 트럼프의 딸이라는 것은 어린 시절부터, 특히 부모의 이혼 동안 극심한 대중의 시선을 견뎌야 하는 것이었고, 이는 신뢰에 대한 필수적인 경계심을 가르쳐주었습니다. 이후 그녀는 "잡음 속에서 신호를 찾는" 법을 배우고 전투적인 소셜 미디어를 피하며 내면의 평화를 우선시합니다. 이방카는 부모의 깊은 진정성을 언급하며, 자신은 의사소통에서 더 섬세한 접근 방식을 취하지만 스토아 철학의 인도를 받아 진정성 있게 살고 외부 압력에 저항하는 강한 자아의식을 유지한다고 말합니다. > *그 교훈이 없었다면 강인해질 수 있었을지 모르겠어요. 그것은 누구도 믿지 말라고 가르쳐주었어요. [18:53]* > *저는 반격하지 않아요. 왜냐하면... 시간과 집중을 전투적으로 쓰거나 소셜 미디어의 불쾌한 소용돌이에 뛰어드는 것을 믿지 않거든요. [26:19]* ## [23:36] 권력과 명성에 둘러싸인 채 자아를 찾는 법 권력과 명성에 둘러싸인 가운데, 이방카 트럼프는 의도적인 자기 성장과 자신을 "열어젖힌" 모성의 변혁적 경험을 통해 자아를 찾았고, 이는 사랑의 능력을 더 깊게 만들어주었습니다. 그녀는 외부 압력에 저항하고 "군중이 이기지" 않도록 자신을 정의하는 자기 인식의 결정적 중요성을 강조합니다. 이 철학을 육아에 적용하여 자녀의 개성을 키우고 있으며, 자신의 부모가 존중하는 범위 내에서 반대 의견을 허용해준 덕분에 진정한 자아를 지킬 수 있었다고 감사를 표합니다. > *자신이 누구인지 모르면 군중이 이겨요. [29:55]* > *그들은 반대 의견이 괜찮은 환경을 만들어주었어요. [32:44]* ## [30:57] 과소평가가 최대의 무기가 된 이유 이방카 트럼프는 어머니에게서 과소평가당하는 것이 강력한 무기가 될 수 있다는 것을 배웠습니다. 부동산 커리어 초기에 성공한 부모의 자녀이자 남성 중심 업계의 젊은 여성으로서 종종 잘못 평가받았습니다. 그녀는 이 인식을 활용하여 더 열심히 일하고 철저히 준비하는 동기로 삼았고, 궁극적으로 자신을 과소평가한 사람들에게 유리하게 활용했습니다. > *어머니는 과소평가당하는 것이 나쁜 게 아니라고 가르쳐주셨어요. 사실 매우 강력한 거라고요 [00:22]* > *저는 그 두려움, 그 감정을 활용해서 저를 앞으로 나아가게 하는 데 사용했어요. [35:06]* ## [32:59] 채용할 때 실제로 보는 것과 그것이 중요한 이유 채용할 때 이방카 트럼프는 강한 자아의식, 주체성, 좋은 판단력, 그리고 타고난 감각인 "거리의 지혜"를 가진 사람을 우선시합니다. 이런 본질적인 자질은 가르치기 어렵기 때문입니다. 그녀는 신뢰하고 존경하는 "좋은 사람"과 함께 일하는 것의 중요성을 강조하며, 이러한 자질이 성공적인 업무 관계와 전체 팀 역학의 기본이라고 봅니다. > *사람들에게 가르치기 매우 어려워요. 똑똑한 사람이라도 좋은 판단력이 없거나 자기 주도적이지 않으면, 그걸 심어주기 매우 어렵죠. [38:15]* > *좋은 사람이라고 생각하지 않는 사람과 함께 일하고 싶지 않아요. 신뢰하지 않거나 존경하지 않는 사람과 시간을 보내고 싶지 않으니까요. [39:00]* ## [37:49] 패션을 떠나 정부로 향한 이유 Wharton 졸업 후 Anna Wintour로부터 Vogue에서의 명망 있는 직업 제안이 있었음에도 불구하고, 이방카 트럼프는 평생의 열정이었던 부동산을 추구했습니다. 이후 IvankaTrump.com이라는 성공적인 패션 브랜드를 구축하여 연간 약 8억 달러의 매출을 올렸습니다. 그러나 아버지의 행정부에서 일해달라는 요청을 수락하면서 정부 윤리 규정을 준수하기 위해 이 번창하는 사업을 의도적으로 폐업하는 결정을 내렸습니다. 그녀는 이 기회를 막대한 개인적, 직업적 희생에도 불구하고 부정할 수 없는 특권이자 국가에 대한 의무로 보았습니다. > *정부에 들어가면서 사업을 접었을 때 연간 약 8억 달러의 매출을 올리고 있었어요. [42:30]* > *아버지가 우리에게 사랑하는 나라를 위해 봉사할 기회를 준 것에 믿을 수 없을 정도로 감사해요. [43:30]* ## [41:06] 트럼프가 출마를 결심했을 때 실제로 일어난 일 도널드 트럼프의 2015년 대통령 출마 결정은 Bedminster에서의 가족 회의에서 발표되었으며, 1980년대부터 오래되었지만 공개적으로 표명되지 않았던 정치적 야심에도 불구하고 그 신속함에 이방카는 놀랐습니다. 16살 때 아버지가 출마할 것이라는 두려움에 패닉에 빠졌다가 안심시켜 주었던 기억을 회상합니다. 대통령 선거 정치에 뛰어든 것은 가족에게 "급진적인 전환"이었으며, 이방카의 세계관을 뉴욕시의 "버블" 너머로 크게 확장시키고 공직 봉사라는 "특별한 여정"을 시작하게 했습니다. > *한 번 진짜라고 생각한 적이 있어요. 16살이었고 기숙학교에 있었는데 아버지에게 전화해서... '이건 내 인생을 망칠 거야.'라고 했어요. [51:48]* > *아버지의 선거 캠페인이 제게 그걸 열어젖혔고, 제가 있던 버블을 깨달았어요 [48:02]* ## [46:23] 트럼프의 대선 출마가 모든 것을 바꾸다 도널드 트럼프의 대통령 출마 결정은 이방카에게 모든 것을 근본적으로 바꾸었고, 온 가족에게 "급진적인 전환"이었습니다. 전통적인 경력 경로를 우회한 파격적인 정치 진출은 "소방 호스로 물을 마시는 것"과 같았습니다. 선거 캠페인은 이방카가 느꼈던 뉴욕시의 "버블"을 산산조각 내고 세계관을 깊이 확장시켰으며, 국가를 위해 봉사하는 특권을 받아들이게 만들었습니다. > *우리 모두에게 소방 호스로 물을 마시는 것 같았어요. [47:08]* > *아버지의 선거 캠페인이 제게 그걸 열어젖혔고, 제가 있던 버블을 깨달았어요 [48:02]* ## [48:52] Ads 이 부분은 온라인 쇼핑몰 구축, 소셜 미디어 판매, AI 도구를 통한 운영 관리를 간소화하는 전자상거래 플랫폼 Shopify의 광고를 소개합니다. 또한 호스트가 사용하는 지능형 CRM인 Pipe Drive를 소개하며, 판매 프로세스를 한 대시보드에서 시각적으로 보여주는 파이프라인 기능을 강조합니다. > *Shopify는 쉽게 시작할 수 있어요. 스토어를 만들고, 소셜에서 판매하고, 결제를 받고, AI 도구를 사용하고, 모든 것을 한 곳에서 관리할 수 있으니까요. [49:22]* > *Pipe Drive는 사용하기 쉬운 지능형 CRM이에요... 하나의 대시보드로 판매 프로세스를 시각적으로 보여줘요. [50:17]* ## [51:04] 아버지가 정말로 해낼 거라고 생각했을까 도널드 트럼프가 1980년대부터 대통령 출마를 고려했지만, 이방카는 어린 시절 이 야심이 명시적으로 논의된 적은 없었다고 말합니다. 16살 때 아버지가 출마한다고 믿고 패닉에 빠졌다가 안심시켜 주었던 순간을 생생히 기억합니다. 무역 정책 같은 문제에 대한 아버지의 견해는 수십 년간 변하지 않았다고 합니다. > *한 번 진짜라고 생각한 적이 있어요. 16살이었고 기숙학교에 있었는데 아버지에게 전화해서... '이건 내 인생을 망칠 거야.'라고 했어요. [51:48]* > *무역 정책에 대한 아버지의 관점은 시간이 지나도 변함없이 일관되었고 오늘날까지 그대로예요 [52:35]* ## [54:26] 백악관을 떠나는 것은 안도였을까, 아니면 다른 무엇이었을까 백악관을 떠나는 것은 후회의 의미에서의 안도는 아니었습니다. 이방카 트럼프는 "경기장에 모든 것을 쏟았다"고 느끼며 4년간의 공직 봉사에서의 성과에 자부심을 가지고 있습니다. 봉사의 기회를 "놀라운 특권"으로 여기지만 정치로 돌아갈 의향은 없으며, 자녀를 최우선으로 생각하고 더 이상의 공적 생활의 대가를 아이들이 치르게 하고 싶지 않습니다. 자신의 기여에 만족하며 아버지에게는 이제 지지해줄 훌륭한 팀이 있다고 봅니다. > *경기장에 모든 것을 쏟았어요. 돌아보면서... 후회는 없어요. [53:33]* > *저의 첫 번째 책임은 아이들의 엄마가 되는 것이에요. [56:49]* ## [58:08] 백악관 생활에 진정으로 준비된 사람이 있었을까 이방카 트럼프는 고위 정치와 백악관 생활의 강렬한 경험에 진정으로 준비시켜주는 것은 아무것도 없다고 인정합니다. 부와 마찬가지로 권력은 사람들의 본질적인 특성을 증폭시키는 경향이 있다고 관찰했습니다. 군주부터 선출된 지도자까지 세계 지도자들과의 교류를 통해 그들의 신비가 벗겨졌고, 그들의 핵심은 평범한 고민을 가진 "그냥 사람"이라는 것을 깨달으며 느꼈던 위압감이 사라졌습니다. > *그 경험을 위해 준비시켜주는 것은 아무것도 없어요. [58:26]* > *결국 사람은 사람이라는 걸 깨닫게 되죠. [59:03]* ## [59:44] 암살 시도가 영원히 바꿔놓은 것 2024년 7월 아버지에 대한 암살 시도는 이방카 트럼프의 삶을 근본적으로 바꿔놓았고, 보안 우려를 강화시켜 미국 비밀경호국의 보호가 필요하게 되었습니다. 아이들과 함께 실시간으로 사건을 목격하며 첫 반응은 아이들을 돌려세우는 것이었지만, 아버지가 괜찮을 것이라는 직감이 있었습니다. 이 충격적인 경험과 다른 가족의 건강 위기는 삶의 소중함에 대한 믿음을 강화시켰고, 공직과 폭력의 우려스러운 상관관계에도 불구하고 긍정을 선택하고 매 순간을 소중히 여기기로 한 결심을 굳혔습니다. > *제 첫 반응은 아이들을 돌려세우는 것이었어요. [62:02]* > *인생에서 어떻게 반응할지에만 선택권이 있어요. 저는 긍정적인 결과를 보기로 해요. [66:05]* ## [1:07:20] 정치에서 물러난 후의 삶 2022년 정치에서 물러난 후, 이방카 트럼프의 삶은 이제 어린 자녀들과 가정생활을 최우선으로 합니다. 정치의 "어두운 세계"가 자신의 본성과 맞지 않았기 때문입니다. 대중의 비판을 "독수리와 까마귀" 비유를 사용해 헤쳐나가며, 싸우기보다 부정적인 것 위로 날아오르기를 택합니다. 아버지의 거의 죽을 뻔한 경험을 포함한 극심한 공적 시선의 시기는 개인적 성장의 "약"이 되었고, 통제할 수 있는 범위 안에서 내면의 평화와 조화를 찾고 삶의 축복에 감사하는 데 집중하도록 가르쳐주었습니다. > *정치는 꽤 어두운 세계예요. 어둠이 많고, 부정적인 것이 많고, 인간으로서 좋다고 느끼는 것과 정말 맞지 않아요. [67:45]* > *독수리의 반응은... 비틀거나 돌려서 까마귀를 떨어뜨리거나 방어하는 게 아니에요... 그냥 더 높이 나는 거예요. [69:28]* ## [1:11:04] Ads 이 챕터는 팟캐스트 내의 짧은 광고 시간입니다. ## [1:14:24] 심리 치료가 모든 것을 보는 방식을 바꾸다 이방카 트럼프는 "성장 지향적 사고방식"과 중요한 삶의 사건들을 처리하려는 욕구에서 비롯되어 성인 심리 치료를 시작했으며, 이를 "내면의 재고 조사" 도구로 봅니다. 주요 계기에는 남편 Jared의 두 번째 갑상선암 진단, 워싱턴에서의 퇴임, 그리고 어머니의 갑작스러운 타계가 포함됩니다. 치료는 감정을 구획화하기보다 자신을 돌보고 감정을 처리하는 데 도움을 주었으며, 궁극적으로 자기 이해와 앞으로 나아가는 것에 대한 관점을 바꾸었습니다. > *저는 매우 성장 지향적인 사고방식을 가지고 있어요... 항상 자신과 세상에 대해 배우려고 해요 [74:35]* > *Jared가 두 번째로 갑상선암 진단을 받았어요. 그리고... 어머니가 돌아가셨어요 [75:59]* ## [1:20:28] 어머니의 상실이 가르쳐준 것 이방카 트럼프는 2022년 어머니 이바나 트럼프의 갑작스럽고 비극적인 죽음을 회상하며, 예기치 않은 부모의 상실이 주는 독특한 충격을 이야기합니다. 그녀는 불편함에 직면하고 감정을 처리하며 제대로 된 애도 과정을 거치기로 결심했습니다. 부모로서 이제 자녀들에게 어머니의 긍정적인 면을 보여주면서 어머니의 어려움을 전하지 않도록 의식적으로 노력하며, 어머니의 삶에 대해 더 명확한 어른의 시각을 갖게 되었습니다. > *그래도 좋은 삶을 사셨어요. [81:07]* > *그녀를 완전히 우상화하던 아이의 눈이 아니라, 명확하게 보는 어른의 눈으로 어머니를 생각하는 시간을 정말 가졌어요. [83:15]* ## [1:26:28] 성공과 행복을 정의하는 3가지 원칙 이방카 트럼프는 진정한 성공과 행복이 세 가지 핵심 원칙으로 정의된다고 믿으며, 특히 기업가 정신에 있어 딸 Arabella에게 전하고 싶은 것이라고 합니다. 첫째, 하는 일을 진정으로 사랑해야 합니다. 열정이 헌신의 필수 조건이기 때문입니다. 둘째, 진정성이 가장 중요합니다. 자기 자신이 되어 자신만의 길을 개척하는 것이 핵심이며, 모방은 패배로 이어집니다. 셋째, 가장 근본적으로, 세상이 믿기 전에 자신을 먼저 믿어야 합니다. 이것이 모든 성취의 출발점이기 때문입니다. 또한 전통적인 "일과 삶의 균형"은 잡기 어려우며, 대신 우선순위와의 조화를 추구한다고 말합니다. > *정상에 있으면서 자기가 하는 일을 정말로 사랑하지 않는 사람을 본 적이 없어요. [92:46]* > *세상이 당신을 믿기 전에 자기 자신을 먼저 믿어야 해요. [94:48]* ## [1:28:37] Planet Harvest란 무엇이며 왜 생각보다 중요한가 Planet Harvest는 음식물 쓰레기를 줄이고 미국 농부를 지원하는 것을 목표로 한 이방카 트럼프의 사명 중심 사업입니다. 이 사업은 코로나19 팬데믹 동안 공급망 문제로 대량의 부패하기 쉬운 농산물이 버려지는 것을 목격하면서 영감을 받았습니다. Planet Harvest는 엄격한 외관 기준을 충족하지 못한다는 이유로 소매업체에 의해 거부되는 완벽하게 좋은 식품의 지속적인 문제를 해결하여 농부들에게 추가 수입을 제공하고 환경에도 이롭습니다. > *Planet Harvest는... 팬데믹 초기에 보았듯이 사람들이 식품을 필요로 할 때 밭의 음식이 갈아엎어져 낭비되지 않도록 하는 데서 탄생했어요. [89:18]* > *매년 4억 파운드의 딸기가 밭에 버려져요... 불완전해서가 아니에요. 정말 엄격한 외관 규격에 맞지 않을 뿐이에요. [90:57]* ## 등장인물·개념 - **Ivanka Trump** (인물): 도널드와 이바나 트럼프의 딸, 사업가이자 전 정부 관료. - **The Diary Of A CEO** (단체): 인터뷰를 진행한 팟캐스트. - **Donald Trump** (인물): 이방카 트럼프의 아버지, 전 미국 대통령. - **Ivana Trump** (인물): 이방카 트럼프의 어머니, 전 체코슬로바키아 스키 선수. - **Michael Jackson** (인물): 미국의 유명 가수, 작곡가, 댄서. - **O.J. Simpson** (인물): 전 미국 풋볼 선수, 방송인, 배우, 유죄 판결 범죄자. - **Marcus Aurelius** (인물): 로마 황제이자 스토아 철학자. - **Shopify** (단체): 온라인 쇼핑몰 구축을 위한 전자상거래 플랫폼. - **Pipe Drive** (단체): 지능형 CRM(고객 관계 관리) 소프트웨어. - **Anna Wintour** (인물): Vogue 편집장. - **Vogue** (단체): 패션 및 라이프스타일 잡지. - **Wharton School of Business** (단체): 펜실베이니아 대학교 경영대학원. - **Office of Government Ethics** (단체): 이해 충돌 방지를 담당하는 미국 정부 기관. - **Jared Kushner** (인물): 이방카 트럼프의 남편, 역시 정부에서 근무함. - **US Secret Service** (단체): 이방카 트럼프와 그 가족의 보호를 담당하는 정부 기관. - **Planet Harvest** (단체): 이방카 트럼프가 공동 창립한 음식물 쓰레기 감소 및 농부 지원 사업. - **Arabella** (인물): 이방카 트럼프의 장녀. - **Stoicism** (철학): 고대 그리스의 스토아 철학. - **Buddhism** (철학): 동양 철학. - **Daoism** (철학): 동양 철학. - **Czechoslovakia** (장소): 중앙유럽의 구 국가. - **New York City** (장소): 미국의 주요 도시. - **Bedminster, New Jersey** (장소): 이방카 트럼프가 아버지에 대한 암살 시도 소식을 들은 장소. - **Child Tax Credit** (정책): 자녀가 있는 가정을 위한 미국 세액공제. - **Great American Outdoors Act** (정책): 이방카 트럼프가 지지한 법안. - **Human Trafficking Legislation** (정책): 이방카 트럼프가 공직 재직 중 추진한 법안. - **Vocational Education and Skills Training** (정책): 이방카 트럼프가 미국 근로자의 직업 교육 및 재교육을 위해 추진한 프로그램. - **Meditations** (서적): Marcus Aurelius의 개인적 저술.

Anthropic's three-minute walkthrough of the loop they consider the single most important habit when working with Claude Code: research first in plan mode, define what "done" looks like before any file is touched, then have a subagent review the diff before you push. ## [00:03] Why explore-plan-code-commit beats jumping straight in The opening pitch is blunt — if you only adopt one habit from the course, make it this workflow. The failure mode it's fighting is the reflex of pasting a task into Claude and watching it generate code immediately, which front-loads speed but back-loads correction cost. > *Without this, most people jump straight to pasting in Claude to write code, which means more course correcting later on.* ## [00:21] Plan mode: read-only research before any edits Plan mode is how you collapse explore and plan into a single move. Claude can read files and run web searches but is forbidden from writing — Shift+Tab cycles into it from the prompt. The narrator demos with a real ask (add WebP conversion to an image upload pipeline, figure out where it belongs, what dependencies are needed, how to approach it). Claude returns a plan; you read it, ask for revisions if it misses something. This is the cheapest place in the whole cycle to change direction, because nothing has been written yet. > *With plan mode, Claude can't edit files. It just reads files to gather research on how to tackle this implementation.* ## [01:11] Approve the plan, then course-correct as Claude codes Once the plan looks right, Approve hands execution back to Claude to tick through the checklist. You choose whether file edits auto-accept or prompt every time. Claude will troubleshoot on its own, but expect to intervene — and the reason plan mode pays off here is that the agent now carries the research context that produced the plan, so mid-flight corrections land in the right place instead of starting from scratch. > *This is the benefit of working with plan mode because after the plan is finished, we also have the context of how it got to the results to help it guide its next decision.* ## [01:39] Make success criteria explicit and give Claude real tools A plan without a definition of "correct" leaves Claude guessing. Spell out what success looks like, then equip the agent to actually verify it: the Claude+Chrome extension lets it drive a browser tab to test a UI it just built; a test suite gives it something to validate against on every loop, and Claude can author the tests too — but only if you've already vetted them as ground truth. A quick durability tip: when Claude keeps re-hitting the same problem, have it persist the fix into the CLAUDE.md file so it stops relearning. > *In order for Claude to be confident in its results, it has to be clear on what it deems correct.* ## [02:24] Subagent review, commit, recap Before pushing, spin up a subagent code reviewer over the diff — a second pass with no attachment to the implementation. Then have Claude draft the commit message in your style and ship it. The recap reframes each step: Explore feeds context, Plan defines success, Code is the back-and-forth that converges on the plan, Commit reviews and pushes so you can move on. > *A tip before you commit, run a sub agent code reviewer to look at your code.* ## Entities - **Anthropic Tutorial Narrator** (Person): Anthropic's official voice-over for the Claude Code 101 course. - **Claude Code** (Software): Agentic terminal coding tool whose recommended day-to-day loop is the subject of this episode. - **Plan mode** (Feature): Read-only mode toggled with Shift+Tab — Claude researches and proposes a plan but cannot edit files. - **Claude + Chrome extension** (Software): Lets Claude Code drive a Chrome tab to verify UI changes before declaring a task done. - **CLAUDE.md** (File): Project memory file used here as a persistence target for recurring fixes Claude keeps relearning. - **Subagent code reviewer** (Pattern): Pre-commit Claude subagent that reviews the diff before the human pushes.

Anthropic's Claude Code 101 walkthrough on context — what fills the window, when auto-compaction kicks in, and the practical levers (/compact, /clear, /context, claude.md, MCP toggles, skills, sub agents) for keeping a session lean enough to keep working. ## [00:03] Why context is finite — and why it matters Context is Claude's working memory: every prompt, every file read, every tool call result lands in the same window. The window is large but finite, so optimizing what goes in is non-negotiable once you start running multi-step sessions. > *Every file it reads, every command it runs, every message you send, it all takes up space in the context window.* ## [00:39] Auto-compaction and the /compact command As you near the limit, Claude Code auto-compacts: it summarizes the important bits and drops noisy tool-call results to free space. You can also trigger `/compact` manually — useful when you want headroom but still want to remember what you've been working on. Tradeoff: compaction can lose detail from earlier turns. > *Compaction will summarize important details and remove the unnecessary tool call results and free up a lot of space in your context window.* ## [01:11] /clear and /context: starting over, seeing what's used If you want a true reset with no memory of the prior session, `/clear` wipes everything. To see where your space is actually going, `/context` shows total size, the categories eating the most, and a graphic of the breakdown — the diagnostic before you decide between compact and clear. > *To check the state of your context, run the /context command.* ## [01:35] The rule of thumb: compact mid-feature, clear between features The narrator gives a clean heuristic: still working on one feature and bumping the ceiling? Compact — you want the relevant history to carry over. Done with the plan, moving to something new? Clear — old conversation will bias the new work. > *If you have finished the plan and want to start on a new feature, then clear. You don't want the previous conversation to present bias in anything new that you want to create.* ## [01:57] claude.md, prompt specificity, and writing less by writing more Anything Claude should remember across sessions belongs in `claude.md` so it doesn't rediscover the same facts every time. And counterintuitively, terse prompts cost more context: when the ask is vague, Claude grep-walks the codebase and reasons more, all of which fills the window. A sentence or two of specificity buys back a lot of space downstream. > *The irony behind writing a smaller prompt is that it in the long run, it will take up more context.* ## [02:26] MCP servers, skills, and sub agents as context tools MCP servers load every tool they expose into context by default — fine if relevant, expensive if not, so turn off the ones unrelated to the project. Skills behave like MCP servers but don't dump the whole surface into context. Sub agents run in parallel with their own separate window, so for fact-finding tasks ("where are the auth endpoints?") you can dispatch a sub agent and get back just the answer instead of the whole journey. > *Sub agents run in parallel with your main agent but has a complete separate context window.* ## [03:06] Recap Managing context in Claude Code is the difference between a long productive session and a stalled one. Use `/compact` to summarize long sessions, `/clear` to start fresh, be specific in prompts, check `/context` to see what's eating the window, and delegate answer-only work to sub agents. > *Managing context within cloud code is crucial. Use slash compact to summarize long sessions and slashclear to start fresh.* ## Entities - **Anthropic Tutorial Narrator** (Person): Anthropic's official voice-over for the Claude Code 101 tutorial series. - **Claude Code** (Software): Anthropic's agentic terminal coding assistant whose context window is the subject of this episode. - **Context window** (Concept): Claude's working memory — finite, filled by prompts, file reads, and tool-call results. - **/compact** (Command): Slash command (and auto-trigger) that summarizes history and drops tool-call noise to free space. - **/clear** (Command): Slash command that wipes the session entirely for a clean start on new work. - **/context** (Command): Slash command that reports total context size and which categories are consuming it. - **claude.md** (File): Project-level memory file Claude reads across sessions so it doesn't rediscover the same facts. - **MCP servers** (Software): Tool providers that load all exposed tools into context by default — toggle off when unrelated. - **Skills** (Feature): Lighter-weight alternative to MCP servers that avoids loading the whole tool surface into context. - **Sub agents** (Feature): Parallel agents with their own context windows used to answer scoped questions without polluting the main window.

Terence Tao는 수학에서 AI가 맡게 될 진화하는 역할을 논하면서, AI가 많은 정형 업무를 자동화하겠지만 인간 수학자를 완전히 대체하지는 않고 오히려 그들이 새로운 영역에 집중하도록 만든다고 주장한다. 그는 인간과 AI의 협업이 열어갈 미래, 그리고 AI가 과학적 발견에 미칠 장기적 영향의 예측 불가능성을 강조한다. ## [00:10] 프런티어 수학에서 AI의 현재 역할 Terence Tao는 AI가 이미 인간은 할 수 없는 '프런티어 수학'을 수행하고 있지만, 그 프런티어는 우리가 익숙했던 것과는 다른 종류라고 설명한다. 그는 이를 과거에 계산기가 인간의 능력을 뛰어넘는 작업을 전문화된 방식으로 처리하며 수학의 가능성을 확장했던 방식에 비유한다. > *어떤 면에서 그것들은 이미 인간이 할 수 없는 초지능적인 프런티어 수학을 수행하고 있지만, 우리가 익숙한 프런티어와는 다른 종류의 프런티어입니다.* ## [00:52] AI는 대체가 아닌 자동화 도구 Tao는 10년 안에 AI가 현재 수학자들이 수행하는 많은 정형 업무를 대신 처리하면서, 인간은 더 복잡하고 중요한 문제에 집중할 수 있게 될 것이라고 전망한다. 그는 과거에 컴퓨터가 '인간 계산수'의 업무를 자동화했거나, 유전체 분석이 자동화된 뒤에도 유전학이라는 학문이 새로운 규모로 계속 진화한 역사적 전환을 예로 든다. > *10년 안에 지금 수학자들이 하는 많은 일들이… AI에 의해 수행될 수 있을 것입니다. 하지만 그것이 우리 작업에서 가장 중요한 부분은 아니었다는 것을 우리는 알게 될 것입니다.* ## [02:46] 수학에서의 인간-AI 협업의 미래 Dwarkesh Patel은 AI가 밀레니엄 난제를 자율적으로 풀 수 있는지 묻는다. Terence Tao는 '인간+AI 하이브리드'가 앞으로도 오랫동안 수학을 지배할 것이라고 본다. 현재의 AI는 지적 작업을 완전히 대체할 모든 요소를 갖추지 못했기에 보완적 도구로 기능한다는 설명이다. > *인간과 AI의 하이브리드가 앞으로도 오랫동안 수학을 지배할 것이라고 저는 믿습니다.* ## [03:43] 과학적 발견에 미칠 예측 불가능한 영향 Tao는 AI가 과학과 새로운 발견을 가속화하는 동시에, '우연성을 파괴함'으로써 특정 유형의 진보를 저해할 가능성도 있음을 인정한다. 그는 AI가 과학적 발견에 미칠 미래의 영향은 매우 예측 불가능하다고 결론짓는다. > *AI가 어떤 식으로든 우연성을 파괴함으로써 실제로 특정 유형의 진보를 저해할 가능성도 있습니다.* ## 등장인물·개념 - **Terence Tao (테렌스 타오)** (인물): 게스트이자 당대를 대표하는 수학자. - **Dwarkesh Patel** (인물): 해당 팟캐스트의 호스트. - **AI** (개념): 인공지능. 수학과 과학적 발견에서의 역할을 논의함. - **Mathematica / Wolfram Alpha** (소프트웨어): 수학 자동화의 예시로 언급된 계산 도구. - **밀레니엄 난제 (Millennium Prize Problems)** (개념): 수학의 7대 미해결 난제. 각 문제에 100만 달러의 상금이 걸려 있음.

서브에이전트는 중간 작업이 메인 스레드에 속하지 않을 때 강력한 도구가 됩니다. 하지만 무분별하게 위임하면 오히려 상황이 나빠집니다. 이 튜토리얼은 유용한 위임(리서치, 코드 리뷰, 도메인별 시스템 프롬프트)과 컨텍스트를 소모하고 꼭 필요한 정보를 잃게 만드는 안티패턴(전문가 페르소나, 순차 파이프라인, 테스트 러너) 사이의 선을 명확히 그어줍니다. ## [00:03] 도입: 서브에이전트가 도움이 될 때와 역효과가 날 때 시리즈에서는 지금까지 서브에이전트를 만들고 설계하는 법을 다뤘습니다. 마지막 편은 배포 관점의 질문으로 넘어갑니다. 어떤 작업이 별도 에이전트를 띄울 때 진짜 이득이 되고, 어떤 작업이 오히려 손해를 보는가? 답은 하나의 검증으로 귀결됩니다. 중간 작업이 메인 스레드에 중요한가? 탐색과 실행이 분리되어 있을 때 서브에이전트는 값어치를 합니다. 각 단계가 이전 단계의 발견에 의존할 때는 인계 비용이 꼭 필요한 세부 내용을 앗아갑니다. > *"간단히 말해, 중간 작업이 메인 스레드에 중요한지 여부가 핵심 차이입니다."* ## [00:32] 리서치 작업: 탐색을 격리된 상태로 유지하기 인증 추적은 구체적인 예시입니다. 메인 스레드가 알아야 할 것은 JWT 검증이 어디서 일어나는가 — 중간에 읽은 수십 개의 파일이 아닙니다. 리서치 서브에이전트는 코드베이스 전체를 스캔하고, 파일을 넘나들며 함수 호출을 추적해 정확한 답 하나를 돌려줄 수 있습니다. JWT 검증은 middleware/auth.js의 42번째 줄에 있고, route/api.js에서 호출됩니다. 그 모든 탐색은 서브에이전트의 컨텍스트 안에 고스란히 남습니다. 메인 스레드는 결론만 받고, 검색 기록이 컨텍스트 창을 어지럽히지 않은 채 앞으로 나아갑니다. > *"메인 스레드는 이렇게 받습니다: JWT 검증은 middleware/auth.js의 42번째 줄에 있고, Express 라우터와 route/api.js에서 호출된다 — 뭐 이런 식으로."* ## [01:15] 코드 리뷰 서브에이전트: 새로운 시각으로 피드백 받기 Claude가 스스로 작성에 참여한 코드를 리뷰하면 편향이 생깁니다. 모든 결정 과정에 있었기 때문에 외부 시각에서 무엇이 이상해 보이는지 쉽게 포착하지 못합니다. 리뷰어 서브에이전트는 이를 완전히 우회합니다. 코드가 어떻게 발전해왔는지에 대한 이력 없이, diff와 수정된 파일만 봅니다. 이 깨끗한 출발점은 두 번째 이점도 만들어냅니다. 프로젝트 고유의 리뷰 기준 — 명명 규칙, 보안 패턴, 아키텍처 규칙 — 을 서브에이전트의 시스템 프롬프트에 한 번 새겨두면 매번 메인 스레드가 기억에 의존하지 않고도 일관되게 적용됩니다. > *"리뷰어 서브에이전트는 별도 컨텍스트에서 변경 사항을 봅니다. git diff를 실행하고 수정된 파일을 읽은 뒤, 코드가 작성된 이력 없이 전문화된 리뷰 기준을 적용합니다."* ## [01:59] 커스텀 시스템 프롬프트: 카피라이팅과 스타일링 Claude Code의 기본 프롬프트는 간결하고 기술적인 출력에 최적화되어 있습니다. 랜딩 페이지나 마케팅 이메일에는 정반대가 필요합니다. 카피라이팅 서브에이전트는 톤, 대상 독자, 구조에 대해 완전히 다른 지침을 받아 메인 스레드의 기본값으로는 절대 나오지 않을 결과물을 만들어냅니다. CSS에도 같은 논리가 적용됩니다. 디자인 시스템 파일을 언급하는 스타일링 서브에이전트는 한 줄을 쓰기 전에 컬러 변수, 간격 규칙, 컴포넌트 패턴을 자동으로 컨텍스트에 불러옵니다. 모든 스타일 결정이 합리적 추측이 아닌 실제 시스템을 반영하도록 보장합니다. > *"Claude Code의 기본 프롬프트는 간결하고 기술적인 글쓰기 쪽으로 치우쳐 있어서, 랜딩 페이지나 이메일 캠페인에는 어울리지 않습니다 — 고객을 잠들게 하고 싶지 않다면요."* ## [02:57] 안티패턴: 전문가 주장, 파이프라인, 테스트 러너 세 가지 패턴이 반복적으로 상황을 악화시킵니다. 첫째, 페르소나 프롬프트 — "당신은 Python 전문가입니다" 또는 "당신은 Kubernetes 전문가입니다" — 는 아무것도 더하지 않습니다. Claude는 이미 그 지식을 갖고 있기 때문입니다. 전문가 레이블을 붙이기 위해 서브에이전트를 띄우는 것은 메인 스레드가 할 수 있는 일을 위해 격리 비용만 낭비하는 셈입니다. 둘째, 순차 파이프라인은 단계들이 진정으로 독립적이지 않을 때마다 무너집니다. 세 에이전트 흐름 — 버그 재현, 디버그, 수정 — 은 깔끔해 보이지만 실제로는 실패합니다. 디버그 에이전트에게는 재현 에이전트의 라이브 컨텍스트가 필요하지, 압축된 요약이 아닙니다. 셋째, 테스트 러너 서브에이전트는 정보를 능동적으로 숨깁니다. 테스트가 실패하면 무엇이 잘못됐는지 파악하려면 날것의 출력이 필요합니다. "테스트 실패"만 돌려주는 서브에이전트는 직접 출력에서 바로 보였을 세부 정보를 얻기 위해 추가 디버그 스크립트를 강요합니다. > *"'테스트 실패'만 반환하는 서브에이전트는 직접 출력에서 바로 보였을 세부 정보를 얻기 위해 추가 디버그 스크립트를 만들게 합니다."* ## [04:10] 시리즈 정리와 핵심 판단 기준 시리즈 전반에 걸쳐: 서브에이전트는 요약을 돌려주는 격리된 스레드이며, /agents로 만들고, 구조화된 출력과 구체적인 설명으로 설계합니다. 리서치, 코드 리뷰, 커스텀 시스템 프롬프트가 필요한 작업에 활용하세요. 전문가 페르소나, 다단계 의존 파이프라인, 테스트 실행에는 쓰지 마세요. 모든 판단 틀은 한 가지 질문으로 수렴합니다. 중간 작업이 중요한가? 답이 아니라면 위임하세요. > *"핵심 질문: 중간 작업이 중요한가? 그렇지 않다면 위임하세요."* ## 등장 인물 - **Anthropic Tutorial Narrator** (인물): Claude Code 서브에이전트 튜토리얼 시리즈 진행자, Anthropic - **Claude Code** (소프트웨어): Anthropic의 AI 코딩 어시스턴트; 서브에이전트가 만들어지고 조율되는 환경 - **Subagent** (개념): 메인 컨텍스트에서 실행되는 격리된 Claude 스레드. 전체 작업 컨텍스트를 노출하는 대신 압축된 요약을 반환함 - **JWT (JSON Web Token)** (개념): 코드베이스 전반의 인증 로직을 추적하는 리서치 서브에이전트의 실습 예시로 사용됨 - **System prompt** (개념): Claude Code의 기본 프롬프트와 다른 도메인 특화 동작을 가능하게 하는 서브에이전트별 지침 세트 - **Anthropic** (조직): Claude 및 Claude Code 서브에이전트 튜토리얼 시리즈 개발사

Claude Code에는 기본 내장 서브에이전트가 있지만, 커스텀 서브에이전트를 만들면 특정 작업에 맞는 전문화된 동작을 직접 설정할 수 있습니다. 이 튜토리얼은 코드 리뷰 서브에이전트를 처음부터 만드는 과정을 다룹니다. `/agents` 명령어, 도구 선택, 모델 결정, 그리고 Claude가 언제 어떻게 위임하는지 제어하는 설정 파일의 필드까지 차례로 살펴봅니다. ## [00:03] 커스텀 서브에이전트란 무엇인가 Claude Code에는 내장 서브에이전트가 포함되어 있지만, 특정 작업을 전담하는 서브에이전트를 직접 만들 수도 있습니다. 커스텀 서브에이전트는 YAML front matter가 있는 마크다운 파일입니다. front matter는 Claude에게 해당 에이전트로 라우팅할 시점과 에이전트가 가진 기능을 알려주고, 마크다운 본문은 서브에이전트가 실행되는 시스템 프롬프트가 됩니다. > *"Custom sub aents are markdown files with YAML front matter. These markdown files contain configuration that helps claude understand when to use the sub aent and provides directions to the sub aent itself."* ## [00:28] /agents로 서브에이전트 만들기 `/agents` 명령어를 실행하면 에이전트 관리 패널이 열립니다. "새 에이전트 만들기"를 선택하면 두 가지를 묻습니다. 범위(현재 프로젝트 또는 머신의 모든 프로젝트에서 공유)와 생성 방법입니다. 권장 방법은 Claude가 자동으로 에이전트를 생성하도록 맡기는 것입니다. 내레이터가 코드 품질과 보안 문제를 검토하는 서브에이전트를 평범한 말로 요청하면 Claude가 나머지를 처리합니다. > *"Now, the easiest way to create a sub agent is with the / agents command. Next, you can create a sub agent manually, but we recommend using claw code to automatically generate it for you."* ## [00:56] 도구, 모델, 색상 설정하기 Claude가 파일을 작성하기 전에 서브에이전트가 접근할 수 있는 도구를 선택합니다. 코드 리뷰 에이전트는 편집 도구가 꼭 필요하지 않지만, 실행을 활성화해 두면 대기 중인 변경 사항을 더 쉽게 확인할 수 있습니다. 도구 선택 후 모델을 고릅니다. 속도 우선이면 haiku, 깊이 있는 분석은 opus, 그 중간은 sonnet입니다. 마지막으로 색상을 선택합니다. UI에서 서브에이전트를 한눈에 알아볼 수 있게 해주는 색상입니다. > *"Now, given that our sub agent is only responsible for reviewing code, you might decide to disallow tools for editing, but I'll leave an execution to allow the sub agent to more easily identify pending changes."* ## [01:43] 설정 파일 이해하기 생성된 파일은 요약 창에 표시된 경로의 프로젝트 내에 저장됩니다. 핵심 필드는 네 가지입니다. `name`은 고유 식별자로, 메시지에 `@agent-code-quality-reviewer`를 입력해 참조할 수 있습니다. `description`은 Claude가 읽고 위임 여부를 결정하는 내용으로, 한 줄로 작성해야 합니다(이스케이프된 `\n`은 그대로 literal 문자입니다). description에 "proactively"를 추가하면 Claude가 에이전트를 더 자주 사용하고, 예시 대화를 추가하면 라우팅이 더 정확해집니다. `tools`는 생성 시 부여된 접근 권한을 반영하지만 파일에서 직접 편집할 수 있습니다. > *"If you want Claude to use the sub agent automatically more often, add in the word proactively to the description."* ## [02:41] 시스템 프롬프트와 Claude의 활용 방식 `model` 필드는 `haiku`, `sonnet`, `opus`, `inherit` 중 하나를 받습니다. `inherit`는 서브에이전트를 상위 대화와 같은 모델로 실행합니다. front matter 아래의 모든 내용이 시스템 프롬프트입니다. 서브에이전트가 작업을 수행하는 방법과 결과를 메인 에이전트에게 돌려주는 방법을 안내합니다. > *"The system prompt will provide guidance to the sub agent, helping it understand how to complete its task and how it should return information back to the main agent."* ## [03:15] 서브에이전트 테스트하기 설정을 저장한 뒤 코드를 수정하고 Claude에게 검토를 요청합니다. 서브에이전트가 예상대로 트리거되지 않으면 `description` 필드부터 조정하세요. 더 구체적인 예시를 추가하면 Claude가 언제 위임해야 할지를 더 정확하게 파악합니다. > *"If the sub agent isn't being used when you expect, check your description. Adding more specific examples helps Claude understand when to delegate."* ## 등장인물 - **Anthropic Tutorial Narrator** (사람): 이 에피소드의 단독 진행자; Anthropic 공식 YouTube 채널에서 Claude Code 서브에이전트 튜토리얼 시리즈를 진행함 - **Claude Code** (소프트웨어): Anthropic의 AI 코딩 어시스턴트; 내장 서브에이전트와 사용자가 만든 커스텀 서브에이전트를 모두 지원 - **커스텀 서브에이전트** (개념): YAML front matter가 있는 마크다운 파일로, Claude Code가 특정 작업을 전문화된 에이전트 인스턴스에 위임하도록 설정 - **/agents command** (개념): 서브에이전트를 만들고 관리하는 Claude Code UI 진입점; 프로젝트 범위 또는 전역 범위 제공 - **시스템 프롬프트** (개념): 서브에이전트 설정 파일의 마크다운 본문; 런타임에 서브에이전트에게 작업 지침과 출력 형식 안내를 제공 - **Anthropic** (조직): Claude 및 Claude Code 플랫폼의 개발사